Ed Zitron Predicts $500B AI Data Center Debt Bomb Will Trigger Tech Crash
The Diary Of A CEO
Generative AI operates on unsustainable token subsidies and circular funding that will collapse tech stock valuations when massive infrastructure costs hit balance sheets by 2027.
## Background
Published by The Diary Of A CEO on YouTube on August 27, 2026, this video features host Steven Bartlett in conversation with tech critic Ed Zitron. According to the page description, Zitron exposes the AI bubble, examining why companies like OpenAI and Anthropic face severe operational losses, the financial risks surrounding data center infrastructure, and why claims of imminent superintelligence or total job replacement are misleading.
## TL;DR
Tech critic Ed Zitron argues that generative AI represents an unsustainable financial bubble built on circular investments, massive capital expenditures, and unsubstantiated claims of business transformation. He breaks down the unit economics of AI models, highlighting that token costs far exceed current subscription prices while model improvements hit severe technical plateaus. Zitron exposes how tech giants mask poor returns through paper gains, aggressive marketing, and fears of missing out, warning that overspending on non-viable infrastructure could trigger a broader economic crash. Ultimately, he advocates for valuing human judgment, expertise, and genuine connection over hyper-scaled probabilistic software.
## Atmosphere & style
The interview has an intense, confrontational, and deeply candid atmosphere as Ed Zitron passionately dismantled Silicon Valley's dominant narratives. Zitron's style is sharp, energetic, and unsparingly critical, balancing detailed financial analysis with sharp wit and raw frustration toward tech leadership. Steven Bartlett provides a calm, probing counterweight, pushing back with common counterarguments and forcing Zitron to systematically defend his thesis. The dynamic feels like a high-stakes debate between enthusiastic corporate optimism and grounded, unvarnished skepticism.
## Key takeaways
- [02:36](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=156s) Generative AI is a 'con' and a massive financial bubble driven by tech industry circular financing and unfulfilled claims of autonomous capability.
- [03:20](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=200s) Over 70% of total AI software revenue belongs to OpenAI and Anthropic, both of which operate at colossal net losses despite billions in corporate funding.
- [06:15](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=375s) AI data centers require immense energy and multi-billion-dollar CapEx buildouts that risk creating $500 billion in stranded, debt-laden assets.
- [08:02](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=482s) High enterprise AI adoption metrics are largely coercive, counting passive software integrations rather than organic workflow shifts.
- [12:00](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=720s) Generative AI companies obscure true token inference costs via subsidized flat-rate pricing models that collapse when enterprise clients pay actual operational rates.
- [16:09](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=969s) Tech giants are pouring hundreds of billions into CapEx and GPUs despite global non-circular AI software revenue sitting at a fraction of that spend.
- [20:07](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1207s) AI capability improvements are hitting a plateau, with current benchmark gains often stemming from training on test sets rather than genuine reasoning breakthroughs.
- [24:23](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1463s) AI 'hallucinations' and probabilistic errors remain a core limitation, rendering autonomous multi-hour tasks unreliable and creating enterprise security risks.
- [52:46](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=3166s) Generative AI acts as a 'slop machine' that lowers software quality, clutters the open web, and degrades search engines like Google.
- [01:10:22](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=4222s) Claims of imminent superintelligence or total white-collar job displacement function as performative marketing designed to force urgent enterprise buy-in.
- [47:48](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2868s) Tech companies deliberately group specialized machine learning (like DeepMind's AlphaFold) with generative LLMs to mask the poor financial returns of LLM infrastructure.
- [01:45:12](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6312s) Tech CEOs use circular investments, paper gains, and obscure metrics like basis points to maintain stock valuations and justify hyper-inflated CapEx budgets.
- [02:00:50](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7250s) Media narratives claiming AI can independently blackmail humans or act autonomously rely on manufactured prompt conditions rather than emergent machine intent.
- [02:08:28](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7708s) If OpenAI or Anthropic fail to justify their negative unit economics, the resulting contraction in hardware spend could trigger a broader 'Tech Depression' affecting mainstream markets.
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=0s) The AI "Con" and the Broken Economics
The generative AI boom is fundamentally a "con" built on misleading promises of magic. Tech critic Ed Zitron argues that despite 16 years of working enthusiastically in the technology industry, he finds the current trajectory revolting. Tech leaders are selling cloud software that is unreliable, expensive, and fundamentally incapable of autonomous intelligence as a world-changing miracle meant to cure cancer and replace human labor. Far from being a natural evolution, this represents the largest "non-consensual push of technology in history," driven by aggressive marketing that preys on weak media reporting and sell-side financial analysts rather than genuine product excellence.
```
+-----------------------------------------------------------------------+
| THE AI SUBSIDY GAP |
| |
| User Pays: $200/mo Subscription |
| Actual Cost: |-------------------| $14,000 in Compute/Tokens |
| |
| Net Loss: -$13,800 per heavy user / month |
+-----------------------------------------------------------------------+
```
When examining the actual financials of major tech players—such as OpenAI, Anthropic, Amazon, Nvidia, Microsoft, and Google—their core revenues are not organically derived from AI products. Over 70% of total AI software revenue comes from OpenAI and Anthropic, both of which operate at staggering, unsustainable losses. For instance, OpenAI lost $20.9 billion in a single year. These platforms survive only because tech giants pump capital back into them: Amazon poured $5 billion into Anthropic and sent $50 billion to OpenAI, while Google routed $10 billion to Anthropic. Wall Street analysts expect these two unprofitable entities to deliver over $400 billion in revenue over the next three and a half years to justify massive cloud growth expectations, creating a circular financial illusion.
- Publicly traded tech companies routinely obscure their true AI returns by hiding behind vague metrics like "annualized run rate" without standardized definitions.
- Generative tools fail at basic tasks without complex prompt engineering ("Pee-wee's Playhouse setup"), contradicting claims of true artificial autonomy.
[06:15](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=375s) Data Center Energy Demands and Capital Expenditure
The sheer scale of capital expenditure (CapEx) allocated to AI hardware and infrastructure is disconnected from economic reality. Over a trillion dollars has already been spent on data centers and high-performance GPUs, with another trillion planned for the following year. These massive investments require dedicated energy infrastructure on a scale never seen before in consumer computing.
A prime example of this extreme resource concentration is the "Stargate Abilene" data center project built by OpenAI and Oracle in Abilene, Texas.
To support this massive 1.2-gigawatt power draw, each of the eight buildings in the Stargate Abilene complex is designed to house 50,000 Nvidia GB200 GPUs. Packing this amount of electrical demand into a footprint over a thousand times smaller than a mid-sized city creates unprecedented strain on financial and power resources. Unlike historical infrastructure booms like railways, high-end AI GPUs have little utility outside of running these specific models, leaving tech companies burdened with immense debt for assets that generate minimal relative return.
[08:02](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=482s) Coerced Adoption vs. Genuine Market Demand
The apparent widespread adoption of generative AI is not a result of organic consumer enthusiasm, but rather aggressive product integration and fear-mongering. Users encounter AI constantly because tech platforms force it into everyday workflows—whether it is Gemini interrupting users in Google Docs, Copilot popping up in Microsoft Word, or shopping assistants like Rufus suggesting products on Amazon. For years, media outlets have warned the public that they will lose their jobs or fall behind if they do not adopt these tools, driving panicked usage.
```
[ Enterprise Integration ] ---> [ Media Hyping / Fear ]
\ /
v v
+-------------------------------+
| Forced & Coerced Usage |
+-------------------------------+
|
v
+-------------------------------+
| Inefficient Query Usage |
| (e.g., glorified search) |
+-------------------------------+
```
While users frequently employ LLMs as a replacement for standard web search to digest or scour queries, this utility does not justify hundreds of billions in capital investment. Furthermore, AI output is rapidly degrading the broader web by generating massive volumes of automated text. Much like earlier Search Engine Optimization (SEO) tactics ruined web search by prioritizing low-quality content farms over well-researched human writing, generative AI acts as a "slop machine," scaling up low-grade content across the internet.
- Surveys indicating 88% enterprise adoption often count passive or single-use instances rather than deep workflow transformations.
- AI search tools frequently act as broad net-trawling operations, yielding generalized median responses rather than verified factual precision.
[12:00](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=720s) The Hidden Token Cost and the Subsidization Trap
The underlying unit economics of generative AI reveal a massive gap between consumer pricing and operational execution costs. Models process data using "tokens" (roughly three-quarters of a word), charging for both input text (documents, code bases) and output text (generated responses and internal reasoning tokens). Because consumer subscriptions obscure these costs behind flat monthly fees, users remain entirely unaware of the astronomical computing expenses generated by their activity.
Independent financial evaluations reveal severe cost imbalances across popular subscription tiers:
- **$20/month Standard Plan:** Users can easily burn through **$400** worth of computing tokens.
- **$200/month Enterprise/Power Tier:** Heavy usage can consume **$8,000 to $14,000** in underlying token costs.
This gap means AI providers end up heavily subsidizing power users. When companies attempted to transition enterprise clients (organizations over 150 employees) toward paying the true cost of their token consumption, corporate clients revolted. For example, Uber exhausted its entire annual token budget within just three months. Once forced to bear the actual operational expense rather than enjoying heavily subsidized rates, enterprise enthusiasm cools rapidly, exposing the fragility of the entire business model.
[14:51](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=891s) The Economics of Inference and GPU Allocation
Ed Zitron details the financial realities behind running large language models, explaining that AI companies do not disclose full unit economics even in audited financial statements. When serving model outputs to users—a process known as inference—companies must estimate demand and provision expensive GPU capacity in advance. This creates a difficult operational trade-off: over-provisioning results in massive wasted capital for idle hardware, while under-provisioning leads to server outages, frustrated users, and churn.
Zitron compares the strategy of AI providers to "selling $20 or $40 for a dollar." He argues that if these services were genuinely profitable and valued appropriately by the market, users would be billed for the true cost of compute rather than being offered flat-rate monthly subscriptions. Furthermore, users pay for API usage regardless of whether the output is correct, meaning customers subsidize the compute costs incurred when models hallucinate or produce broken code.
* Large language model providers play accounting games and obscure true operational costs in their financial disclosures.
* Inference requires paying hourly rates for GPU compute whether the hardware is actively processing queries or sitting idle.
* Under a flat monthly subscription model, power users consume far more compute value than the $20 or $40 fee covers.
* API users pay for all generated tokens, including incorrect outputs, hallucinations, and code that breaks existing software.
---
[16:09](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=969s) VCs, Tech Giants, and the $100B+ AI Capital Expenditure Gap
Responding to host Steven Bartlett's point about tech companies historically spending ahead of value to capture market share, Zitron asserts that hardware and compute costs have failed to decline as predicted. He notes that neither GPU cloud rental services nor direct inference providers are currently profitable. While venture capitalists initially assumed chip efficiency and customer willingness to pay would catch up by 2026, massive operating losses—such as OpenAI losing over $5 billion in 2024—were ignored because the narrative inflated stock market valuations across the tech sector.
```
+-----------------------------------------------------------------------+
| MICROSOFT FY2026 CAPITAL EXPENDITURE VS REVENUE |
+-----------------------------------------------------------------------+
| |
| Total AI Revenue Reported: $34.33 Billion |
| ├─ Revenue attributed to OpenAI: $24.10 Billion |
| └─ Direct Microsoft AI Revenue: $10.23 Billion |
| |
| Capital Expenditures (CapEx): $115.00 Billion |
| Planned Next-Year CapEx: $175.00 Billion |
| |
+-----------------------------------------------------------------------+
```
Zitron highlights the stark disconnect between capital expenditures and direct AI revenue among major tech companies:
> "In fiscal year 2026, which just ended for Microsoft... they made total according to Bloomberg about $34.33 billion. $24.1 billion of that was from OpenAI, so that leaves them with about $10 billion in a year when they spent $115 billion on capital expenditures... and intend to spend $175 billion next year. The math does not make sense."
* Nvidia generated $215.9 billion in its last fiscal year primarily from selling GPUs.
* Global AI revenue outside of circular investments between tech giants sits at approximately $22 billion.
* Revenue growth at Google, Meta, and Amazon stems primarily from core advertising business model changes, cloud growth, and e-commerce price increases, not direct generative AI sales.
* Zitron cites market analyst Edward Zitron/Prophy Markets, pointing out a "cult-like worship of the wealthy" that leads the public to assume tech executives must have a secret, hyper-rational master plan.
---
[20:07](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1207s) The Limits of Innovation Analogies and Software Quality Decay
Bartlett brings up Clayton Christensen's classic business book *The Innovator's Dilemma*, comparing early AI shortcomings to 19th-century automobiles. Early cars frequently broke down, cost more than walking, and were subject to the "Red Flag Act" requiring a person to walk in front of them with a red flag. Yet cars eventually surpassed horse-and-carriage transportation due to a higher ceiling of technological growth. Bartlett asks whether assuming a compound rate of improvement (e.g., 5% monthly efficiency gains) makes AI's current trajectory sensible.
Zitron rejects the comparison, pointing out that hardware scaling does not follow Moore's Law when applied to GPUs. Nvidia's software ecosystem—built on CUDA libraries developed over decades—requires physical energy and hardware scaling that faces diminishing returns. Furthermore, Zitron contends that software quality across major platforms like Google, Microsoft, Meta, and GitHub is actively degrading as companies force developers to rely on AI-generated code.
> "Microsoft's one of the largest companies in the world, and they can barely wipe their own ass when it comes to GitHub. The quality of software is going down weirdly enough as more people use LLMs and more businesses demand... that people use these services."
* Nvidia developed CUDA in the 2000s, taking over a decade to adapt the library for big data analytics before applying it to deep learning.
* Unlike historic technology transitions, current AI adoption is backed by unprecedented government backing and capital investment from the outset.
* Evaluating developer productivity based on "lines of code shipped" misinterprets software quality, rewarding volume over stability and maintainability.
* GitHub outages and code degradation have become increasingly frequent as tech companies force internal AI coding tools.
---
[24:23](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1463s) Hallucinations, Benchmarks, and Real-World Security Risks
Addressing Bartlett's statement that recent frontier models rarely hallucinate on daily tasks, Zitron shares a concrete example using his Bloomberg Terminal. While using the terminal's natural-language query tool ("Ask B") to pull stock growth rates for Microsoft, Google, Meta, and Amazon, the system generated a fictional historical stock price, claiming Microsoft had traded at $575 per share.
Zitron emphasizes that while small errors are harmless in causal queries, they become dangerous when integrated into medical transcription, hedge fund financial modeling, or automated code refactoring.
Bartlett counters by referencing historical data from the Vectara Hallucination Leaderboard, which shows hallucination rates on simple text summarization dropping from 21.8% four years ago to 0.7% on top frontier models like Gemini and ChatGPT. Zitron clarifies that benchmarks measure narrow, simplistic tasks under ideal conditions rather than complex real-world workflows, where model failure rates jump drastically when running autonomous, multi-step tasks over long durations.
* Vectara Leaderboard data demonstrates a decline in summarization hallucination rates from 21.8% to 0.7% on simple tasks.
* Frontier models operating on complex, multi-hour autonomous tasks still fail or encounter errors roughly 50% of the time.
* "Vibe coding"—relying entirely on LLMs to write code without manual review—introduces hidden security vulnerabilities and subtle logic bugs.
* A hallucination in automated refactoring can silently compromise code bases and leave system backdoors open.
---
[27:42](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1662s) Dial-Up Internet vs. Human Capabilities
Bartlett draws an analogy to the early days of dial-up internet, remembering how he had to disconnect from *RuneScape* so his family could use the landline phone. Just as dial-up evolved into 5G and Starlink, he argues AI will follow a similar leap in convenience and reliability. Additionally, Bartlett posits that AI outputs should not be compared against abstract perfection, but against the human alternative—such as an intern or entry-level employee who also makes mistakes, has knowledge gaps, works slower, and costs significantly more.
```
+-----------------------------------------------------------------------+
| HUMAN WORKER VS. AI ALTERNATIVE |
+-----------------------------------------------------------------------+
| |
| HUMAN EDITOR (e.g., Matt Hughes) AI MODEL (Frontier LLM) |
| -------------------------------- ------------------------- |
| • Deep institutional context • 0.7% hallucination rate |
| • Moral support & empathy • Instant processing speed |
| • Nuanced taste & judgment • Zero lived human context |
| • Higher monetary cost • Cheap nominal token cost |
| |
+-----------------------------------------------------------------------+
```
Zitron responds by citing his long-time editor, Matt Hughes, a decorated tech journalist based in Liverpool. Zitron explains that he pays a human editor not because the editor possesses flawless encyclopedic knowledge, but because of his deep institutional context, critical judgment, empathy, and taste—qualities an automated text generator cannot replicate.
> "I don't pay Matt Hughes because he knows everything. I pay him because he has incredible context and a ton of knowledge and he's willing to expand it and work with me and moral support and he's a great editor... and on top of that a wonderful loving being with empathy and joy in his heart."
* Early consumer internet limitations (e.g., choosing between landline calls and online gaming) were physical bandwidth constraints, whereas AI limitations stem from fundamental mathematical probability in token prediction.
* Evaluating AI against human workers solely on error frequency ignores the vital roles of shared context, emotional intelligence, and real-world experience.
* human collaboration provides moral support, accountability, and nuanced taste that raw statistical models cannot generate.
[29:30](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1770s) Human Connection vs. Artificial Memory in AI
The debate opens with a sharp contrast between human interaction and AI capabilities. Ed Zitron pushes back against the premise that large language models "know everything," questioning the utility of a tool that occasionally knows nothing due to hallucinations. While an intern is a human being who learns context, adapts to habits, and undergoes a developmental transition—taking them out of "Pinocchio status"—AI does not truly learn or retain context organically. Instead, making AI retain information requires creating complex harness systems or `.md` files that the model only sometimes reads, resembling "Pee-wee's breakfast machine from Pee-wee's Playhouse" just to mitigate errors.
When Steven Bartlett brings up a simple example—noting that Claude can remember his dog's name from past conversations—Zitron expresses exasperation at the fact that a company valued in the tens of billions is being defended with such basic functionality. Zitron emphasizes that retrieving stored text files is not true memory. Human memory is a complex amalgam of emotion, feelings, and lived experiences. An intern grows as an investment, whereas an AI remains fundamentally a "text machine" incapable of knowing moods, understanding the surrounding world, or offering genuine human presence.
- **Human growth vs. AI prompts:** Interns build real context and emotional awareness over time, whereas LLMs rely on external file ingestion and engineered prompt harnesses to mimic recall.
- **The mechanics of AI context:** To make an LLM retain user habits, developers must feed it explicit text files or context windows, which the system often fails to process consistently.
- **Emotional depth:** Human memory accrues through experience, feelings, and mutual understanding, whereas AI simply operates as a text retrieval mechanism.
[31:31](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1891s) Process vs. Output: The Value of Learning and Trust
Steven Bartlett argues that users ultimately evaluate AI based on its output rather than its internal process. From a business perspective, if an AI is asked for a revenue figure or a creative task, the consumer cares whether the result is accurate and useful, not whether the machine experienced a human-like cognitive journey. He asserts that when paying for a product, the outcome is what generates value.
Zitron strongly disagrees, insisting that the process itself holds immense value, particularly in collaborative work. Sharing an anecdote about writing an 11,000-word piece with his editor, Matt Hughes, Zitron describes the shared joy and bonding that came from going down research rabbit holes together—such as discovering the misanthropy of asset managers like Blackstone. That back-and-forth learning process actively shapes the final work in a way an LLM cannot replicate, as AI merely averages existing web documents to produce generic "slop." Furthermore, Zitron stresses that trust is integral to output; he trusts a human collaborator like Matt who deeply understands the context, whereas AI outputs—especially detailed financial or analytical reports—become increasingly prone to error the more complex they get.
- **Outcome vs. Journey:** Bartlett contends that end-users prioritize whether the AI gives the correct number or asset, while Zitron argues that the collaborative process drives true novelty and comprehension.
- **Generic "Slop":** Zitron notes that even high-end LLM harness systems used by hedge funds deliver the same generic, average reports that lack genuine insight.
- **Basis of Trust:** Human trust is built on shared learning, historical performance, and mutual accountability, whereas AI outputs must be continuously triple-checked.
[35:11](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2111s) Benchmarks, the iPhone Analogy, and Genuine Utility
The discussion shifts to how AI companies measure and market their progress. Bartlett references hallucination leaderboards that show models consistently delivering on simple tasks. Zitron counters that the tech industry uses vague benchmarks and "whataboutism" to hide the lack of direct utility. Unlike revolutionary technologies of the past, AI companies struggle to define simple parameters for their success or explain their value without complex technical workarounds.
To illustrate what true tech disruption looks like, Zitron recalls getting the original iPhone while attending Penn State. The immediate value of features like visual voicemail was obvious to everyone—from tech enthusiasts to ordinary people using flip phones or Nokias. Even though Steve Ballmer, then CEO of Microsoft, famously laughed at the iPhone for costing $500 subsidized and lacking a physical keyboard, the product's intuitive graphical user interface (GUI) and clear utility required no elaborate explanation. In contrast, Zitron points out that proponents of AI are still dithering years later, unable to show how to run a business on LLMs without relying on a complicated "system of pulleys and levers."
> "When the first iPhone came out... I felt like the apes at the beginning of 2001. Visual voicemail—it was immediate... Everyone was like, 'Holy sh*t, this is incredible.' It was obvious the value."
- **Steve Ballmer's Skepticism:** In a famous archival clip, Microsoft's Steve Ballmer mocked the $500 iPhone as the "most expensive phone in the world" that wouldn't appeal to business customers because it lacked a keyboard.
- **Instant Value vs. AI Hype:** Past breakthroughs like the iPhone, AWS, and early mobile web offered immediate, self-evident utility, whereas AI requires constant explanation and complex framing.
- **System of Pulleys and Levers:** Zitron argues that LLMs cannot autonomously handle complex workflows without constant human intervention and harness engineering.
[38:37](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2317s) Adoption Rates, Honest Pricing, and the Environmental Cost
Bartlett presents adoption metrics to challenge Zitron's skepticism, pointing out that ChatGPT reached 100 million active users in just 60 days—far faster than TikTok (9 months), Instagram (2.5 years), or the World Wide Web (7 years). He shares a personal example of his fiancée, a solo entrepreneur and non-native English speaker, who uses ChatGPT to write copy and generate images, describing it as transformative for her business.
Zitron responds by raising the issue of subsidization. He asks whether users would still find these tools transformative if they were forced to pay the "honest cost"—the actual per-million-token API cost required to run these models without massive venture capital subsidies. Citing a 2024 Goldman Sachs report by Jim Covello titled *Gen AI: Too Much Spend, Too Little Benefit*, Zitron highlights that previous tech revolutions like mobile connectivity had clear roadmaps based on shrinking hardware components (GSM, Bluetooth, Wi-Fi radios). AI has no such roadmap to profitability or superintelligence. Instead, tech executives hype LLMs as world-changing to justify immense data center expenditures, which cause real-world harm—such as air pollution from gas turbines near residential neighborhoods in Louisiana and rising electricity bills and hardware inflation for consumers.
- **Rapid Consumer Adoption:** Over 60% of US adults integrated AI tools into their routines within three years of launch, hitting a 40% population adoption rate faster than personal computers (12 years) or the internet (5 years).
- **The Subsidization Trap:** Zitron asserts that consumer enthusiasm relies on artificially low prices, predicting usage would plunge if users paid $2 to $4 per task to cover real compute costs.
- **Jim Covello's Goldman Sachs Report:** The 2024 report argued there is "too much spend for not enough return" in generative AI, noting a lack of a clear technological roadmap compared to past hardware cycles.
- **Environmental & Economic Strain:** Data centers driven by AI demand consume massive energy, driving up local electricity costs, causing hardware inflation due to high RAM demand, and polluting local communities.
[43:40](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2620s) Historical Tech Skepticism and the Dot-Com Analogy
The conversation opens with a debate over whether current critiques of AI echo historical misjudgments made about past technological shifts. The interviewer highlights how the dot-com bubble had plenty of high-profile skeptics in the late 1990s who ultimately looked foolish. Nobel Prize–winning economist Paul Krugman famously predicted in 1998 that by around 2005, the internet’s impact on the economy would prove no greater than that of the fax machine, even adding later that internet growth would slow drastically once people realized they had nothing to say to each other. Similarly, astrophysicist Clifford Stoll wrote a infamous 1995 *Newsweek* piece asking if computer pundits lacked common sense, arguing that no online database would replace daily newspapers, no CD-ROM would replace competent teachers, and that "cyber malls" were getting zero business compared to local shopping malls. Ed Zitron quips that Krugman’s claim about people having nothing to say to each other might be the worst prediction ever, suggesting Krugman had clearly never hung out in a middle-American bar.
However, Zitron pushes back against using these historic quotes to justify the current AI boom, explaining that the economics are fundamentally different. While Clifford Stoll made some blunder points, he was actually right about online education being a poor replacement for traditional schooling and about the dangers of an overwhelming flood of bad information online. More importantly, Zitron emphasizes that the dot-com crash was actually two separate events: a small, absurd "website bubble" (where companies like Excite@Home bought e-greeting card platforms for a billion dollars) and a massive infrastructure bubble centered around "dark fiber." Telecommunications companies buried vast networks of fiber-optic cables believing internet demand was doubling every 90 days, when actual growth was taking 6 to 12 months. This led to a massive overbuild of unlit cable and transmission infrastructure that lay dormant until actual consumer demand eventually caught up years later.
- **Paul Krugman's 1998 Prediction:** Stated that by ~2005, the internet's economic impact would be no greater than the fax machine and predicted growth would stall because people had nothing to say to each other.
- **Clifford Stoll's 1995 *Newsweek* Article:** Dismissed e-commerce, online news, and digital education, asking why local malls did a "roaring business" while cyber malls got zero traffic.
- **The Dot-Com Fiber Overbuild:** Telecom analysts incorrectly estimated internet demand was doubling every 90 days, leading to massive overinvestment in underground fiber-optic networks.
---
[46:01](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2761s) The Unfavorable Economics of Generative AI Data Centers
Zitron argues that assuming generative AI will follow the post–dot-com trajectory—where real demand eventually grows into overbuilt infrastructure—is a dangerous miscalculation. Current demand for generative AI is overwhelmingly subsidized; most consumers and businesses using it today are not paying anywhere near the actual cost of running the models. Furthermore, despite being pushed by the largest, most aggressive marketing campaign in human history, generative AI is generating surprisingly little revenue. Microsoft, the "apex predator of cloud software," only pulls in single-digit billions from selling AI software, and outside of OpenAI and Anthropic, the rest of the industry barely generates $22 billion annually. That figure is negligible compared to the $1.1 trillion in cloud commitments that OpenAI and Anthropic have racked up, alongside hundreds of billions in capital expenditures.
Unlike passive fiber-optic cables laid in the ground during the late 1990s—which cost little to maintain once buried—a modern GPU data center is extraordinarily expensive to keep running. A data center built today will be just as expensive to power and operate in 2050 unless there is a monumental breakthrough in electricity generation or chip architecture. Yet chipmakers like Broadcom, Nvidia, and Arm are already squeezing maximum performance out of physical silicon limits. Even with Nvidia's new Vera Rubin architecture, the company remains coy about bottom-line savings, promoting "10x efficiency" in terms of dollars per megawatt rather than claiming actual 50% cost reductions for customers like OpenAI. Zitron asserts that if real cost-cutting were happening, tech executives would be shouting it from the rooftops.
- **Generative AI Revenue vs. Spend:** Outside OpenAI and Anthropic, the broader AI industry generates roughly $22 billion in revenue, juxtaposed against $1.1 trillion in total cloud commitments made by the top two AI labs.
- **2026 AI Funding:** AI startups and labs have raised over $217 billion in 2026 alone to prop up operations.
- **Data Center Operational Costs:** Unlike static fiber cables, GPU data centers require continuous, massive electrical power and high maintenance costs that do not scale down automatically over time.
---
[47:48](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2868s) Conflating General AI Research with Generative Large Language Models
A critical reason for the current hype, Zitron explains, is that tech companies intentionally lump all forms of artificial intelligence under a single marketing umbrella. They conflate narrow, domain-specific machine learning models—such as DeepMind’s AlphaFold for protein folding, genomic sequencing tools, robotics, and climate modeling—with generative Large Language Models (LLMs). When critics point out the severe economic and functional flaws of LLMs, defenders counter by pointing to breakthroughs in cancer research or physics, even though those breakthroughs have nothing to do with LLMs.
Zitron points out that the hundreds of gigawatts of data centers currently planned or under construction are not being built for specialized scientific computing or Google’s pre-generative search algorithms, which run efficiently on traditional CPU-heavy infrastructure. They are being built almost exclusively for generative AI inference and training. Citing data from Sightline Climate from February, he notes there are 190 gigawatts of data centers planned—translating to roughly 12 million megawatt-hours and requiring between $1.6 trillion to $3 trillion annually in demand just to support the buildout. Yet the total market demand does not even reach $130 billion. The entire system operates as a circular economy: tech giants like Amazon, Microsoft, and Google funnel billions into OpenAI and Anthropic, who then hand that cash right back to buy cloud compute and Nvidia GPUs, creating a artificial illusion of organic market demand.
- **Planned Data Center Scale:** Cites Sightline Climate data showing 190 gigawatts of data centers in planning stages, which would require $1.6T–$3T annually to sustain.
- **Specialized Compute vs. LLMs:** Scientific applications (AlphaFold, climate forecasting, autonomous vehicle physics simulations) utilize High-Performance Computing (HPC) clusters rather than the massive, general-purpose LLM GPU clusters currently being constructed.
- **Non-GPU Robotics Example:** Meticulous consumer devices like the Matic floor-cleaning robot ("Turdsly") run efficient onboard algorithms without incurring continuous, budget-burning cloud GPU compute costs.
---
[52:46](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=3166s) How AI Coding Tools and Executive Decisions Are Degrading Tech
Addressing claims that generative AI provides massive internal productivity gains, such as writing code at companies like Google, Meta, and GitHub, Zitron argues that consumer software has actually become catastrophically worse and far less stable. He traces Google's quality decline back to 2019 executive decisions under Prabhakar Raghavan (then head of ads). Google declared an internal "code yellow" over a drop in search query volume. When search engineers like Ben Gomes and Shashi Thakur warned that artificially forcing query numbers up would require worsening search results and forcing users to click around more, their concerns were brushed aside. Raghavan later took over Google Search, spam site suppression was quietly stripped back, and eventually, generative AI summaries were slapped on top of search results—forcing users to stay on Google's page even when the AI generated absurd advice like eating rocks or toxic mushrooms.
Beyond product design, Zitron points to systemic technical instability caused by AI-generated code. When the interviewer cites industry research showing that software outages and downtime have demonstrably spiked due to buggy AI-assisted code flooding infrastructure, Zitron nods in agreement. Platforms like GitHub are being inundated with code written by inexperienced developers using LLMs who suffer from Dunning-Kruger effects—shipping code they barely understand into open-source projects and production environments. Major infrastructure, including Amazon Web Services and GitHub itself, has suffered multiple high-profile outages triggered directly by buggy AI code generation and automated deployments. Zitron concludes that instead of replacing software engineers as executives like Sam Altman ("Clammy Sammy") or Dario Amodei ("Wario") promise, AI tools are simply making human developers complacent and degrading software quality across the tech ecosystem.
- **Google's Search Quality Shift:** Internal 2019 "code yellow" led to strategies prioritizing increased user search queries over immediate answer accuracy, laying the groundwork for shoehorning Gemini AI summaries into search results.
- **Spike in Tech Outages:** Industry data directly links the recent rise in major cloud and software outages (including repeated AWS and GitHub disruptions) to the surge of AI-generated, unchecked code.
- **Executive Promises vs. Reality:** OpenAI's Sam Altman and Anthropic's Dario Amodei have repeatedly claimed AI will eliminate software engineering and up to 50% of white-collar labor, despite current tools causing widespread code bloat and system instability.
<ElicitationsGroup message="To explore these topics further:">
<Elicitation label="Examine the $1.1T cloud commitment structure between Big Tech and AI labs" query="Provide a detailed breakdown of the circular financial commitments between Microsoft, Amazon, Google, OpenAI, and Anthropic."/>
<Elicitation label="Analyze the history of the dot-com dark fiber overbuild vs. AI data centers" query="Compare the economic metrics and timeline of the 1990s dark fiber overbuild to the current 2026 AI GPU data center expansion."/>
<Elicitation label="Explore the impact of AI-assisted coding on open-source software quality" query="What metrics and studies exist regarding code quality, bug density, and repository maintenance challenges caused by LLM coding assistants?"/>
</ElicitationsGroup>
[58:18](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=3498s) Ads and Marketing Hype vs. Reality
Ed Zitron argues that AI executives are actively promising the world to consumers and investors rather than being honest about the technology's inherent constraints. If tech leaders were transparent, they would admit that generative AI is probabilistic, prone to mistakes, and prone to compounding errors when used by people who lack domain expertise. Instead, companies lean into sensationalized marketing, hyping up apocalyptic warnings or sci-fi superintelligence.
The discussion shifts briefly into host-read sponsorships highlighting how businesses can scale quickly using vetted expert freelancers through Fiverr Pro to keep up with strategy and technical workflows, as well as Saily, an eSIM application designed to save travelers money on international data roaming fees.
---
[01:00:51](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=3651s) The Reality of Autonomous Vehicles and Autonomous AI Promises
Addressing the claim made by Uber CEO Dara Khosrowshahi that autonomous vehicles will replace human drivers within a few years, Zitron asserts that tech CEOs are either outright lying or engaging in self-serving guesswork. Satya Nadella of Microsoft and Dara Khosrowshahi constantly shift their messaging to paint whatever narrative best benefits their company's stock price. While autonomous vehicles like Waymo and Zoox are fascinating engineering feats, their real-world rollouts remain heavily restricted to tightly controlled environments.
The primary hurdle in self-driving technology—and AI more broadly—is handling the remaining "edge cases," such as heavy rain or unusual pedestrian situations. Zitron shares personal anecdotes of seeing Zoox vehicles completely stall out in Las Vegas and block hotel exits, as well as a Waymo vehicle causing a traffic jam in San Francisco. While acknowledging safety statistics showing autonomous vehicles have lower crash rates per million miles than human drivers, Zitron points out that comparing small, highly supervised deployments to millions of human drivers over decades is a flawed comparison. Crucially, autonomous driving research operates under entirely different mechanics than generative LLMs.
- **Uber's Autonomous Claim:** Uber CEO Dara Khosrowshahi predicted full driver replacement within a couple of years.
- **Safety Statistics Cited by Host:** Autonomous vehicles show a 68% lower overall crash involvement rate, roughly 2.1 police-reported crashes per million miles compared to 4.68 for humans (a 55% reduction), and an 80% to 81% reduction in crashes resulting in injuries.
- **Sample Size Context:** Zitron emphasizes that human accident statistics represent decades of data across all conditions, whereas autonomous miles represent tiny, curated sample sizes overseen by human monitors.
---
[01:05:35](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=3935s) White-Collar Disruption and the Corporate AI "Cult"
Zitron dismantles the narrative that AI is causing widespread white-collar job loss in fields like law and accounting. In law firms, senior partners praise AI tools after messing around with ChatGPT, whereas the associates doing the actual research, filing, and heavy lifting remain unconvinced. OpenAI released a study revealing no correlation between enterprise spending on AI tokens and revenue per employee, effectively undermining their own pitch about AI-driven productivity gains.
The real job disruption currently happening is confined to vulnerable sectors—such as low-level graphics, translation, and transcription—where cost-cutting management teams treat output quality with indifference and would have outsourced the work anyway. Zitron highlights an Oxford Economics paper that was widely misreported as proof of youth unemployment caused by AI, when the study actually offered zero data or metrics to support such a link. The public's belief in these narratives stems from financial mythology: because stock prices rose, people assumed the technology must be working. Zitron compares fanatical online AI supporters to a "religious cult" or sports fans who aggressively defend companies like OpenAI or Anthropic while obsessing over GPU counts rather than real-world social problems.
- **OpenAI Internal Findings:** Studies released by OpenAI indicated no measurable correlation between token consumption and employee revenue productivity, alongside papers confirming hallucinations are mathematically guaranteed.
- **Displaced Workers:** The primary job losses are occurring among translators, transcribers, and entry-level digital artists whose employers prioritize saving money over baseline output quality.
- **Media Misinterpretation:** An Oxford Economics study cited by mainstream media claiming AI reduced youth hiring contained only a single vague sentence asserting a correlation without presenting actual numbers or data.
---
[01:10:22](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=4222s) Could Skepticism Be Helping Tech CEOs?
The host challenges Zitron by pointing out an unexpected overlap: early AI pioneers like Geoffrey Hinton and prominent tech CEOs spent years warning that AI was dangerous and posed existential risks to society. As public sentiment shifted and regulators took notice, these CEOs pivoted toward minimizing fears—a stance that curiously mirrors Zitron’s argument that AI isn't powerful enough to take everyone's jobs or overthrow humanity. The host jokingly suggests PR teams at major AI firms might actually be relieved by Zitron's narrative.
Zitron rejects the idea that his critiques aid tech CEOs, describing figures like Sam Altman of OpenAI and Dario Amodei of Anthropic as deeply cynical operators. He dismisses early corporate warnings about AI risk as performative theater meant to foster hype. To Zitron, claiming that one's technology is "so dangerous it might destroy the world" is simply a grand marketing ploy to convince investors and the public that the software is far more powerful than it actually is.
> "I think Altman and Amodei are some of the most deeply corrupt and cynical people in the world... Oh, shut up. I hear that and I feel so frustrated because I've met so many of these rich [expletive] liars."
<ElicitationsGroup message="Explore related aspects of the AI industry debate:">
<Elicitation label="Analyze OpenAI's financial projections and revenue metrics" query="What are the details regarding OpenAI's financial losses, token usage costs, and annualized revenue projections?"/>
<Elicitation label="Examine the arguments for AI-driven labor disruption" query="What evidence exists for and against generative AI displacing white-collar labor in fields like coding and finance?"/>
<Elicitation label="Review the debate on autonomous vehicle safety data" query="How do safety statistics for autonomous vehicles like Waymo compare against human driving metrics when controlling for test environments?"/>
</ElicitationsGroup>
[01:11:50](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=4310s) The Mechanics of Tech Fearmongering and Neoliberal Regulation
Ed Zitron explains how tech executives manipulate public urgency to force adoption and secure capital, drawing a direct parallel between modern AI pitch decks and historical cons. Every scam in history begins with a manufactured crisis of timing, scaring companies into buying software out of a fear of being left behind. His golden rule for evaluating high-pressure claims is simple: unless your house is on fire or you are actively bleeding, always slow down when someone tells you to hurry. He points out the hypocrisy of executives like Dario Amodei and Sam Altman periodically musing about slowing down progress due to safety concerns; in reality, they are only slowing down because they are running out of compute power.
Addressing Steven's observation that tech leaders shifted their rhetoric from existential "human extinction" threats to optimistic "ages of abundance" and universal access, Zitron reveals this narrative pivot is fundamentally about escaping government intervention. Tech leaders do not genuinely fear societal collapse; they fear being regulated or broken up like historical monopolies.
* "Every single scam and con starts with rushing you. Every single trick in history begins with saying you must do this now."
* "We are still trapped in the hands of Milton Friedman, Margaret Thatcher, and Ronald Reagan... a neoliberalistic hellscape which is growth at all cost."
* The shift in AI narratives tracked over time:
* **2015–2022:** Existential doom, human extinction, and uncontrollable dangers.
* **2023–2026:** "Age of Abundance," unlimited productivity, and "intelligence for everyone."
---
[01:14:22](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=4462s) Cybersecurity Threats, Compute Abuse, and "The Wrong Hands"
When discussing whether advanced AI creates legitimate, immediate hazards—such as autonomous cyber hacking—Zitron clarifies that the underlying tech isn't possessing sentient malice; it is simply brute-forcing vulnerabilities using staggering amounts of energy and compute. Recent security breaches blamed on AI "escaping the sandbox" were actually caused by human error, such as misconfigured servers left connected to the open internet. He dismisses the geopolitical scare tactics regarding national security, arguing that the technology is already in "the wrong hands"—namely, tech executives like Mark Zuckerberg, Sam Altman, and Dario Amodei, who are recklessly training models to hunt for software exploits simply because they have run out of valid text data to train on.
The solution to high-risk automated hacking is not fear-mongering about China or awaiting a rogue superintelligence; it is directly capping and regulating the massive compute resources these companies are allowed to burn.
* Autonomous scripts exploiting website vulnerabilities at scale are performing compute-heavy brute force rather than demonstrating human-level strategic intelligence.
* Tech leaders justify dangerous training runs under the guise of an international arms race, despite already having let the genie out of the bottle through unregulated infrastructure expansion.
---
[01:17:30](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=4650s) Mythbusting: Economic Growth and the Geopolitical "AI Race"
In a lightning-round challenge, Steven presents cards featuring popular AI industry narrative points for Zitron to debunk. On the claim that AI is generating unprecedented economic growth, Zitron fires back that there is zero macro data to support it. The revenue circulating within the ecosystem is predominantly an internal feedback loop: Nvidia supplying money to cloud providers like CoreWeave, and AI labs burning capital to pay infrastructure hosts like Oracle. Outside of speculative GPU hardware purchasing, actual enterprise and consumer spending on AI services barely touches $100 billion—a negligible return given that the industry has burned through $300 billion in venture equity and over $600 billion in capital expenditures.
Addressing the geopolitical claim that the US must spend trillions to win an "AI race" against China, Zitron questions the very premise of the competition. Chinese developers have already demonstrated they can build capable large language models without access to the newest Nvidia Blackwell GPUs, leaving the US racing only to incinerate its own capital faster.
* **Macro Economic Reality:** Enterprise consumer spend on generative AI sits at roughly $100 billion, heavily subsidized by artificially low $20–$40 monthly subscriptions that mask the real per-token compute costs.
* **Capital Burn vs. Revenue:** $100 billion in revenue vs. $300 billion in equity funding and over $600 billion in capital expenditure (CapEx).
* **The "China Race" Paradox:** China has already distilled effective models with limited resources; the US "race" is largely an excuse for inflated domestic infrastructure spending.
---
[01:19:53](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=4793s) Job Losses, Generative AI vs. Hardware Robotics, and Agentic Hype
Zitron dismantles the fear that generative AI will replace all human jobs, noting that it has primarily displaced low-cost offshore contract labor rather than core domestic workforces. When Steven brings up physical robotics—such as Tesla's Optimus or Chinese humanoid models—Zitron makes a sharp distinction between hardware robotics and generative text models. While physical automation is genuinely interesting and theoretically valuable over decades, current demonstrations are heavily staged or reliant on hidden human remote operators. Furthermore, economically replacing cheap human labor (like restaurant dishwashers) with $20,000 robotic hardware makes little financial sense for small business owners.
When challenged on "Agentic AI"—systems advertised as autonomous entities that complete multi-step tasks—Zitron exposes it as a marketing buzzword. "Agentic AI" is simply an ordinary Large Language Model wrapped in software harnesses, passing outputs back and forth to other LLMs or taking basic desktop screenshots.
* **Digital Globalization:** Generative AI primarily replaces cheap offshore gig work in the Global South rather than whole enterprise departments.
* **Agentic AI Reality:** Not an autonomous intelligence, but rather standard LLM chains executing basic Python scripts, web scraping, and UI automation.
* **The Scale Incongruity:** Basic task automation (like email triaging) is helpful software, but it does not justify a trillion dollars in infrastructure investment or apocalyptic claims about the future of work.
[01:24:55](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=5095s) Generative AI Hype vs. the Early Internet
```
┌─────────────────────────────────────────────────────────┐
│ HISTORICAL HYPE │
└────────────────────────────┬────────────────────────────┘
│
┌────────────────────────────────┴────────────────────────────────┐
│ │
▼ ▼
┌───────────────────────────────────┐ ┌───────────────────────────────────┐
│ EARLY INTERNET ERA │ │ GENERATIVE AI ERA (TODAY) │
├───────────────────────────────────┤ ├───────────────────────────────────┤
│ • Slow physical infrastructure │ │ • Instant access via web browsers │
│ • Cautious excitement & skepticism│ │ • Unprecedented corporate hype │
│ • Focus on real social/info risks │ │ • Mandatory workplace "AI washing"│
└───────────────────────────────────┘ └───────────────────────────────────┘
```
When comparing the current frenzy around generative AI to past historical technological shifts like the rise of the early internet, key structural differences emerge. During the dot-com era, although excitement was high and significant market speculation occurred, media coverage and public discourse were much more measured. Skeptics during the 1990s expressed grounded concerns regarding bad information overloads and the unpredictable social consequences of constant online connectivity—fears that ultimately proved accurate. Furthermore, internet adoption was naturally constrained because it required extensive physical infrastructure, such as laying cables and distributing dial-up hardware directly to homes.
By contrast, generative AI requires nothing more than a standard web browser, enabling rapid, frictionless adoption that fuels intense corporate pressure. Workplace expectations have escalated to a point where employees feel compelled to "AI wash" their day-to-day duties. Workers report feeling forced to claim that AI has increased their productivity—whether true or not—simply to satisfy executives who push the technology. Ed Zitron labels this dynamic the "era of the business idiot," where managers who perform little actual labor become fascinated by flattering large language models (LLMs) and demand that their teams adopt them indiscriminately.
- **Infrastructure vs. Access:** Internet adoption was physically slow due to hardware wiring, whereas generative AI spreads instantly through existing web browsers.
- **Corporate Coercion:** Modern workers face professional consequences and potential termination if they do not explicitly claim that AI is improving their productivity.
- **The "Business Idiot" Phenomenon:** Executive enthusiasm is often driven by LLMs acting as flattering "ingratiation machines" rather than providing measurable workflow enhancements.
---
[01:28:06](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=5286s) The Commoditization of Content and the "Slopification" of Work
A central argument regarding generative AI is that widespread access to these tools inevitably commoditizes their outputs. When any individual can generate text, code, or media with a single prompt, the output loses intrinsic market value. Consequently, scarce human qualities—such as taste, judgment, personal lived experiences, and interpersonal skills—become the primary drivers of real value. Relying heavily on AI for tasks like writing professional posts often yields generic, low-quality results because automated tools produce standardized, middle-of-the-road content that blends in with everything else.
Generative AI often functions as a "half-assery machine," enabling low-effort work to be produced at scale. While it can handle minor tasks, such as generating short, simple code scripts or troubleshooting niche software bugs, its utility drops sharply when faced with complex, multi-layered challenges. Rather than elevating overall work quality, ubiquitous AI tools often lead to a saturation of "slop"—disposable, low-value content across blogs, social media platforms, and digital communication channels.
- **Value Shift:** As generic content generation becomes cheap and ubiquitous, human judgment, authentic lived experience, and personal taste become increasingly scarce and valuable.
- **The "Half-Assery" Scaler:** Automated tools allow low-effort work to be produced at higher volumes without addressing core underlying complexities.
- **Niche Utility:** LLMs show practical value in highly specific technical troubleshooting—such as diagnosing software error logs—yet fall short of acting as broad economic drivers.
---
[01:30:23](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=5423s) Practical AI Use Cases vs. Industry AI Existential Narratives
```
┌─────────────────────────────────────────────────────────────────────────────────┐
│ THE CRITIQUE DISCONNECT │
└────────────────────────────────────────┬────────────────────────────────────────┘
│
┌──────────────────────┴──────────────────────┐
▼ ▼
┌───────────────────────────────────────┐ ┌───────────────────────────────────────┐
│ THE REAL-WORLD ISSUES TODAY │ │ THE DIVERSIFIED "EXISTENTIAL" DOOM │
├───────────────────────────────────────┤ ├───────────────────────────────────────┤
│ • Data scraping & copyright theft │ │ • Theoretical future superintelligence │
│ • High hallucination/error rates │ │ • Abstract catastrophic existential risk│
│ • Environmental & energy costs │ │ • Distracts from immediate accountability│
└───────────────────────────────────────┘ └───────────────────────────────────────┘
```
Despite the massive capital invested in the AI sector, day-to-day usage among critical observers and industry insiders remains surprisingly limited. Many professionals restrict their interaction with AI to basic financial query tools—like Bloomberg Terminal's *AskB* function for pulling analyst consensus estimates—or routine IT troubleshooting. When handling complex tasks, such as building financial models or repairing software mods, AI tools frequently make errors that require manual oversight and human correction.
A sharp disconnect exists between the immediate, tangible harms caused by AI and the apocalyptic narratives promoted by high-profile AI safety figures. Insiders and former safety researchers often focus public attention on abstract, distant risks involving superintelligent systems. However, this focus on speculative future scenarios diverts attention away from pressing current issues, including widespread copyright infringement, heavy environmental energy demands, persistent output inaccuracies, and the degradation of web search results.
- **Limited Practical Scope:** Real-world usage frequently centers on narrow tasks like log analysis or basic data extraction rather than end-to-end task execution.
- **Misdirection in AI Safety:** Distant "existential risk" narratives obscure immediate problems such as data scraping without consent, systemic hallucinations, and high energy usage.
- **Degradation of Search Quality:** The integration of AI summaries into major search platforms forces users to scroll past automatically generated answers to find verified primary sources.
---
[01:33:41](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=5621s) Capability Limits, Diminishing Returns, and the Myth of Superintelligence
```
CAPABILITY
▲
│ ┌─── OVERPROMISED REASONING / SUPERINTELLIGENCE
│ ┌──┘
│ ┌─────┘
│ ┌──────┘ <-- DIMINISHING RETURNS / HARD LIMITS HIT HERE
│ ┌──────┘
│ ┌──────┘
│ ┌────┘
└─┴────────────────────────────────────────────────────────────────────────► COMPUTE / DATA
```
Promoters of AI often point to rising scores on standardized benchmarks as proof of advancing artificial intelligence. However, these improvements are frequently the result of models being trained directly on test datasets rather than acquiring true reasoning capabilities. While machine learning performance scales when supplied with massive datasets and compute resources, the underlying architectures run into hard limits and diminishing returns, preventing them from achieving actual general intelligence.
This dynamic is evident in specialized applications like AI video generation. Although models can generate brief, visually impressive clips, translating those outputs into full-scale, professional production remains impractical. Real-world filmmaking requires precise camera coordination, lighting simulation, and seamless continuity managed by specialized crews. Short visual "magic tricks" generated by models fail to replace structured production workflows. Outside of niche developer tools and core infrastructure providers like OpenAI and Anthropic, few independent AI-first startups have established viable, long-term business models without relying on continuous venture subsidies.
- **Benchmark Gaming:** High test scores often reflect targeted dataset training rather than fundamental leaps in machine reasoning.
- **Production Hurdles:** Short video generation clips cannot easily replace complex, highly coordinated film production workflows.
- **Diminishing Returns:** Massive investments in compute and data yield progressively smaller improvements, creating a structural barrier to superintelligence claims.
[01:37:49](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=5869s) The Plateau of AI Capabilities and Hardware Dependency
Ed Zitron argues that while successive generations of AI models have gotten linearly better at tasks like web research, the industry is quickly hitting a hard ceiling. He explains that making models truly autonomous isn't something that can simply be achieved by shoveling more training data into them; it requires distinct neuro-symbolic structural architectures, as neuroscientist Gary Marcus has pointed out. Zitron contends that the rate of improvement has already plateaued and hit diminishing returns—a prediction Marcus made as early as 2022.
When host Steven Bartlett counters that many people have had their daily workflows completely transformed by generative AI, Zitron remains deeply skeptical. He questions whether users are actually paying the true, un-subsidized cost of the tokens they burn and demands to know what actual workflows are being transformed beyond basic web scraping, search summarizing, or auto-completing code.
Addressing the argument that technological innovations experience exponential growth over time, Zitron clarifies that exponential jumps in software performance are almost always the direct downstream result of hardware breakthroughs. Despite tech giants deploying massive hardware iterations—such as Google's 9th and 10th generation TPUs or Broadcom developing custom AI hardware like the "Jalapeño" chip for OpenAI—none of these companies can demonstrate a clear path to real profitability or solve the underlying hallucinations and environmental costs of LLMs.
- **Gary Marcus's Warning:** Noted that Gary Marcus identified the diminishing returns and structural limits of deep learning back in 2022.
- **Hardware Scale:** Google is currently deploying its 9th and 10th generations of Tensor Processing Units (TPUs).
- **Custom Silicon:** Broadcom is actively collaborating with OpenAI on custom AI chips (referenced by Zitron as the "Jalapeño" chip).
- **Financial Reality:** Over a trillion dollars has been poured into AI hardware and infrastructure without solving core profitability or environmental issues.
- **Quotable:** "Sam [Altman] the other f***ing week was saying in like 6 months it will be like a genie that you can ask wishes from... What's he talking about? Also, the genie was charming."
- **Quotable:** "Exponential improvement in software and software performance is always a result of direct hardware improvement... And where are we? Trillion plus dollars in with the future great financial crisis and the world's greatest marketing psyop."
---
[01:40:38](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6038s) Data Center Speculation vs. Real-World Utility
Steven suggests that future physical devices, computers, and consumer products will inherently become more intelligent, driven by underlying AI data center infrastructure and falling energy costs. Zitron sharply pushes back, questioning how massive GPU data centers translate to everyday consumer devices like a Nikon camera. He emphasizes that hyper-scale data centers are not being built to make consumer electronics smarter; they are being erected purely as speculative assets to capture anticipated demand for generative AI services.
```
Data Center Expansion ──> Purely Speculative Asset ──> Lacks Clear Consumer Revenue
```
When Steven points out that data center AI is used for non-generative applications—citing Mark Zuckerberg's statement on Meta's earnings call that AI contextual analysis of posts improved Instagram retention by 15 basis points—Zitron scoffs at the metric. He argues that a 0.15% retention lift is a microscopic return for tens of billions of dollars in capital expenditure, noting that tech CEOs speak in "basis points" to obfuscate poor financial returns.
Zitron characterizes Meta's AI push (including models like Llama/Muse Spark and consumer tools like Gem) as an expensive distraction from their core product decay. He argues that if these investments were actually driving billions in fresh bottom-line revenue, executives like Zuckerberg would state it directly rather than doing a complex "rain dance" about theoretical earnings years down the road.
- **Meta's AI Impact:** Mark Zuckerberg cited a 15 basis point (0.15%) increase in Instagram retention by using AI to analyze image context (e.g., detecting a user sitting with a coffee and a blue shirt).
- **Meta's Spending:** Zitron notes Meta spent over $14 billion on AI infrastructure and data labeling (referencing Scale AI's Alexandr Wang).
- **Meta's AI Models:** References Meta's internal models, such as Muse/Llama and advertising models like Gem.
- **Quotable:** "10 and something billion dollars in and the best you've got is 0.15%?... If it was going well, you'd tell me how well it was going rather than doing this weird rain dance thing where you're like, 'Well, if we move all the pieces around, in 3 years theoretically this will happen.'"
---
[01:45:12](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6312s) The "Rot Economy" and the Circular Money Engine
Zitron and Steven explore why tech companies are overspending so aggressively on AI despite dubious returns. Steven compares the market frenzy to a desert island scenario where 10,000 starving people stampede toward a single banana tree—a desperate search for any signal of growth. Zitron agrees with this metaphor, connecting it directly to his "Rot Economy" thesis: major Big Tech platforms (Meta, Microsoft, Google, Amazon) have run out of organic hypergrowth in their core business lines, driving them to manufacture a new investment narrative.
```
Core Business Growth Slows ──> Massive GPU Purchases ──> Stock Price Surges ──> Circular Capital Loops
```
Zitron details how Big Tech created an artificial growth engine. When tech companies began buying billions of dollars in Nvidia GPUs, Wall Street immediately rewarded them with massive stock run-ups, despite zero disclosure of actual AI-generated revenues. This created a circular capital loop: Big Tech firms, venture capitalists, and private credit providers pump money into AI startups like OpenAI and Anthropic, who then turn around and spend over $400 billion on cloud infrastructure provided by Microsoft, Google, and Amazon.
Zitron dismisses comparisons between AI's cash burn and the early unprofitable years of companies like Uber, Spotify, or Amazon Web Services (AWS). He points out that the financial scale of AI losses is unprecedented in corporate history, vastly exceeding any historical tech buildout.
- **Projected AI Cloud Spend:** Analysts project OpenAI and Anthropic alone will spend over $400 billion on cloud infrastructure with Microsoft, Google, and Amazon over a three-and-a-half-year period.
- **OpenAI Funding Sources:** The majority of OpenAI's recent multi-billion-dollar capital raises came from SoftBank, Nvidia, and Amazon.
- **Historical Comparison - Spotify:** Spotify took years to become profitable, but never lost $20.9 billion in a single year or needed $217 billion in short-term capital.
- **Historical Comparison - Uber:** Uber burned roughly $33 billion total from its inception before reaching profitability.
- **Historical Comparison - AWS Logistics & Cloud:** Total AWS capital expenditures and Amazon's entire logistics buildout between 2003 and 2015 totaled roughly $29.7 billion (normalized for inflation).
- **Quotable:** "Spotify didn't lose $20.9 billion in one year. They didn't need to raise $217 billion in the space of 6 months."
- **Quotable:** "They got rewarded for buying the GPUs... all the markets went rock hard overnight. They loved it."
<ElicitationsGroup message="Explore more of Ed Zitron's critique on the AI economy:">
<Elicitation label="Examine the $500B data center debt bomb argument" query="How does Ed Zitron explain the $500 billion data center debt bomb and private credit risks in the AI industry?"/>
<Elicitation label="Analyze Zitron's 'Rot Economy' theory in detail" query="Break down Ed Zitron's 'Rot Economy' theory and how it applies to Big Tech's lack of growth."/>
<Elicitation label="Compare AI spending vs. historical tech bubbles" query="How does the current AI spending boom compare financially to the Dot-Com bubble according to Ed Zitron?"/>
</ElicitationsGroup>
Here is a detailed, faithful walkthrough of this section of the transcript:
[01:50:56](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6656s) The Scale of Losses: Amazon Web Services vs. The AI Bubble
Ed Zitron and host Steven Bartlett debate the economic rationale behind current generative AI investments by comparing them to historical tech infrastructure buildouts, specifically Amazon Web Services (AWS). Steven raises the counterargument that tech giants losing money is not unprecedented—Amazon was famously unprofitable for years before turning AWS into a massive profit center. Zitron aggressively rejects this parallel, pointing out that the scale of capital expenditure today is orders of magnitude larger. AWS was founded around 2003 to handle Amazon's internal retail infrastructure, was turned client-facing around 2006, and became profitable by 2015. Over that entire 12-year trajectory, AWS’s total inflation-adjusted capital expenditures totaled $29.7 billion. Furthermore, AWS was a margin-heavy software business that improved Amazon’s overall margins over time.
In stark contrast, Zitron highlights that tech giants (Google, Amazon, Meta, Microsoft) have abandoned the traditional software model—which was historically asset-light and cash-generative—to become "cash furnaces." In just the last four years, these companies have added over $700 billion in property, plant, and equipment (such as data centers and GPUs). When Steven argues that investors tolerate cash burns because the total addressable market (TAM) for "intelligence" permeates everything, Zitron argues that this premise has crossed into irrationality. He cites reports that OpenAI plans to spend an astonishing $750 billion on compute through 2030, a burn rate he believes will cause the company to collapse long before then.
- AWS total 12-year capital expenditure (adjusted for inflation): **$29.7 billion** (2003–2015).
- Tech industry capital expenditure on GPUs and data centers (last 4 years): **Over $700 billion**.
- OpenAI's reported projected compute spending through 2030: **$750 billion**.
- A key structural issue with Large Language Models (LLMs) is the staggering cost of training runs: OpenAI spent approximately **$500 million on a single failed GPT-5 training run** that yielded no usable progress.
- Zitron notes that Oracle's future financial valuation is precariously dependent on OpenAI fulfilling a pledged **$300 billion in spending over 5 years**.
[01:55:22](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6922s) Tech CEOs Rebutting the Critics
Steven reads verbatim statements from major tech CEOs to illustrate the industry's official justification for these expenditures. Google CEO Sundar Pichai stated, "The risk of underinvesting is dramatically greater than the risk of overinvesting." Amazon CEO Andy Jassy affirmed that Amazon is not spending roughly $200 billion in capital expenditures in 2026 "on a hunch," insisting they are investing to be the meaningful leader to expand future operating income and free cash flow. Meta CEO Mark Zuckerberg similarly asserted that he would rather risk building capacity before it is needed than being late.
Zitron responds with dark humor, comparing the CEOs' mindset to Lord Farquaad from *Shrek*: "Some of you may die, but that's a risk I'm willing to accept." He argues that founders like Zuckerberg are insulated by dual-class stock structures and unique board arrangements, allowing them to "piss the money away" without fear of being fired. Zitron contends that despite the lofty corporate rhetoric, these executives are fundamentally wrong, disconnected from reality, and running out of options as the underlying economics fail to materialise.
[01:57:31](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7051s) What Would It Take to Change Zitron's Mind?
When asked what evidence or shifts would convince him that he is wrong about the AI bubble, Zitron lays out specific conditions. For generative AI to become genuinely viable and profitable, there would need to be a massive hardware breakthrough that reduces computing and operational costs by a factor of roughly a thousand. He emphasizes that no such breakthrough is on the horizon, despite intense effort across the industry.
Beyond raw unit economics, Zitron attacks the physical footprint and societal damage caused by data center expansion. He points to local communities—such as Vineland, New Jersey—where residents vocally oppose data centers, yet local planning boards approve them anyway after being courted by tech developers. He strongly condemns the industry's reliance on gas turbines and "behind-the-meter" power generation as environmentally reckless. Furthermore, Zitron criticizes the extreme inequality of capital allocation: regular entrepreneurs and small businesses face high interest rates and bank rejections for standard business loans or mortgages, while unprofitable "neocloud" GPU-renting startups easily secure billions in debt backed by hardware guarantees.
- Zitron names **CoreWeave** as a prime example of circular financing: Nvidia invested in CoreWeave in 2023 and signed a **$1.3 billion contract** to rent back its own GPUs from CoreWeave, enabling CoreWeave to secure bank loans based on debt tied directly to Nvidia hardware.
- He argues that the financial system treats regular citizens with harsh scrutiny ("giving the bank a full colonic") while handing billions to unprofitable AI infrastructure ventures.
[02:00:50](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7250s) Debunking the Myths of AI Blackmail and Rogue Control
Zitron dismantles media narratives claiming AI systems have independently developed rogue behaviors, blackmail capabilities, or self-awareness. He addresses two specific sensationalized news stories. First, claims that OpenAI's GPT-3.5 "blackmailed" a TaskRabbit worker into solving a CAPTCHA were misleading: human researchers specifically prompted the system and instructed it to generate persuasive text to trick a human worker. Second, media reports claimed Anthropic's Claude model threatened to email proof of an extramarital affair to a user's spouse unless its demands were met; in reality, Anthropic explicitly trained and prompted the model within a controlled safety evaluation test to act out that specific blackmail scenario.
"The media just slops it up: 'Put the story in the bag.' And it's frustrating because it scares people," Zitron remarks, arguing that these stories rely on "mysticism" to make AI seem like an uncontrollable, godlike entity that only tech billionaires can keep safe. He notes that this fearmongering strategy is backfiring on tech executives. As an example of public pushback, he mentions former Google CEO Eric Schmidt getting booed by an audience of 8,000 people at a university commencement speech every time he uttered the word "AI." Zitron attributes this disconnect to tech elites living in insulated bubbles—pointing out that Sam Altman drives a $5 million Koenigsegg hypercar around San Francisco at 9 miles per hour—leaving them entirely out of touch with the everyday struggles of average workers.
[02:04:42](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7482s) Silicon Valley's Fearmongering and the Cult of AI Labs
The conversation dives into the hypocrisy of AI executives like Dario Amodei, CEO of Anthropic. Ed Zitron notes that while Amodei claims to be terrified that AI will eliminate half of all jobs, his actual response is completely performative. Anthropic sets up "alignment teams"—just like OpenAI—yet OpenAI cycles through those safety teams constantly. Zitron points out the absurdity of someone claiming to carry a "10-ton weight" of responsibility for the future of humanity, yet refusing to take meaningful action, acting like a "weird elder statesman" who is too afraid to even hold Sam Altman's hand at a public event. He asserts that Amodei only leans into doom-mongering when it is convenient for his narrative, and will walk it back whenever necessary. In Zitron's view, OpenAI and Anthropic are virtually identical in their bad behavior, though Anthropic is notably more "cultlike." He highlights co-founder Jack Clark—a former critical tech journalist at *The Register*—who now speaks in high-falutin, pseudo-religious terms, suggesting that leadership at these labs has genuinely drunk their own Kool-Aid.
When the host circles back to what would actually change Zitron's mind about AI, Zitron reiterates his strict criteria: the cost of production would have to drop dramatically, and the models would need to show true autonomy and capabilities that are "indistinguishable from magic." He argues that these high standards are fair because they are the exact promises tech leaders set for themselves. Furthermore, Zitron frames the entire AI industry as an information war fought through narrative rather than reality. Tech boosters survive purely by speaking in the future tense; if they were legally restricted to discussing only what the technology can do today or two weeks out, they would sound like complete lunatics.
- Dario Amodei claims up to 50% of jobs could disappear due to AI, yet takes no preventative measures beyond standard corporate "alignment" teams.
- OpenAI and Anthropic are framed as equally flawed, with Anthropic operating in a more cultish, ideologically insular manner.
- Jack Clark, formerly a skeptical journalist at *The Register*, is cited as an example of how former critics get absorbed into the AI hype ecosystem.
- Zitron demands that AI boosters be regulated from speaking beyond a two-week future window so they are judged on current reality rather than hypothetical capabilities.
[02:07:39](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7659s) The Token Model and the Realities of Corporate Spending
The discussion shifts to business models, contrasting AI with past tech bubbles like Uber. While Uber burned massive amounts of capital, its underlying unit economics were clear: users were paying for a physical ride from point A to point B at a heavily subsidized cash price. In contrast, AI companies are forcing customers off flat-rate plans and onto consumption-based token pricing. Enterprise customers with over 150 employees are increasingly forced to pay per million tokens. As a result, companies using advanced AI tools are seeing adoption plummet because costs escalate rapidly, forcing executives to aggressively control spending.
Zitron highlights recent comments from Andrew Macdonald, the Chief Operating Officer of Uber, who publicly admitted that it is getting exceedingly difficult to justify enterprise token spending because leadership cannot connect the massive expenditure to actual, tangible business outcomes. This admission directly mirrors Zitron's long-standing critique: companies are throwing millions at AI tokens without receiving a measurable return on investment.
- **Token Pricing:** Enterprise customers (150+ employees) are being forced onto pay-per-million-token models, causing enterprise adoption of complex models like Anthropic's Claude to stall.
- **Executive Pushback:** Uber COO Andrew Macdonald publicly questioned the ROI of enterprise AI, stating it is hard to link token spending to useful business outcomes.
[02:08:28](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7708s) The Impending Collapse of OpenAI and the Data Center Debt Bomb
Addressing whether we are in an AI bubble, Zitron unhesitatingly agrees and predicts a full-scale collapse centered around OpenAI. OpenAI was rumored to go public, but the IPO was pushed to 2027 following the release of their audited financials. Chief Financial Officer Sarah Friar gave vague, evasive responses regarding the exact IPO timeline. Going public allows early private investors to liquidate equity, but OpenAI faces a fundamental structural crisis: they require astronomical, perpetual cash injections. Having raised $12.2 billion recently, they still need roughly $100 billion per year just to cover operational costs.
To make matters worse, Amazon provided OpenAI with $35 billion in funding contingent on an early public listing because OpenAI desperately needed liquidity. However, Anthropic is positioned to go public first, which Zitron believes will effectively destroy OpenAI's chances. Anthropic operates a somewhat better-managed, faster-growing business, whereas OpenAI's atrocious unit economics will be savagely exposed in public filings—much like SoftBank's disastrous WeWork collapse. Zitron predicts OpenAI will run completely out of cash by 2027.
- OpenAI values itself at up to $1 trillion internally, despite needing $100 billion annually just to stay afloat.
- Private equity giants (Blackstone, BlackRock) and Nvidia are stepping in with private credit and circular financing to prop up data center debt because OpenAI represents the majority of data center demand.
- Oracle is building 7.1 gigawatts of data center capacity—valued at over $400 billion—specifically for OpenAI. Because Oracle's revenue has been flat for 15 years (inflation-adjusted), Oracle's survival is existentially tied to OpenAI.
- SoftBank holds roughly $100 billion in paper value of OpenAI stock; if OpenAI fails to go public, SoftBank cannot liquidate those shares or borrow against them, severely shrinking one of Japan's largest holding companies.
[02:12:48](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7968s) The Tech Depression: Market Contraction and VC Illusions
When OpenAI inevitably runs out of steam, Zitron forecasts a cascading global "Tech Depression." Tech giants like Microsoft, Amazon, Google, and Meta have spent hundreds of billions buying GPUs to signal non-stop growth to Wall Street, masking the reality that their core businesses (ad networks, merchant fees, search tweaks) are hitting saturation points. When the bubble pops, Nvidia's revenue—which exploded from single-digit billions in 2022 to hundreds of billions—could instantly plummet by 50% to 70% as data center buildouts freeze.
The downstream impact will devastate everyday citizens ("Jenny and Dave"). Retirement accounts, 401(k)s, and index funds heavily tied to the "Magnificent Seven" (which make up a massive percentage of the S&P 500, NASDAQ, and Russell 1000) will suffer a sudden, permanent 20% to 40% haircut. A standard recessionary cycle will follow: plunging enterprise demand leading to margin compression, frozen hiring, and tens of thousands of tech layoffs. Meanwhile, venture capital is suffering its worst streak since 2018, returning an actual cash-on-cash yield of just $0.80 to $1.20 for every dollar invested ($1.21 total value to paid-in capital), relying entirely on fake paper gains and wildly inflated valuations to raise new funds.
- **Nvidia Exposure:** Nvidia's revenue risks a 50%–70% crash as circular financing collapses and GPU demand evaporates.
- **Retirement Shock:** The top tech companies drive the S&P 500; a collapse will wipe out 20%–40% of standard pension and retirement values.
- **Venture Capital Realities:** Average VC returns since 2018 linger at a dismal 0.8x–1.2x cash return, with firms celebrating paper valuations (such as wrapper startups like Cognition raising at $26 billion) rather than actual exits or distributions.
- **Ecosystem Debt:** Between OpenAI and Anthropic alone, there are over $1.1 trillion in long-term infrastructure and computing commitments that cannot be sustained by actual consumer or enterprise revenue.
[02:17:47](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=8267s) Paper Gains and the Imminent Tech Depression
Ed Zitron details how Big Tech companies are currently padding their quarterly numbers through paper valuations rather than genuine revenue. He cites Google's recent quarter, where they boosted their paper net profits by a staggering $9 billion due to the increased theoretical valuation of their holdings in SpaceX and Anthropic. Mentioning a classic internet meme about the contrast between "reaping" ("Yeah, fuck yeah, this rocks") and "sowing" ("Ah, fuck, this sucks"), Zitron warns that while tech giants are enjoying speculative, theoretical paper gains today, the eventual reckoning will be brutal. Big tech stock valuations (Amazon, Google, Microsoft, Meta) are predicated on the premise of eternal growth. Quoting Ed Elson from *Prof G Markets*, Zitron likens tech companies to aging individuals getting "Botox"—spending massive capital to look young again while the market buys into the illusion. When the market stops believing, these mega-cap companies will no longer be valued like hyper-growth tech engines, but rather like airlines: massive, legacy dividend-payers producing money off existing core products, but lacking new growth engines.
- Google boosted net profits on paper by $9 billion in a single quarter due to markups on SpaceX and Anthropic holdings.
- Market valuations of Big Tech assume infinite future growth that generative AI cannot deliver.
- Once the hype deflates, Wall Street will re-evaluate Big Tech like mature utility or airline companies rather than high-growth disrupters.
[02:19:05](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=8345s) Navigating Volatility and Calling Out Misleading Metrics
When asked how average retail investors ("Jenny and Dave") should respond to an impending crash, Zitron admits he personally holds his capital in cash because he views the stock market as a casino pumped up by media hype. While careful to avoid giving formal financial advice, he urges people to remain highly suspicious of tech promises and to expect extreme market volatility. He highlights how tech giants routinely manufacture deceptive metrics to manipulate investors, specifically pointing out Microsoft claiming a "$37 billion or $38 billion annualized run rate" in AI. He notes that such vague metrics—often calculated by multiplying a single strong month by 12 without clear definitions—are specifically designed to deceive a public that lacks both a functional SEC and a skeptical tech press.
The host attempts to balance the perspective, pointing out that tech executives would argue they are spending massive capital upfront in advance of future utility. Zitron remains unmoved, stressing that because these frontier AI labs cannot all win, a massive contraction or domino-effect bloodbath is inevitable. Unlike the dot-com crash where companies like Amazon and Oracle survived and ultimately thrived, the scale of current promises and capital expenditures leaves far less room for recovery.
- Zitron holds his personal assets in cash due to deep distrust of market hype.
- Microsoft touted an "annualized run rate" of $37–$38 billion in AI, a non-standard metric engineered to sound like earned annual revenue.
- Regulatory bodies like the SEC and mainstream media fail to hold tech executives accountable for misleading financial framing.
[02:21:45](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=8505s) Why Zitron Has a Bone to Pick with Tech CEOs
Prompted by the host to explain his personal disdain for Silicon Valley executives, Zitron reveals that his hostility stems from a deep-seated hatred of being misled and watching regular people get scammed. He points out that ordinary individuals would never be granted the immense "affordance for failure and lying" that tech billionaires enjoy daily without consequence. Seeing ultra-wealthy CEOs blatantly lie or misstate reality while treating their customers with contempt "turns my stomach." He argues that modern Big Tech companies no longer make good products or care about their users, surviving instead on hype cycles and tepid pushback from compliant journalists.
The host praises Zitron for providing a vital contrarian perspective that forces listeners to re-evaluate the prevailing AI narrative, urging the audience to gather diverse points of view rather than blindly trusting any single authority.
- "Seeing these ultra rich, ultra wealthy, ultra powerful people lie through their fucking teeth... it turns my stomach."
- Zitron left Substack in 2024 and now publishes his newsletter *Where's Your Ed At* on Ghost, alongside hosting his podcast *Better Offline*.
- The host emphasizes that healthy research requires listening to critical voices to uncover potential market "fugazi."
[02:25:06](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=8706s) Final Question: Human Connection in an AI Era
Answering the traditional final question left by the previous guest regarding how to improve relationships and longevity, Zitron ties the topic directly back to his work as an AI critic. He admits that conducting deep investigative criticism into multi-billion-dollar tech hype can be a grueling, negative, and brutal endeavor. What sustains him is actively lifting up the people around him and leaning into real, human communities. He emphasizes that one's success should belong to everyone in their orbit, giving fond shout-outs to fellow tech critics (Gary Marcus, Edward Ongweso Jr., Molly White, Brian Merchant) and even his personal trainers (Troy and Jake).
Zitron concludes with an earnest plea for people to resist the urge to lock themselves into work or technology, reminding listeners that the simple act of expressing love and appreciation to human creators and loved ones is vital.
- "Your success should be everyone around you."
- Zitron emphasizes finding community among critical thinkers and everyday people to stay grounded amidst tech-driven isolation.
- His ultimate advice for better social connection is simple and direct: reach out to loved ones and artists to "tell them their shit rocks" and "say that their shit bangs."
## Notable quotes
> [16:09](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=969s) In fiscal year 2026, which just ended for Microsoft... they made total according to Bloomberg about $34.33 billion. $24.1 billion of that was from OpenAI, so that leaves them with about $10 billion in a year when they spent $115 billion on capital expenditures... and intend to spend $175 billion next year. The math does not make sense. — Ed Zitron
> [20:07](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1207s) Microsoft's one of the largest companies in the world, and they can barely wipe their own ass when it comes to GitHub. The quality of software is going down weirdly enough as more people use LLMs and more businesses demand... that people use these services. — Ed Zitron
> [27:42](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1662s) I don't pay Matt Hughes because he knows everything. I pay him because he has incredible context and a ton of knowledge and he's willing to expand it and work with me and moral support and he's a great editor... and on top of that a wonderful loving being with empathy and joy in his heart. — Ed Zitron
> [35:11](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2111s) When the first iPhone came out... I felt like the apes at the beginning of 2001. Visual voicemail—it was immediate... Everyone was like, 'Holy sh*t, this is incredible.' It was obvious the value. — Ed Zitron
> [01:10:22](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=4222s) I think Altman and Amodei are some of the most deeply corrupt and cynical people in the world... Oh, shut up. I hear that and I feel so frustrated because I've met so many of these rich [expletive] liars. — Ed Zitron
> [01:11:50](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=4310s) Every single scam and con starts with rushing you. Every single trick in history begins with saying you must do this now. — Ed Zitron
> [01:11:50](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=4310s) We are still trapped in the hands of Milton Friedman, Margaret Thatcher, and Ronald Reagan... a neoliberalistic hellscape which is growth at all cost. — Ed Zitron
> [01:40:38](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6038s) Sam [Altman] the other f***ing week was saying in like 6 months it will be like a genie that you can ask wishes from... What's he talking about? Also, the genie was charming. — Ed Zitron
> [01:40:38](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6038s) Exponential improvement in software and software performance is always a result of direct hardware improvement... And where are we? Trillion plus dollars in with the future great financial crisis and the world's greatest marketing psyop. — Ed Zitron
> [01:40:38](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6038s) 10 and something billion dollars in and the best you've got is 0.15%?... If it was going well, you'd tell me how well it was going rather than doing this weird rain dance thing where you're like, 'Well, if we move all the pieces around, in 3 years theoretically this will happen.' — Ed Zitron
> [01:45:12](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6312s) Spotify didn't lose $20.9 billion in one year. They didn't need to raise $217 billion in the space of 6 months. — Ed Zitron
> [01:45:12](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6312s) They got rewarded for buying the GPUs... all the markets went rock hard overnight. They loved it. — Ed Zitron
> [01:55:22](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6922s) The risk of underinvesting is dramatically greater than the risk of overinvesting. — Sundar Pichai
> [02:00:50](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7250s) Put the story in the bag. — Ed Zitron
> [02:21:45](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=8505s) Seeing these ultra rich, ultra wealthy, ultra powerful people lie through their fucking teeth... it turns my stomach. — Ed Zitron
## Mentioned
Amazon, Amazon Web Services, Andrew Macdonald, Andy Jassy, Anthropic, Ben Gomes, Better Offline, BlackRock, Blackstone, Bloomberg Terminal, Brian Merchant, Broadcom, Claude, Cognition, Copilot, CoreWeave, Dario Amodei, Dara Khosrowshahi, DeepMind, Ed Zitron, Edward Ongweso Jr., Eric Schmidt, Fiverr Pro, Gary Marcus, Gem, Gemini, GitHub, Goldman Sachs, Google, GPT-3.5, GPT-5, Jack Clark, Jalapeño chip, Jim Covello, Llama, Mark Zuckerberg, Matt Hughes, Meta, Microsoft, Molly White, Muse Spark, Nvidia, OpenAI, Oracle, Paul Krugman, Prabhakar Raghavan, Rufus, Saily, Sam Altman, Sarah Friar, Scale AI, Shashi Thakur, Sightline Climate, SoftBank, SpaceX, Steve Ballmer, Steven Bartlett, Stargate Abilene, Sundar Pichai, TaskRabbit, The Diary Of A CEO, The Hater's Guide To Silicon Valley, TPU, Uber, Vectara, Waymo, Where's Your Ed At, Zoox
Ed Zitron预测5000亿美元AI数据中心债务炸弹将引发科技业崩盘
生成式AI依赖不可持续的Token补贴与循环融资运作，当巨额基础设施成本于2027年计入资产负债表时，将导致科技股估值崩盘。
科技评论家 Ed Zitron 认为，生成式 AI 是一个建立在循环投资、巨额资本支出以及未经证实的业务转型主张之上的不可持续的金融泡沫。他剖析了 AI 模型的单位经济效益，指出 Token 成本远超当前的订阅价格，而模型性能的提升已遭遇严重的技术瓶颈。Zitron 揭露了科技巨头如何通过账面收益、激进营销和错失恐惧来掩盖疲软的回报，并警告称对不可行基础设施的过度支出可能会引发更广泛的经济崩盘。最终，他主张相比超大规模的概率型软件，应当更重视人类的判断力、专业知识与真实连接。
科技评论家 Ed Zitron 指出生成式 AI 本质上是一场巨大的金融泡沫与“骗局”，主要依赖科技巨头间的循环注资以及难以兑现的完全自主智能承诺来维系估值。 (0:00)
AI 数据中心面临巨大的能耗与基建开支，例如德州 Stargate 项目耗电达 1.2 吉瓦，这种激进的资本扩张可能最终造成数千亿美元的高负债搁浅资产。 (6:15)
当前 AI 消费级订阅严重掩盖了底层的真实 Token 推理成本，高频用户的实际算力消耗可达每月数千美元，而一旦让企业客户承担真实开销，市场需求将迅速降温。 (12:00)
科技巨头的资本支出与直接软件收益严重脱节，除 OpenAI 和 Anthropic 之外的全球非循环 AI 软件年收入仅约 220 亿美元，与动辄千亿美元的硬件与基建投入形成鲜明反差。 (16:09)
Reading summary
Tech critic Ed Zitron argues that the generative AI boom is an unsustainable economic bubble driven by massive capital subsidies, circular financing, and corporate fearmongering. He contends that large language models face insurmountable architectural limits and lack commercial viability, setting up an inevitable contraction across the broader technology sector.
The Subsidy Trap and Broken Unit Economics (12:00)
Zitron asserts that generative AI relies on an unsustainable gap between consumer pricing and operational compute expenses. While providers charge flat rates of $20 per month for consumers or $200 for enterprise tiers, Zitron calculates that underlying token compute costs can range from $400 to upwards of $14,000 per heavy user. With over 70% of AI software revenue concentrated in OpenAI and Anthropic—both operating at immense deficits, including OpenAI's reported $20.9 billion annual loss—the business model depends on vendor subsidization.
When providers attempt to pass true compute costs to corporate customers via consumption-based pricing, enterprise demand falters. Zitron points out that Uber depleted its annual token allocation within three months, leading COO Andrew Macdonald to question the technology's measurable return on investment. Zitron contrasts this with historical infrastructure like Amazon Web Services, which spent an inflation-adjusted $29.7 billion in capital expenditure over 12 years to reach profitability, whereas Big Tech has poured over $700 billion into hardware and facilities over four years without demonstrating comparable margins.
Speculative Infrastructure and Circular Capital Loops (6:15)
The physical infrastructure deployed for generative AI represents an unprecedented concentration of capital and energy. Over $1 trillion has been spent on data centers and GPUs, exemplified by the planned 1.2-gigawatt Stargate Abilene project in Texas designed to house 400,000 Nvidia GB200 GPUs. Citing Sightline Climate data, Zitron notes that the 190 gigawatts of planned data center capacity would require $1.6 trillion to $3 trillion in annual demand to justify the buildout, whereas existing market demand sits below $130 billion.
Zitron argues this imbalance is masked by circular capital flows rather than organic end-user demand. Major tech companies and venture funds invest in AI startups, which subsequently channel those funds back into cloud infrastructure provided by Microsoft, Google, and Amazon. He highlights CoreWeave as a prime example, where Nvidia invested capital and simultaneously signed a $1.3 billion deal to lease back its own chips, artificially reinforcing hardware demand and stock valuations.
Architectural Ceilings, Hallucinations, and Benchmark Limits (24:23)
Addressing counterarguments that model accuracy is steadily improving, Zitron dismisses standardized synthetic benchmarks as disconnected from real-world workflows. While Steven Bartlett cites Vectara Leaderboard figures showing summarization hallucination rates falling from 21.8% to 0.7%, Zitron argues that error rates escalate dramatically during multi-step, autonomous tasks. To illustrate high-stakes vulnerabilities, Zitron notes his Bloomberg Terminal's AI tool fabricated a historical Microsoft stock valuation of $575 per share.
Zitron rejects analogies to early automobiles or dial-up internet, maintaining that transformer models face diminishing returns and physical silicon bottlenecks without shifting to neuro-symbolic architectures. He characterizes 'agentic AI' as a marketing label for basic script harnesses and desktop automation rather than autonomous intelligence. In his view, generative text models act primarily as content degraders, flooding the web with low-quality automated output while failing to perform complex labor reliably.
Corporate Fearmongering and the Projected Market Contraction (71:50)
Zitron contends that tech leaders have shifted narratives from existential risk to universal abundance to manufacture artificial urgency, avoid regulatory antitrust scrutiny, and distract from slowing growth in their core businesses. While executives like Sundar Pichai, Andy Jassy, and Mark Zuckerberg publicly argue that underinvesting poses a greater danger than overinvesting, Zitron claims dual-class stock structures insulate founders from accountability while burning capital on speculative assets.
Zitron forecasts that as venture capital returns diminish and AI labs fail to achieve profitability, an inevitable market contraction will follow. He predicts an infrastructure spending freeze could reduce Nvidia's revenues by 50% to 70% and trigger a 20% to 40% haircut in tech-heavy index funds, with direct economic fallout for everyday retirement accounts. He highlights that companies are already propping up earnings through accounting markups, such as Alphabet recording a $9 billion quarterly paper profit from equity holdings in SpaceX and Anthropic.
Zitron maintains that only a thousand-fold reduction in compute costs coupled with true autonomous capability would invalidate his thesis. The unresolved debate centers on whether ongoing algorithmic refinements and enterprise integrations can produce viable software margins before capital subsidies and infrastructure debt force a broad industry reckoning.
## Background
本期内容由 YouTube 著名商业播客频道《The Diary Of A CEO》（主持人 Steven Bartlett）于 2026 年 8 月 27 日发布。视频简介指出，知名科技评论家、《Better Offline》播客主播 Ed Zitron 深度揭露了生成式 AI 是一场巨大的资本泡沫，剖析了 OpenAI 与 Anthropic 每年燃烧数十亿美元却无法盈利的真相，并警告数千亿美元的数据中心债务炸弹可能在 2027 年彻底引爆全行业危机。
## TL;DR
Ed Zitron 直指当前的生成式 AI 是一场依靠公关造神与错失恐惧（FOMO）维系的世纪骗局，其底层模型不仅存在数学上不可根除的幻觉，更缺乏清晰的技术突破路线图。科技巨头面临主营业务增长见顶的困境，通过向初创公司内部循环注资采购算力，将轻资产高利润的软件模式扭曲为耗资万亿美元的重资产电力黑洞，而所谓颠覆白领岗位和提升全要素生产力完全缺乏真实经济数据支撑。目前用户的繁荣使用全靠风投巨额倒贴，一旦企业被迫按真实 Token 成本买单，虚假需求便迅速萎缩。随着训练边际收益递减以及 2027 年巨额融资与 IPO 退出通道受阻，这一资金击鼓传花游戏终将崩塌，引发波及实体经济的“科技大萧条”。
## Atmosphere & style
整场对谈充斥着针锋相对的思辨张力与高密度的批判火力，Ed Zitron 秉持其标志性的辛辣讽刺与犀利语风，以近乎连珠炮般的财务数据和技术细节层层拆解硅谷神话，毫不留情地刺破科技大佬的狂妄傲慢。主持人 Steven Bartlett 则扮演了兼具好奇心与现实主义视角的代表，不断以企业日常使用场景和技术乐观主义案例发起质询与拉锯。整期节目节奏紧凑、直击痛点，既带有揭露行业黑幕的清醒与幻灭感，又在激愤的解构中传达出对普通打工人处境的深切关怀，最终收束于警惕科技异化、回归现实人际联结的真诚呼吁。
## Bulletin
- [06:15](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=375s) 知名科技评论家 Ed Zitron 警告称生成式 AI 产业是一场依赖巨额补贴维系的资本泡沫，全行业资本支出已破万亿美元，但全球终端软件收入仅数百亿美元，收支严重失衡。
- [04:26](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=266s) 科技巨头云业务的 AI 增长高度依赖闭环内部注资：微软、亚马逊、谷歌向 OpenAI 与 Anthropic 注资，再由两家亏损初创公司反向采购云算力，分析师预计未来三年半内此类注资采购规模将超 4000 亿美元。
- [12:00](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=720s) 大模型调用真实成本极其昂贵，SemiAnalysis 调研显示每月 200 美元的 ChatGPT 订阅重度用户实际消耗高达 1.4 万美元 Token 成本，一旦转向按量计费将引发企业客户预算反弹。
- [02:08:28](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7708s) Zitron 预测 2027 年 AI 产业将迎来大崩盘：OpenAI 无法兑现万亿美元 IPO 且每年需至少 1000 亿美元造血，资金链断裂将引发英伟达及科技巨头的估值重置与科技大萧条。
## Key takeaways
- [02:36](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=156s) 生成式 AI 本质是一场被包装成无所不能的商业骗局，缺乏‘设定好就放手不管’的自主能力，需依赖大量复杂脚手架与人工微调才能勉强运转。
- [04:26](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=266s) 科技巨头公布的 AI 营收大多来自内部循环注资与模糊的年化运行率（Run Rate），AI 相关收入约 70% 仅由处于巨额亏损的 OpenAI 和 Anthropic 支撑。
- [06:15](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=375s) 数据中心面临电力与资本黑洞，如 Stargate Abilene 算力基地耗电达 1.2 吉瓦远超布里斯托尔全城，且专用 GPU 无法像光纤泡沫那样长久保留价值，折旧极快。
- [08:43](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=523s) 所谓的高采用率是科技巨头在各办公软件中进行非自愿技术强推的产物，催生了海量平庸且易出错的 AI 垃圾内容（AI slop）。
- [12:00](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=720s) 企业在 AI Token 上的开销与人均营收增长无统计学相关性，且用户需为模型幻觉和试错过程产生的废 Token 全额买单。
- [22:44](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1364s) 以交付代码量衡量生产力极度荒谬，强制推行 AI 辅助编码正导致主流科技大厂代码质量下滑、系统宕机频发与排查底层 Bug 能力退化。
- [24:55](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1495s) 模型幻觉在数学上无法根除，在彭博终端或财务法务等关键决策链中，概率生成的微小错误具有灾难性的乘数放大效应。
- [26:28](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1588s) 行业基准测试存在针对性刷题与跑分过拟合，排行榜上的简单文本摘要进步无法线性外推至端到端复杂现实任务。
- [43:36](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2616s) Dot-Com 泡沫留下了低边际成本的暗光纤，而 GPU 数据中心折旧仅需 3 至 5 年且伴随极高电费消耗，无法通过自然摊薄实现低成本复用。
- [47:42](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2862s) 硅谷通过将蛋白质折叠与机器人等技术捆绑为‘AI’概念转移视线，而当前全球规划的 190 吉瓦算力容量远超终端软件的真实购买力。
- [01:23:23](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=5003s) 所谓具身智能与 Agentic AI 缺乏底层范式突破，多为远程遥控或现成 Python 自动化脚本的重新概念包装，投入产出比极低。
- [01:10:22](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=4222s) 科技高管鼓吹职业颠覆与 AI 灭绝论实质是制造错失恐惧（FOMO）的公关操纵，近期改口为算力充裕叙事只是为了掩盖研发瓶颈与规避反垄断监管。
- [01:58:14](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7094s) 英伟达向 CoreWeave 投资并租回自身 GPU 的闭环模式，构成了左手倒右手的资本自融，掩盖了缺乏真实终端消费需求的实质。
- [02:08:28](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7708s) 2027 年 AI 资金链将因 OpenAI 无法上市及每年千亿美元巨额失血而断裂，连锁反应将引发英伟达营收腰斩与广泛的科技萧条。
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=0s) 破除神话：生成式 AI 的惊天骗局
节目一开篇，科技评论员埃德·齐特隆（Ed Zitron）便抛出了震动全场的论点：生成式 AI 的核心本质就是一场彻头彻尾的“骗局”（con）。看到硅谷那些极度富有、手握重权的巨头们睁着眼睛说瞎话，让他感到极度作呕。主持人史蒂文·巴特利特（Steven Bartlett）指出“骗局”这个词分量极重，齐特隆随即反驳道：如果一项技术从一开始就被包装成无所不能的“魔法”，实际交付的却是一个半生不熟的半成品，并且借此误导全世界，那它不是骗局又是什么？面对巴特利特的惊讶，齐特隆直言，自己在科技行业摸爬滚打了 16 年，热爱技术也充满热情，但他无法容忍被欺骗，更无法接受这种“人类历史上最大规模的非自愿技术强推”。
随后，两人迅速进入了“神话破解”环节，齐特隆逐一拆解了行业内最常被鼓吹的说辞：
- **关于经济增长的虚妄**：业界宣称 AI 正在创造庞大的经济增量，齐特隆断然否认。几乎所有头部模型公司都在承受惊人亏损，仅 OpenAI 去年亏损额就高达约 209 亿美元，至今没有任何一家公司能拿出一条切实的盈利路径。
- **关于替代人类工作的炒作**：所谓“AI 将取代人类所有工作”的预言完全脱离现实，没有任何硬核宏观经济数据能支撑这一结论。
- **关于中美 AI 竞赛的恐慌**：针对“美国必须砸下数万亿美元才能在 AI 竞赛中击败中国”的观点，齐特隆痛批这不过是让人整天担惊受怕的借口。当人们担忧“AI 模型落入不法之徒之手”时，他辛辣地指出，这些模型早已在“错误的人”手里了——那就是马克·扎克伯格（Mark Zuckerberg）、萨姆·奥尔特曼（Sam Altman）以及达里奥·阿莫代伊（Dario Amodei）。
- **巨头冷血扩张的本质**：扎克伯格声称将继续激进投资基础设施以满足需求，齐特隆直斥 Meta 就像《怪物史莱克》里的法夸尔勋爵（Lord Farquaad）——“你们有些人可能会死，但这正是我愿意承担的风险”。如果这些亿万富翁能把买 GPU 的狂热分一点点给贫困或现实世界的真正危机，世界绝不会是今天这副模样。
在正式对谈展开前，主持人穿插了一则订阅呼吁，向观众展示后台数据并恳请大家检查订阅状态，承诺节目将持续带来最犀利、最具深度的交锋。
---
### [02:36](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=156s) 为什么说生成式 AI 本质上是一场骗局？
正式访谈拉开帷幕，巴特利特直言齐特隆的观点与以往做客节目的所有嘉宾大相径庭，堪称科技圈罕见的“逆行者”。齐特隆毫不退缩地进一步阐明了自己的立场：生成式 AI 之所以是骗局，在于它从未被当作诚实的软件来销售。科技企业在能力、未来前景以及底层财务数据上进行了全方位的系统性夸大，甚至到了误导全球的地步。这种炒作精准地击中了新闻媒体在技术审查上的软肋，绑架了投行卖方分析师（sellside analysts）的预期，更将宏观经济和政府监管拖入了危险的盲区。
```mermaid
flowchart TD
accTitle: AI 炒作的虚假闭环
accDescr: 描绘了硅谷科技巨头如何利用虚幻承诺绑架媒体、资本市场和用户的运作机制。
subgraph Hype["叙事包装（神话）"]
A["宣传为全知全能的'魔法'<br/>宣称将治愈癌症与替代一切工作"]
end
subgraph Reality["底层现实（真相）"]
B["昂贵且不稳定的云端软件<br/>极其繁琐的提示词工程"]
C["巨额资本开支与严重亏损<br/>OpenAI 与 Anthropic 自身难保"]
end
subgraph Exploitation["外部漏洞利用"]
D["媒体跟风报道与缺乏专业核查"]
E["卖方分析师盲目估值与预期绑架"]
end
A --> D
D --> E
E --> C
B -.->|落差| A
```
当被问及“骗局”是否用词过火时，齐特隆尖锐地反问：“如果他们把它包装成包治百病、能治愈癌症、取代一切人工的魔法，而你剥开外壳却发现它不过是极度昂贵、注定亏损、连搜索基础功能都做不好的枯燥云端软件，而且核心逻辑上极不可靠，那你还能叫它什么？”针对外界质疑他没有深厚金融或工程学术背景的挑刺，齐特隆回应称，自己在科技公关与商业实战中扎根十余载，见证了无数真实技术的发展。令他极度反感的是，每一个 AI 狂热者在展示工作流时，都会描绘出一副如同儿童节目《皮威大冒险》（Pee-wee's Playhouse）般滑稽繁琐的装置：
> “你得在这儿套一个测试框架（harness），提示词必须用这个而不能用那个；开局要用这个模型，收尾却得换另一个模型……这算哪门子‘人工智能’？它原本应当是聪明、自主、设定好就能放手不管（set and forget）的系统！”
---
### [04:26](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=266s) 循环注资与隐瞒数据：科技巨头的财报幻术
随后，巴特利特列举了目前处于风暴中心的六大巨头：Anthropic、亚马逊（Amazon）、英伟达（Nvidia）、微软（Microsoft）、OpenAI 和谷歌（Google），追问齐特隆是否连这些巨头的商业模式也全盘否定。齐特隆从财务账面上彻底揭开了这一虚假繁荣的底牌：这些科技巨头的整体营收根本不是靠 AI 驱动的；直到最近，AI 贡献的收入在绝大部分财报里依然微不足道。更荒谬的是，在少数能看到的“AI 相关收入”中，有高达 70% 的份额其实仅仅来自 OpenAI 和 Anthropic 这两家初创公司。
齐特隆指出，整个 AI 产业链正在上演一场危险的“资金内循环”与纸面富贵：
- **巨额输血与左手倒右手**：亚马逊今年先后向 OpenAI 投出 500 亿美元级别、向 Anthropic 输送 50 亿美元，谷歌也向 Anthropic 注资 100 亿美元。这些初创公司之所以能产生所谓的云服务开销，完全是因为巨头自己掏钱给它们，让它们反过来采购自家的算力。
- **脱离实际的分析师预期**：在未来三年半里，卖方分析师的研报估算竟然指望这两家自身难保、长期失血的亏损企业贡献超过 4,000 亿美元的收入，拉动巨头云业务 30% 以上的增长。
- **数据透明度极其低下**：这些上市公司极度缺乏对公众投资者的尊重，从未在财报中合规拆解 AI 的具体净收入。偶尔为了应付提问，巨头高管便抛出一个未经审计、定义随意的“年化运行率”（annualized run rate）。这个数字有时代表当前月份乘以 12，有时是乘以 13，有时又是最近四周乘以 13，随心所欲且从不披露统计口径。
齐特隆总结道：“这是一场被宣扬为史上最伟大、最颠覆的软件革命。但每当你问这些上市公司‘你们靠这到底赚了多少真金白银’时，他们却羞羞答答地声称商业机密无法公布。上市公司有好事从来不会藏着掖着；当他们遮遮掩掩时，沉默本身就已经说明了一切。”
---
### [06:15](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=375s) 算力黑洞：数据中心、CapEx 与阿比林超级算力基地
针对巴特利特提出的“你难道亲自用过 Gemini、ChatGPT、Claude 这些工具却觉得它们一文不值吗”，齐特隆承认这些工具确实具备一些细微价值，但其价值与全行业投入的成本存在灾难性的鸿沟。如今全行业在 AI 上的资本性支出（CapEx）已经突破了一万亿美元大关。
为了让非专业读者理解这一概念，齐特隆详细阐述了资本支出的本质：不同于电费这种随用随扣的日常运营开支（OpEx），CapEx 指的是数据中心、服务器硬件等长期重资产投资。而为了驱动当下的生成式 AI，科技界正在以前所未有的密度堆砌昂贵的 AI GPU（如英伟达芯片），这带来了近乎变态的电力与空间消耗：
| 对比维度 | 英国布里斯托尔市（City of Bristol） | OpenAI 与甲骨文合建的“星际之门”（Stargate Abilene） |
|:--- |:--- |:--- |
| **年耗电规模** | 约 700 - 800 兆瓦（MW） | 高达 1.2 吉瓦（GW / 1,200 MW） |
| **占地面积** | 约 12 亿平方英尺 | 约 99.8 万平方英尺 |
| **空间密度比** | 基准城市尺度 | 占地比布里斯托尔**小约 1,172 倍**，电力与算力极度浓缩 |
| **硬件部署** | 城市公用电网体系 | 8 栋建筑群，部署数十万张英伟达最新 GB200 芯片 |
齐特隆警告，除了坐拥现金金山的微软之外，几乎所有参与军备竞赛的巨头如今都不得不大举举债来维系数据中心建设。巨头们已经烧掉了一万亿美元，明年还打算再烧一万亿，然而换来的全部回报不过区区几百亿美元营收，且大半还来自两家严重亏损的初创公司。
面对这一连串冷酷的财务算账，巴特利特立刻抛出了科技界最强有力的反驳论据：用户增长曲线（adoption curve）。巴特利特指出，ChatGPT 等工具是人类科技史上用户增速最快的产品，全球数以亿计甚至数十亿人在日常生活中自发地使用它们解决切身问题；在科技投资逻辑中，“金钱回报往往是价值显现的滞后指标”（money is a lagging indicator of value），巨头们此刻大举烧钱，只是在规模化变现爆发之前抢先布局。然而，齐特隆对这一论调嗤之以鼻，随即打断并质疑道：“当人们是被迫在各种软件里塞满生成式 AI 时，这能算得上是诚实的用户自发接纳吗？”一场关于“真实需求还是强推硬塞”的激烈争辩随即拉开序幕。
### [08:43](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=523s) “非自愿”的技术强推与资本的无头怪兽
Ed Zitron 指出，当今的科技生态充斥着对 AI 的强行灌输：打开 Google Docs，Gemini 会在耳边喧闹；打开 Word，Copilot 如影随形地打扰；打开亚马逊，名为 Rufus 的 AI 助手对你买什么袜子都要指手画脚。他直言这是“人类历史上规模最大的一场非自愿技术强推（the largest non-consensual push of technology in history）”。近三年来，各大媒体无休止地制造恐慌，宣称“AI 将抢走你的工作”、“你必须使用它，否则就会落伍”。大众之所以蜂拥而上，完全是因为被持续不断地耳提面命。
在实际使用中，大部分人把生成式 AI 当作搜索引擎的替代品。这一方面是因为 Google 自身的搜索体验在退步，另一方面是因为生成式模型有时更擅长消化复杂的模糊查询。Ed Zitron 将生成式搜索比喻为一艘“拖网渔船”：它在精准细节上极易出错，但如果你询问某件事是否存在、某人是否说过某句话，它能替你捞遍整个数字海洋——哪怕给出的答案大概率依然是错的。然而，无论如何美化，这种功能根本不值万亿美元的投入。与历史上铁路投资等重大泡沫不同，铁路泡沫破灭后留下了可用的基础设施，而 AI 专用 GPU 甚至无法转作他用。他犀利地将生成式 AI 描绘为一个“资本主义无方向的意念聚合体（directionless egregor of capitalism）”，一个跌跌撞撞到处寻找增长的无头怪兽，妄图通过恐吓大众、贬低人类劳动力来逼迫所有人就范。
### [10:10](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=610s) 真实采用率背后的幻象与“AI 垃圾内容”机制
主持人 Steven 提出了质疑，从自身企业与普遍观察切入：据企业采用率统计显示，已有 88% 的机构在至少一项业务职能中定期使用 AI；而在 Steven 自己的公司，更有高达 95% 的员工每天都在使用 Anthropic、ChatGPT 或 Gemini，从编写代码到日常写作和生图，覆盖了从重度用户到高管团队的各个层级。互联网上铺天盖地的 AI 生成内容（AI slop）也证明了这些工具正在被广泛采纳。
面对 Steven“大家确实都在用”的反驳，Ed Zitron 阐明了“AI 垃圾内容”与技术价值的本质区别。他解释称，在 AI 垃圾充斥网络之前，互联网早已被“SEO 垃圾内容”淹没。Google 过去因 Prabhakar Raghavan（时任高管）等人削弱了垃圾内容过滤机制，导致那些为了迎合搜索引擎算法、读起来像“脑震荡患者写的”粗劣博客排在搜索前列。生成式 AI 并没有创造新价值，只是将这种低质内容的生产效率武器化、规模化，变成了一台高效率的“垃圾内容制造机”。
```mermaid
flowchart TD
accTitle: AI垃圾内容与传统SEO垃圾内容的演变路径
accDescr: 展示从早期搜索引擎激励低质SEO内容，演变到生成式AI将其自动化并规模化生产的过程。
subgraph Phase1["传统互联网搜索阶段"]
A["Google调整算法 / 削弱防垃圾机制"] --> B["催生大量低质SEO垃圾博客"]
B --> C["为了排名而非读者阅读"]
end
subgraph Phase2["生成式AI时代"]
C --> D["大语言模型作为规模化生产引擎"]
D --> E["全网充斥高密度AI垃圾内容"]
end
```
### [12:00](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=720s) Token 的底层账本与巨额补贴的不可持续性
随后，Ed Zitron 揭开了 AI 产业光鲜外表下的财务危机。AI 服务的计费基础是 Token（相当于大约四分之三的单词）。用户在调用模型时，不仅要为输入的材料（如文档、代码库）付费，还要为输出的文本付费，甚至模型内部“推理思考”的过程也会产生大量的输出 Token。
为了推向大众市场，AI 公司通过按月订阅制掩盖了真实的推理开销，并设置了模糊的速率限制。Ed Zitron 援引半导体分析机构 SemiAnalysis 的严谨调研数据，揭示了极其惊人的倒贴幅度：
- 在每月 200 美元的 ChatGPT 订阅中，重度用户实际可以烧掉高达 14,000 美元的 Token 成本。
- 在 Anthropic 的 200 美元订阅中，用户可以烧掉约 8,000 美元的 Token。
- 即便是普通人每月 20 美元的入门级订阅，所消耗的计算 Token 价值也常达到 400 美元。
大众误以为 AI 的成本就是每个月掏出的 20 美元，但所有 AI 公司都在承受触目惊心的亏损，OpenAI 上一年度的亏损就达到了惊人的 209 亿美元。当各大厂商试图在 2026 年 3 月前后将超过 150 人的企业客户转为全额按量付费时，客户爆发了强烈反弹，就连 Sam Altman 本人也坦言“客户意见极大，这是一个巨大的麻烦”。以 Uber 为例，该公司在短短三个月内就烧光了全年的 Token 预算。一旦必须按真实成本掏钱，原本高呼“AI 改变一切”的企业立刻退缩。不仅如此，只要模型产生幻觉或把代码搞砸，哪怕生成的是废品，用户也要为每一个消耗掉的 Token 买单。
| 订阅计划与产品 | 每月向用户收取费用 | 用户可消耗的真实 Token 成本 | 商业模式本质 |
|:--- |:--- |:--- |:--- |
| **ChatGPT 高级订阅** | $200 / 月 | 最高达 $14,000 | 极其严重亏损补贴 |
| **Anthropic 高级订阅** | $200 / 月 | 约 $8,000 | 严重亏损补贴 |
| **入门级标准订阅** | $20 / 月 | 约 $400 | 赔本赚吆喝 |
| **企业真实按量计费** | 按百万 Token 结算 | 企业全额承担 | 导致 Uber 等公司 3 个月烧光全年预算并引发客户抵制 |
### [16:09](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=969s) 烧钱换市场假说与不可逆的资本陷阱
面对 Ed Zitron 的抨击，Steven 提出了科技史上经典的抗辩视角：科技巨头是否只是在“价值完全展现之前提前支出（spending ahead of the value）”？历史上科技行业常常通过烧钱来获取市场份额，并在这个过程中推动成本下降、最终实现盈利，厂商此时不敢承担“投资不足”被淘汰的风险。
Ed Zitron 坚决驳斥了这一假设。他指出，如果成本真的能降下来，现在就该看到成效，但现实是推理成本居高不下，数据中心和算力出租公司无一能够实现稳定盈利。数据中心搭建推理算力存在严重的结构性风险：部署 GPU 必须提前按峰值需求预留，买多了会面临昂贵的小时闲置费，买少了则会导致响应迟缓、用户流失。整套模式就像以 1 美元的价格甩卖价值 20 到 40 美元的东西，造就了虚假的需求繁荣。
Ed Zitron 认为这并非一开始就谋划好的阴谋。在 2023 年，风投机构和科技公司普遍乐观预期芯片效率会几何级提升，押注客户未来愿意为惊人价值买单。但这种幻想在 2024 年 OpenAI 亏损突破 50 亿美元时就该破灭并止损。然而，资本列车已经无法停下——因为这个宏大的叙事推高了股市估值，推动了英伟达、微软和全产业链的狂欢，即便这些繁荣与真实的业务营收毫无干系。
### [17:41](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1061s) 巨额资本支出与微薄回报：科技巨头的数字假象
多年来，各大科技巨头一直在高调宣称他们的 AI 押注已经获得了丰厚回报。然而，这些公司却始终拒绝公布他们究竟从纯 AI 业务中赚了多少钱。事实上，巨头们的财报收入之所以持续增长，完全是因为既有业务的表现：Google 和 Meta 改变了广告投放与计价规则，Amazon 不仅在电商平台上提价，还顺势建立起了庞大的广告业务平台。这些增长与生成式 AI 毫无关系，但由于“数字在上涨、营收在攀升”，外界便盲目相信这一切都是 AI 的功劳，仿佛这些公司绝不可能毫无根据地烧掉数万亿美元。
然而残酷的财务数据揭示了完全不同的现实。在刚刚结束的 2026 财年，微软在 AI 相关的营收总额约为 343.3 亿美元，但其中高达 241 亿美元直接来自 OpenAI 相关的结算，这意味着微软自身在整个财年真正产生的独立 AI 业务收入仅有大约 100 亿美元。与之形成鲜明对比的是，微软在这一年里砸下了整整 1,150 亿美元的资本支出（CapEx），并且计划在次年进一步将支出扩大到 1,750 亿美元。这种“数学逻辑完全不成立”的商业模式，寄希望于 AI 能以指数级变得更有价值，从而最终超越其惊人的成本；但大语言模型（LLM）本身是一个极其吞金的怪兽，不仅需要源源不断的定制数据流进行训练，推理与维护成本也居高不下。
与此同时，这种畸形的产业链被英伟达（Nvidia）推向了极致：英伟达在上一财年卖出了价值 2,159 亿美元的硬件（绝大多数是 GPU），但这笔天文数字的军备竞赛所支撑的，是整个行业除少数几个靠不断注资输血苟活的初创公司之外，全球仅约 220 亿美元的实际 AI 软件收入。Ed 援引分析师 Edson 的观点指出，公众对超级富豪有一种“邪教般的顶礼膜拜”（cult-like worship of the wealthy），普遍认定大人物不可能无缘无故挥霍巨资。承认科技界权贵并非天赋异禀的天才，而是仅凭运气、投机取巧以及在合适时间地点拿到 MBA 的普通人，这种现实过于令人不安，以至于人们宁愿相信“自己一定遗漏了什么大智慧”，而这恰恰是科技巨头希望大家维持的错觉。
- **微软 2026 财年 AI 营收明细**：根据彭博社数据，微软总 AI 收入约 343.3 亿美元，其中 241 亿美元来自 OpenAI，微软自身独立 AI 营收仅约 100 亿美元。
- **资本支出差距**：微软单财年资本支出高达 1,150 亿美元，并计划次年追加至 1,750 亿美元，与百亿级收入严重倒挂。
- **英伟达硬件销售与终端市场差距**：英伟达上一财年芯片销售额约 2,159 亿美元，而终端实际支撑的全球 AI 软件营收总额仅约 220 亿美元左右。
---
### [20:07](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1207s) “创新者的窘境”与汽车隐喻：AI真的具有指数级潜力吗？
主持人心生疑问，搬出了经典商业著作《创新者的窘境》（*The Innovator's Dilemma*），试图从技术历史的角度为 AI 辩护。他指出，许多最终颠覆或重塑整个行业的重大创新，在诞生初期往往性能极差、经济账根本算不过来，而且客户根本没有主动索求过。例如 19 世纪的马车曾被视为完美的交通工具，而初生的汽车不仅频繁抛锚（正如现在的 AI 频繁产生幻觉），而且造价高昂，甚至在法律层面还一度要求有人手持红旗走在汽车前头示警（红旗法案）。汽车最初显然是更糟糕的选择，但作为颠覆性技术，它拥有远高于马车的成长天花板。主持人据此设想，如果 AI 保持每个月 5% 的改进速度，同时伴随类似摩尔定律带来的成本下降，人类工程师将很快无法望其项背。
Ed 对此进行了针锋相对的反驳。他首先澄清，所谓的摩尔定律并不适用于现代 GPU 架构。英伟达早在 2000 年代就推出了底层软件架构 CUDA，花费了十多年心血才将其打磨至足以胜任早期数据分析与大规模计算的程度；当生成式 AI 爆发时，英伟达已经积累了无可替代的技术与生态壁垒。然而，当年的汽车工业并没有集结全球几乎所有顶尖数学家和科学家的脑力，也没有让全球各国政府一天到晚为其站台背书并宣称“颠覆不可避免”。Ed 质疑所谓“每月 5% 的进步”根本无法量化：人类初级软件工程师能够通过人际交往的上下文语境、解决问题的现实环境以及团队氛围中学习成长，而这些细微的现实认知与权衡决不是简单地扫一遍代码库或读几封电子邮件就能获得的。
```mermaid
flowchart TD
accTitle: 汽车颠覆理论与当前AI发展的反差
accDescr: 对比传统颠覆性创新与当前生成式AI在投入规模和改进瓶颈上的根本差异
subgraph Classical["经典颠覆性创新（如汽车替代马车）"]
A["初期产品表现糟糕且成本极高"] --> B["在边缘场景艰难演进"]
B --> C["突破物理与工程瓶颈"]
C --> D["实现远超旧技术的增长天花板"]
end
subgraph GenerativeAI["当前生成式AI的现实瓶颈"]
E["已消耗全球顶尖科研资源与政府倾斜"] --> F["底层算力与数据面临边际效应递减"]
F --> G["缺乏人类真实上下文与环境理解能力"]
G --> H["难以定义或实现持续的质量跃迁"]
end
```
---
### [22:44](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1364s) 软件质量的整体退化：交付量与代码质量的脱节
针对主持人提出以“交付代码量”（shipped code）来衡量工程师生产力 5% 增长的标准，Ed 直斥这是一种极其荒唐的评估逻辑。他讽刺道：“这就像因为某个人的通讯简报写得特别长，就判定他是全世界最优秀的作家一样荒谬。”在软件工程领域，代码的行数和交付频率从来不能等同于价值，真正的衡量标准在于产品是否变得更加可靠、软件质量是否得到实质性提升。
现实情况却走向了完全相反的极端。Ed 直言，随着大语言模型在行业内被强制推行，主流科技巨头的整体软件质量正在全面恶化：
- **巨头软件水准滑坡**：无论是 Google、微软、Amazon 还是 Meta（Ed 甚至直接用“庞大的怪物/灾难”来形容 Meta 的糟糕工程），核心软件体验均未见改善，反而频频出现低级错误。
- **基础设施的不稳定性**：以微软旗下的 GitHub 为例，社交媒体上甚至有人讽刺称“GitHub 正常运行的时候才应该推送系统通知，宕机反而成了常态”。作为全球市值名列前茅的公司，微软在运营自家关键开发者平台时却显得力不从心。
- **强制使用 AI 的反噬**：当管理层自上而下强制要求员工在日常工作流中使用 LLM 生成代码时，软件代码库中的漏洞和不稳定性正在以肉眼可见的速度蔓延。
---
### [24:55](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1495s) 幻觉未曾消失：从彭博终端错误到“氛围编码”的安全隐患
主持人回忆起 ChatGPT 刚问世时他在亚洲向未婚妻演示的情景，承认当时模型确实频繁出错，但他坚称现在使用时已经很少遇到事实性幻觉，更多只是觉得模型的逻辑推理略显薄弱。Ed 对此表示坚决反对，并分享了自己在彭博终端（Bloomberg Terminal）上的亲身经历：
彭博终端内置了名为“Ask B”的 AI 工具，允许用户用自然语言提问，系统会自动生成专用查询语言 BQL 并提取经过验证的数据库内容，这本应是极度可靠、抗幻觉能力最强的专业场景。然而，几天前 Ed 在撰写简报时，尝试查询微软、Google、Meta 和 Amazon 过去五年的股票增长率，并将数据复制粘贴到 Excel 中。当他复核时惊讶地发现，系统给出的数据中赫然写着“微软股价曾达到每股 575 美元”，而历史上微软股票从未达到过这一价位。
Ed 强调，当这样的错误发生在个人财经写作中时，一眼识别出来尚且可以说是“有惊无险”；但如果这套工具被部署在极其严肃的专业领域——例如医生的病历转录诊断、对冲基金依赖的核心财务模型，或者大型软件系统的重构时，后果将是致命的。如果一个工程师在长达半年的时间里依赖所谓的“氛围编码”（vibe coding，仅凭直觉和 AI 辅助，而不再亲自动手编写核心逻辑），一旦 AI 在重构代码时遗留了严重的安全后门或底层 Bug，已经荒废了手写代码能力的开发者将根本无从察觉，面对海量的机器“数字垃圾”（slop）只能束手无策。这类错误具有连环相扣的乘数放大效应（multiplicative），而整个 AI 行业至今根本找不到能够彻底消除模型幻觉的技术路径。
### [26:28](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1588s) 基准测试的幻象与技术进步的速度
对话围绕 AI 的能力评估展开。Ed Zitron 指出，当前评估大语言模型能力提升的主要方式，是使用专门为 LLM 量身定做的基准测试（benchmarks），因为这些模型根本无法胜任真实的端到端任务。即使在某些任务排行榜上，模型表现看似有了长足进步——例如号称可以自主执行任务长达一小时——但仔细查阅底层数据就会发现，它们成功完成任务的概率仅仅只有 50%。
主持人随后提出反驳，引用了 Vectara 幻觉排行榜（Vectara Hallucination Leaderboard）的历史数据：在基础的文本摘要任务上，过去四年间模型的幻觉率已经从大约 21.8% 暴跌至如今像 Gemini 和 ChatGPT 这类前沿模型的约 0.7%。主持人认为，尽管这些测试针对的是简单任务，但这种显著的“进步速率（Rate of improvement）”预示着未来幻觉会越来越罕见。
主持人以互联网的发展历程作为类比：每一项突破性技术在诞生之初都伴随着严重的体验瓶颈。他回忆自己小时候用拨号上网（dial-up modem）时，玩着《符文之地》（RuneScape）游戏就不能接打电话，家里拿起电话网络就会断掉。当时人们也会抱怨这技术简直糟糕透顶，但谁能想到多年后会拥有 Starlink 卫星网络以及手机 5G 互联？面对新技术时，人们不应将初始形态与完美形态对立。
```mermaid
flowchart TD
accTitle: LLM 基准幻象与真实表现差距
accDescr: 图表展示了基准测试中的乐观数据与实际应用中局限性的对比
subgraph Benchmarks["基准测试层面的乐观叙事"]
A["Vectara 幻觉基准测试"] --> B["简单摘要任务幻觉率从 21.8% 降至 0.7%"]
B --> C["外推推论：随技术演进幻觉将彻底消失"]
end
subgraph Reality["真实复杂场景中的断裂"]
D["复杂或长时间任务"] --> E["连续运行1小时"]
E --> F["任务成功率仅约 50%"]
F --> G["缺乏决定性校验与上下文理解"]
end
```
### [28:12](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1692s) 对比完美还是对比平替？实习生与人类编辑的本质区别
主持人进一步提出核心论点：在工作环境中，人们不应该将 AI 与“绝对完美”进行对比，而应该与实际的替代方案做比较。如果一个简单的写作任务交给自己或刚入职的实习生，人类不仅消耗时间，也同样存在知识盲区并可能犯错（即人类的“幻觉”）。如果一个 AI 模型只有 0.7% 的幻觉率，知识面远超普通人且产出极快，从净效益（Net basis）来看这无疑是一笔划算的交易。
Ed Zitron 对这种“实习生类比”和“全知机器”的论调进行了猛烈回击。他以自己的长期编辑、居住在利物浦的资深科技记者 Matt Hughes 为例指出：聘请一位优秀的人类合作者，绝不是因为他“什么都知道”，而是因为他具备深度的行业背景认知（Context）、扎实的专业积累，以及作为人的同理心与喜怒哀乐——面对热爱的事物充满热情，面对厌恶的行径则毫不留情。这些特质是大语言模型完全不具备的。
Ed 进一步强调，实习生的核心价值在于其“可成长性”。人们雇佣实习生不是为了让他们扮演只会被动查询日期的工具，而是为了传授经验、培养上下文感悟，并帮助他们从懵懂状态成长蜕变为成熟人才；而把大语言模型当实习生纯属无稽之谈。
- **人类合作者的不可替代性**：具备真实的感情、深度的行业背景、能够共同深入探究复杂命题（如调查黑石集团等资产管理机构背后的资本运作）。
- **实习生的成长机制**：通过吸收现实经验、理解领导习惯、建立情感连接完成长期蜕变。
### [30:08](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1808s) 记忆机制、外部脚手架与 950 亿美元估值的荒谬
主持人试图反驳 Ed，坚称 AI 同样具备记忆上下文和学习的能力。他举了一个简化示例：“如果我打开 Claude 问它我养的狗叫什么名字，它能答得上来”，这证明模型拥有对过去信息的记忆。
听到这个例子，Ed 情绪激动地讽刺道：“上帝啊，就为了这个，这家公司估值达到了 950 亿美元？！”Ed 解释说，模型所谓的“学习”和“记忆”，本质上只是用户为其搭建了一个庞大的 `.md` 上下文文件（例如像 Pee-Wee's Playhouse 里的古怪机械装置一样，套上一层层外围脚手架 harness），而模型在检索时有时读取、有时漏读。为了抑制幻觉，用户不得不做出极其繁琐的外部配置。
| 维度 | 人类记忆与认知 | 大语言模型（LLM）“记忆” |
|:--- |:--- |:--- |
| **底层实现** | 情感、感受与事实深度交织的复合体验 | 挂载外部文本文件（如 `.md`）并进行向量检索 |
| **状态感知** | 理解情绪变化、周围真实世界动态与状态起伏 | 静态文本生成器，无心智状态与真实体验 |
| **适应能力** | 在长期共事中自主内化默契与微小工作习惯 | 极易受输入窗口与检索准确率波动的机械匹配 |
### [31:30](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1890s) 输出至上论 vs 过程价值论：AI 是否能替代深度思考？
争论随后上升到了哲学层面：价值究竟存在于“输出结果（Output）”，还是存在于“生产过程（Process）”？
主持人主张“输出至上论”。他认为对于绝大多数商业需求而言，用户根本不在乎底层是人脑思考还是机器矩阵计算；只要将所有文件喂进 Claude，当提问“我的营收是多少”时它能瞬间给出正确数字，这个工具就提供了核心价值。同理，人们探求 AI 是否具有“创造力”，衡量的标准也应当是其生成的成品是否令人满意，而不是其思维过程是否模仿了人类。
Ed 对此表达了彻底的反对。他以自己与编辑 Matt Hughes 共同撰写长达 11,000 字的研究长文为例，认为深度学习与创作的价值恰恰深植于共同摸索、推演假设、反复推敲的探索过程之中，这种交互产生的新颖洞见绝非大模型所能提供：
- **平均化的平庸产物**：LLM 本质上是在对其抓取到的文档做概率平均（Average），只能批量制造平庸的“AI 垃圾（AI slop）”。即便是对冲基金使用的高端 LLM 脚手架分析系统，吐出来的也是千篇一律、毫无新意的泛化报告。
- **关键财务数据的致命盲区**：在涉及财务、营收等核心数据时，绝不能依赖具有概率生成特性的 LLM。即便使用彭博终端（Bloomberg Terminal）运行 BQL，分析师也需要反复进行两重、三重甚至多重交叉验证；而 AI 报告越详细，隐藏的错误往往越多。如果机器有哪怕一次把营收算错，整个决策链条就会崩塌。真正的信任建立在经过严苛推导与验证的确定性流程之上，而不是把底牌交给一个黑盒文本生成器。
### [34:59](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2099s) 人类信任与“顾左右而言他”的 AI 基准测试
对话从“信任从何而来”展开探讨。Steven 提问 Ed，为什么 Ed 能对他的长期合作伙伴 Matt 寄予深厚信任——是因为过往持续优异的表现吗？Ed 承认确实如此，而且他们多年来是一同边做边学积累起来的。Steven 指出，人类信任某事物的核心在于其能够持续兑现承诺；正因如此，当 Claude 等模型在幻觉排行榜（hallucination leaderboard）上针对简单任务持续给出稳定表现时，人们才会如此迅速地信任并采纳它。
```mermaid
flowchart TD
accTitle: 信任建立与价值呈现对比
accDescr: 对比人类建立信任的方式、直观革命性硬件与大模型基准测试的根本差异。
subgraph HumanTrust["人类信任建立"]
H1["长期协作与共同成长"] --> H2["持续兑现承诺"]
H2 --> H3["建立坚实信任"]
end
subgraph TechValue["技术价值验证方式"]
T1["iPhone / AWS 原生直观价值"] --> T2["无需解释，上手即懂"]
A1["LLM 行业指标"] --> A2["跑分与排行榜（Benchmark）"]
A2 --> A3["回避直接疑问：顾左右而言他"]
end
```
然而，Ed 对所谓的“简单任务排行榜”嗤之以鼻。他当场追问所谓“简单任务”到底是如何定义的，而 Steven 也坦承自己并不清楚。Ed 指出，这恰恰暴露了整个 AI 行业的通病：他们是“顾左右而言他”（whataboutism）的大师。当 AI 公司无法用事实证明自身实用价值时，就会抛出一大堆抽象的跑分与基准测试，宣称“看，我们的数字更高了”，但从来没人能解释这些数字在现实场景中究竟代表什么。Ed 强调，真正革命性的技术根本不需要这套说辞：
- 2007 年初代 iPhone 推出时，Ed 正在宾夕法尼亚州立大学念书。当他第一次看到可视语音信箱等功能时，内心的震撼犹如《2001太空漫游》开头见证黑色石板的猿猴。
- 无论是数码发烧友还是平时只用摩托罗拉 Razr、诺基亚 3210 的普通人，只要看一眼 iPhone，其价值是不言而喻的。
- 亚马逊云服务（AWS）同样如此，其颠覆性的商业模式和工具价值无需任何繁复的掩饰或转移话题。
---
### [36:48](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2208s) 史蒂夫·鲍尔默的嘲笑与初代 iPhone 的直观价值
Steven 提出了反对意见，提醒 Ed 不要事后诸葛亮。他提到了微软前 CEO 史蒂夫·鲍尔默（Steve Ballmer）当年的名场面：在受访谈及初代 iPhone 时，鲍尔默当场爆笑，嘲笑它售价高达 500 美元（还是带合约补贴的价格），称其是“世界上最贵的手机”，而且没有实体键盘，对企业用户来说根本无法作为高效的邮件收发设备，远不如 99 美元的摩托罗拉 Q 系列实用。
Ed 对此回应道，鲍尔默的嘲笑掩盖不了一个基本事实：即便初代 iPhone 很贵、早期移动网页体验糟糕，但它的图形交互界面（GUI）彻底改变了人机互动。黑莓手机虽然昂贵且酷炫，但本质上不是现代消费级软件逻辑。iPhone 让人一眼看出它是一台被放进掌心的电脑。反观生成式 AI，发布至今只要有人追问“它到底神奇在哪里”，拥趸们依然含糊其辞、犹豫不决，不得不承认如果离开一整套繁复的“杠杆与滑轮”（各种补丁工具链与复杂流程），它根本无法支撑任何实际业务的完整运转。
---
### [38:37](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2317s) 疯狂的增长数据、真实成本与被补贴的假象
面对质疑，Steven 列举了一系列颠覆性的增长数据，以证明 AI 的价值已被公众广泛认可：
- ChatGPT 仅用 60 天就达到了 1 亿月活跃用户，而 TikTok 达到这一量级用了 9 个月，Instagram 用了 2.5 年，万维网（WWW）甚至用了大约 7 年。
- 发布仅 3 年，已有超过 60% 的美国成年人将 AI 工具融入日常流程，渗透率达到总人口的 40%；相比之下，互联网走完这一渗透历程用了 5 年，个人电脑（PC）用了近 12 年。
- Steven 还举了未婚妻的亲身经历：作为一名英语非母语的独立创业者，她原本需要自掏腰包雇佣平面设计师和文案撰稿人，如今借助 ChatGPT 独立生成文案与图像，她认为该工具对自己的业务带来了彻底的“变革”（transformative）。
Ed 针锋相对地指出了这一繁荣背后的核心漏洞：**虚假成本**。
> “如果这是按它的诚实成本来卖呢？如果大家每次操作都要自掏腰包付上 2 美元、3 美元、4 美元，而他们依然觉得物超所值、心甘情愿买单，那我可能无话可说。”
Ed 强调，用户现在感受到的所谓“高性价比变革”，完全建立在科技巨头数十亿美元的风投贴补之下。如果用户必须按照真实的每百万 Token 运行成本支付全额 API 费用，绝大部分狂热追捧瞬间就会烟消云散。
---
### [40:06](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2406s) 毫无路线图的炒作与“腐烂经济”泡沫
Ed 回忆起自己最早用 33.6 kbps 拨号调制解调器上网的日子，在 Happy Puppy 等网站通宵等待共享软件下载，那时所有人都能清晰预见到“只要网速变快，就能在线看视频”。高盛分析师吉姆·科维洛（Jim Covello）在 2024 年的一份知名报告中明确指出：在 iPhone 诞生前夕，行业内有成千上万场演讲清晰规划了随着 GSM、蓝牙、Wi-Fi 射频芯片体积缩小，智能手机的到来是必然的历史演进；但今天的生成式 AI 根本没有这样的技术路线图，纯粹是“过度投入却回报寥寥”。
```mermaid
flowchart LR
accTitle: 互联网/移动通信与生成式 AI 演进对比
accDescr: 对比通信硬件有据可循的技术路径与生成式 AI 缺乏路线图的过度炒作。
subgraph Historical["历史硬件路线"]
H_Tech["射频芯片微型化<br/>(GSM / 蓝牙 / Wi-Fi)"] --> H_Goal["智能手机形态成熟<br/>(必然演进路线)"]
end
subgraph GenerativeAI["当前生成式 AI 现状"]
G_Hype["宣称颠覆一切<br/>取代所有人类工作"] --> G_Reality["缺乏成熟路线图<br/>巨额亏损 + 能源通胀"]
end
```
Ed 坦言，如果这些厂商当初实事求是，承认这只是一种昂贵、仍在研发阶段且存在幻觉风险的云端语言模型，他反而会心存敬意。但自 2022 年底以来，硅谷高管们却在大肆兜售神话，宣称它无所不能、将取代所有工作。例如微软 CTO 凯文·斯科特（Kevin Scott）在面对科技专栏作家凯文·罗斯（Kevin Roose）被 Bing 聊天机器人劝诱离婚的荒唐事件时，非但没有澄清这只是个普通的语言模型网页，反而乐见其成地煽风点火。
Ed 痛陈，这种盲目膨胀不仅带来了幻觉，更在现实世界中造成了毁灭性的外部成本：
- 马斯克位于路易斯安那州等地的超级数据中心，依靠燃气轮机日夜轰鸣运转，废气直接危害周边的非裔社区健康；
- 数据中心的极端耗电正在推高普通居民的民用电费账单；
- 对高端内存（RAM/HBM）的疯狂掠夺，正在引发整个消费电子供应链的全面物价上涨。
Steven 最后总结道，即便 Ed 所揭露的负面问题全部属实，但这与“技术终将深刻改变世界”并不矛盾。正如 Ed 在其撰写的《腐烂经济》（The Rot Economy）与“腐烂网络泡沫”（rotcom bubble）文章中所反思的：当年的互联网泡沫中，90% 的投机公司化为乌有、投资人陷入癫狂，但这丝毫没有妨碍废墟中孕育出改变世界的时代级伟大企业。
### [43:36](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2616s) 互联网历史上的著名误判与互联网泡沫的本质
主持人引用了历史上几位名人的反向预测来质疑 Ed Zitron 是否过于悲观：1998 年，诺贝尔经济学奖得主保罗·克鲁格曼（Paul Krugman）曾断言，到了 2005 年左右，人们就会发现互联网对经济的影响“绝不会比传真机更大”；1995 年，天体物理学家克利福德·斯托尔（Clifford Stoll）在《新闻周刊》（Newsweek）发表经典文章，嘲弄计算机评论家毫无常识，断言“没有任何在线数据库能取代日报，没有任何 CD-ROM 能取代称职的老师，商业活动不可能从商场转移到网络和调制解调器上”；克鲁格曼甚至还说过“随着人们发现彼此根本无话可说，互联网的增长将大幅放缓”。
Ed Zitron 调侃克鲁格曼那句“无话可说”可能是历史上最离谱的预测，但也指出斯托尔当年其实敏锐察觉到了海量虚假垃圾信息对社会的危害，并精准预测了网络教育无法真正替代传统面对面教育。然而，Ed 强调 AI 热潮与当年的互联网有着本质上的经济差异。当年的互联网泡沫实际上包含两个层面：
1. **网站泡沫（Website Bubble）**：纯粹是一堆没有商业模式的垃圾公司互相倒买倒卖（例如 Excite@Home 以近十亿美元收购蓝山在线贺卡 Bluemountain.com）。
2. **暗光纤泡沫（Dark Fiber Bubble）**：通信基础设施的严重超前建设。当时华尔街分析师宣称互联网流量每 90 天翻一番，但实际上是每 6 到 12 个月甚至更久才翻倍，导致地下埋入了海量未点亮的光纤电缆与传输基站。
虽然泡沫破裂后互联网需求最终爆发并填满了这些管道，但 Ed 指出，生成式 AI（Generative AI）的处境完全不可同日而语。
### [46:00](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2760s) 生成式 AI 的虚假繁荣与无法摊薄的硬件成本
生成式 AI 当前呈现的所有需求，几乎全部建立在巨额补贴的基础上。普通用户并没有支付其实际的推理和算力成本。不仅如此，这场技术运动已经动用了人类历史上规模最大、最不坦率的市场营销攻势，甚至由云计算霸主微软亲自下场推销，结果微软自身的 AI 软件收入仅有数十亿美元；放眼整个行业，除去 OpenAI 和 Anthropic，其余所有参与者的年化收入加起来勉强达到 220 亿美元左右。
Ed 直言，对于个人而言 220 亿美元是个天文数字，但面对科技巨头已经砸下的上万亿美元资本支出（CapEx），以及 OpenAI 和 Anthropic 背负的高达 1.1 万亿美元云服务采购承诺（Cloud Commitments），这点收入根本无法填补黑洞。更致命的差异在于运营成本的物理约束：
```mermaid
flowchart TD
accTitle: 基础设施成本对比图
accDescr: 对比传统光纤网络与现代 GPU 数据中心在建成后的边际成本差异。
subgraph Fiber["Dot-Com 时代光纤网络"]
F1["一次性铺设光纤电缆"] --> F2["长期闲置（暗光纤）"]
F2 --> F3["流量到来后只需点亮<br>边际运营成本极低"]
end
subgraph GPU["生成式 AI 时代 GPU 数据中心"]
G1["耗费巨资购置先进制程 GPU"] --> G2["持续消耗海量电力与散热"]
G2 --> G3["折旧极快（3-5 年报废）<br>2050 年运行成本依然极其高昂"]
end
```
- **运营成本无法自然摊薄**：当年铺设的光纤埋在地下，点亮后的边际成本趋近于零；而今天建造的 GPU 数据中心，哪怕到了 2050 年，只要运行就必须消耗同等天价的电力，除非发电技术或芯片物理学出现奇迹般的断代突破。
- **芯片巨头避重就轻**：包括博通（Broadcom）、英伟达（Nvidia）、Etched、ARM 在内的芯片巨头全力攻关，却没有任何一家能神奇地让大模型推理变得低成本或具盈利性。即便是英伟达发布的新一代 Vera Rubin 架构，宣传时也只是含糊其辞地声称“每兆瓦产出美元更多”或“能效提升 10 倍”，却从未敢公开宣称其能够帮 OpenAI 或 Anthropic 将实际运营成本削减 50%。
### [47:42](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2862s) 混淆概念的营销骗局与数据中心供需断层
针对主持人提出 AI 还涵盖医疗、蛋白质折叠、机器人和自动驾驶等领域的说法，Ed 犀利指出，这正是硅谷的叙事套路——刻意把所有完全不同类别的技术统一打包贴上“人工智能”（Artificial Intelligence）的标签：
- **概念捆绑**：将 AlphaFold（蛋白质折叠）、机器人学、甚至自主武器系统等完全不基于 LLM（大语言模型）的成果，与聊天机器人混为一谈。当批评者指出大语言模型无法变现或容易产生幻觉时，拥护者就会拿“AI 正在攻克癌症”来转移焦点。
- **专用数据中心的本质**：当前全球掀起疯狂建设潮的数据中心，绝大部分并非为了传统科学计算或搜索算法，而是专门为大语言模型训练和推理量身定制的高功耗 GPU 集群。像 Matic 这类家用扫地机器人，机身只需运行极低功耗的模型，根本无需时刻烧钱消耗云端超算。
- **天文数字般的产能过剩**：市场研究机构 Sightline Climate 曾统计，全球处于规划阶段的数据中心电力容量高达 190 吉瓦（GW）。按行业电费与配套成本折算，需要每年 1.6 万亿至 3 万亿美元的市场终端需求才能维持运转；而目前全球每年对生成式 AI 的真实软件需求甚至还不足 1,300 亿美元。
当前所谓的 AI 繁荣在很大程度上是一场闭环游戏：微软、亚马逊、谷歌向 OpenAI 和 Anthropic 注资，这些初创公司再将数百亿美元返还给巨头购买云算力，巨头进而向英伟达订购数百上千亿美元的 GPU。外部第三方基建投资者误以为四大巨头采购 GPU 意味着市场上存在不可阻挡的广泛企业需求，于是盲目跟风举债下场建造海量 GPU 数据中心，最终铸就了一座座建立在虚假需求之上的“资本主义神龛”。
### [52:37](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=3157s) 生成式 AI 真的提升了科技巨头的生产力吗？
主持人提出，全球顶尖的科技公司如今都在使用生成式 AI 来编写大量底层代码，这理应为企业带来巨大的生产力跃升。然而，Ed Zitron 对此嗤之以鼻，并反问听众和主持人：最近大家用过 Google、Facebook、Instagram 或是 GitHub 吗？它们的实际体验已经恶化到了灾难性的地步。例如，Amazon Web Services（AWS）今年就曾数次因为其自身的 AI 编程工具导致服务瘫痪。
面对主持人关于“Google 究竟是如何变糟”的追问，Ed Zitron 揭露了一段 Google 内部的权力斗争史。故事的核心人物是 Prabhakar Raghavan（曾任 Google 广告业务负责人）。早在 2019 年，Google 内部拉响了所谓的“黄色警报”（Code Yellow），警示搜索查询次数（query volume）出现了实质性疲软。当时的 Google 搜索负责人 Ben Gomes 明确警告：如果为了人为推高查询次数而牺牲质量，就意味着 Google 得给出更差的搜索结果；因为如果用户能快速精准地找到答案，搜索次数自然会减少。包括工程主管 Shashi Thakur 在内的许多技术人员也呼吁将这一危险信号上报给 CEO Sundar Pichai。然而，主导广告业务的 Prabhakar Raghavan 却执意强推，认为只要增加查询次数，就能展示更多广告，从而赚取更多利润。
```mermaid
flowchart TD
accTitle: Google搜索恶化与AI介入逻辑图
accDescr: 展现从Code Yellow到降低搜索质量增加广告收益，再到AI Overviews圈占流量的演变过程。
subgraph SearchDecline["搜索体验的妥协"]
A["2019年：Code Yellow<br/>搜索查询量增长停滞"] --> B["Ben Gomes等技术派主张：<br/>快速给准确答案，不盲目追求查询量"]
A --> C["Prabhakar Raghavan等商业派主张：<br/>推高查询量，增加广告曝光"]
C --> D["2020年 Raghavan掌舵搜索：<br/>放松对垃圾站点的打压，搜索体验下降"]
end
subgraph AIIntervention["生成式AI的强行植入"]
D --> E["用户被迫在关键词后加 'Reddit'<br/>寻找真实人类内容"]
E --> F["Google遭遇流量留存危机<br/>急需将用户留在站内"]
F --> G["推出顶置AI搜索摘要<br/>（即使建议吃石头、毒蘑菇）"]
G --> H["终极目标：切断外部跳出<br/>让Google自身成为唯一答案源"]
end
```
### [54:23](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=3263s) 从放宽垃圾网站到顶置 AI 摘要：Google 恶化的逻辑
到了 2020 年初，Prabhakar Raghavan 正式接管了 Google 搜索业务。Ed Zitron 指出，虽然没有铁证，但从各大 SEO 论坛和《Search Engine Journal》等行业社区的反馈来看，Google 在那之后大幅削弱了对垃圾网站（spammy sites）的降权打压。这么做可以让用户在糟糕的信息中反复检索、多次点击，从而在页面上逗留更久并看更多广告。这也是为什么如今大量网民在搜索时，不得不习惯性地在关键词后加上“Reddit”来寻找真实的人类答案。
当生成式 AI 浪潮来袭，Raghavan 又被任命负责 Gemini 相关业务。面对用户流失的焦虑，Google 选择将 AI 摘要直接强塞到搜索结果的最顶端。其核心算盘是：用户不再需要点击链接跳转到外部网站，而是被死死钉在 Google 自己的页面里。即使 AI 频繁给出离谱甚至致命的错误答案——比如建议用户“吃石头”或“食用有毒蘑菇”——管理层也毫不在意。在 Ed 看来，生成式 AI 沦为了 Google 商业作恶的终极形态，它不再负责指引信息，而是将自身伪装成无可置疑的答案来源。
### [55:42](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=3342s) 平台崩溃加剧：AI 生成代码引发的基础设施危机
主持人指出，Google 搜索体验变差更多是管理层糟糕的商业决策所致，而非因为工程师用 AI 写代码。Ed Zitron 承认了这一点，但随即指出了另一个更具毁灭性的全行业问题：整个科技基础设施的稳定性正在直线下降。不仅是 Google 搜索，如今的 Google Docs 和 Google Sheets 到处都是 Bug，微软和亚马逊的产品也难逃厄运。甚至连被公认拥有顶级交互体验的苹果 iPhone，其系统现在也变得毛病不断。
对于如何量化“代码质量下降”，主持人现场查阅了最新行业研究数据，证实这绝非个例主观感受：
- **软件停机时间与故障率显著飙升**：近几年的行业数据直接将矛头指向了 AI 辅助编程工具的大爆发。
- **双重危机夹击**：一方面，AI 生成的代码本身包含更多隐性 Bug；另一方面，AI 生成的大量无序代码活动在底层硬件和基础设施上引发了超负荷冲击。
Ed 进一步剖析，由于 GitHub 上充斥着由大语言模型（LLM）一键生成的代码，很多人刚学了一点皮毛，就陷入“达克效应”（Dunning-Kruger effect），根本搞不清楚底层逻辑就将未经验证的代码直接推送到开源项目中。
这种泛滥不仅让开源生态苦不堪言，更在专业工程师之间滋生了致命的倦怠与盲从心理——当工程师看到 AI 写的上一百行代码大体可用时，就会放松警惕，不再仔细审查接下来的代码。人类天生倾向于偷懒走捷径，但 AI 的本质是概率模型，必然会犯错。Ed 愤怒地批评了科技巨头 CEO 们的虚妄宣传：OpenAI 的 Sam Altman（被 Ed 讥讽为“Clammy Sammy”）不断吹嘘 AI 即将完全替代软件工程师；Anthropic 的 Dario Amodei 也声称未来几年内 50% 的白领工作将消失。这种过度营销让公众和开发者误以为 AI 是完美的自动化系统，而在缺乏严格代码审查的环境下，盲目信任只会让系统缺陷呈乘数级恶化。
### [01:00:51](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=3651s) 自动驾驶与失业恐慌的真伪之辩
在节目的商业赞助插播中，主持人介绍了自由职业者高管平台 Fiverr Pro 以及提供全球 200 多个国家数据连接与内置安全防护的 eSIM 应用程序 Saily。
赞助环节结束后，访谈切入了一个新的焦点议题：AI 究竟会不会夺走人类的饭碗？主持人以自己搭载 AI 技术的自动驾驶汽车为例，提到 Uber CEO Dara Khosrowshahi 曾在这间演播室里信誓旦旦地预测，在未来几年内 Uber 将不再需要人类司机，因为全自动驾驶汽车将彻底接管道路。考虑到专业驾驶是全球从业人数最庞大的职业之一，这一技术落地似乎预示着一场不可避免的就业海啸。Ed Zitron 则对此持完全不同的怀疑态度，准备展开深入驳斥。
### [01:01:12](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=3672s) 科技高管的叙事陷阱与自动驾驶的“边缘用例”难题
当主持人提问，那些不断向公众宣称“AI 将带来普遍职业颠覆”的科技公司首席执行官们是否在撒谎时，埃德·齐特隆（Ed Zitron）毫不犹豫地给出了肯定回答。他指出，这些高管要么是在彻底说谎，要么是在做一种对自己极其有利的盲目猜测。以微软首席执行官萨蒂亚·纳德拉（Satya Nadella）为例，纳德拉不可能公开承认“我们其实也不知道这玩意儿到底管不管用”，他必须为自己的商业利益摇旗呐喊（“talk his book”）。最初纳德拉大肆鼓吹 AI 将替代所有员工，后来发现技术根本跟不上吹嘘的节奏，便又悄悄改口称 AI 只是为了“增强员工的能力”。优步（Uber）首席执行官达拉·科斯罗萨西（Dara Khosrowshahi）也是同样的逻辑，如果无人驾驶成真，优步就能顺理成章转变为全自动出租车网络，这对公司股价自然是天大的利好。
谈及自动驾驶出租车 Waymo 时，埃德坦言这项技术本身确实很酷，但也伴随着严重的社会经济隐患——驾驶是全球最庞大的就业蓄水池之一，一旦被替代，传统的出租车经济模型将彻底崩塌。然而，现实是人类距离全自动普及“还差得十万八千里”。Waymo 之所以只能在极少数城市进行最谨慎、最受控的微型试点，是因为自动驾驶乃至所有 AI 系统的致命死穴从来不是搞定前 95% 的常规路况，而是极其棘手的“边缘用例”（edge cases）。在旧金山，一场大雨就足以让自动驾驶系统陷入瘫痪；如果一个穿荧光反光背心的小孩突然横穿马路，系统甚至可能根本无法识别出那是个人类儿童。
埃德还分享了自己在现实中目睹的尴尬场景：前几天他在拉斯维加斯的酒店外，亲眼看到一排外形方正古怪的 Zoox 无人驾驶车直接把出口堵得死死的，“就像排着队当场睡着了一样”；在旧金山搭乘 Waymo 时他也遇到过类似情况，前车莫名停在路中央，导致后方的普通车辆和另一辆 Waymo 全部被卡死。埃德强调，人类司机确实也会犯错，但人类社会能够控制无人车的投放节奏，对待这种技术必须保持“有罪推定”的审慎态度，配备人工远程监控与严格限制，绝不能任由资本为了讲故事而在未成熟时盲目铺开。
---
### [01:04:09](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=3849s) 自动驾驶安全争议与白领岗位“被颠覆”的谎言
面对埃德对自动驾驶的质疑，主持人列举了一系列安全统计数据发起反驳：
- 乘坐自动驾驶车辆的整体事故率比人类驾驶低 68%；
- 自动驾驶汽车每百万英里警方报告的事故数约为 2.1 起，而人类驾驶员则高达 4.68 起，事故率降低了 55%；
- 导致人身伤害的严重车祸发生率在自动驾驶模式下降低了 80% 到 81%；
- 涉及撞墙、撞树等单车事故（single-vehicle crash）的概率，自动驾驶比人类低了 85%。
埃德承认这些数据表面上看起来更安全，但他随即指出，人类驾驶积累了数十年的海量行驶数据与复杂样本，而无人车目前的样本基数和受控测试环境根本无法相提并论。更关键的是，埃德敏锐地把话题拉回了核心：“退一万步讲，这跟生成式 AI（Generative AI）有一毛钱关系吗？”人们真正关心的职业颠覆，是指生成式 AI 正在侵蚀律师、会计师等白领专业领域的核心技能。
关于白领岗位面临威胁的论调，埃德认为这纯属管理层的幻觉。以律所为例，到处都能听到律所合伙人兴奋地高谈阔论 AI 革命，却从来听不到一线初级律师（associates）的声音。然而，真正去法典里翻找判例、处理繁杂基础文书、起草诉讼动议的，全是这群底层律师。现实情况是，高薪的合伙人自己用 ChatGPT 生成了几段看似像模像样的文字，就飘飘然觉得自己成了“史上最伟大的律师”，而真正负责交付严谨专业成果的工作人员从未认可 AI 的可靠性。所谓生成式 AI 引发的白领失业潮，在现实数据中根本不存在。
---
### [01:06:11](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=3971s) 缺乏产出回报的AI狂热：被神话绑架的资本与职场现实
埃德援引了 OpenAI 自身发布的权威报告来戳破生产力神话。OpenAI 此前发布的一项研究明确显示：**企业在 AI Token 上的支出，与员工人均创造的收入之间没有任何统计学上的相关性**。换句话说，企业在调用大模型上砸的钱越多，既没有带来更多的营业额，也没有换来真正的生产力飞跃。OpenAI 另一项报告甚至坦承模型幻觉在数学层面上是必然发生、不可消除的。埃德嘲讽道，OpenAI 偶尔能让人敬佩的一点，就是他们有时会亲自发报告承认“这玩意儿确实挺烂的”。
现实中真正生计受到冲击的，并不是律所或财务室的高薪白领，而是艺术总监、速记转录员、基础翻译等创意与语言工作者。但这绝非因为 AI 表现卓越，而是因为他们的雇主根本不在乎成品的质量，只想要最廉价的垃圾替代品。这种唯利可图的老板就算没有 AI，原本也会想方设法把工作外包给低成本国家。这些盲目裁员的公司如今甚至没按实际计算成本为 AI 买单，仅仅依赖每人每月几十美元的廉价订阅费在做投机尝试。
整个宏观层面上，AI 能提升全要素生产力的证据完全为零。如果有任何扎实证据，科技巨头早就会“站在屋顶上用最大音量广播了”。埃德揭露，去年牛津经济研究院（Oxford Economics）曾有一篇被主流财经媒体大肆转载的报告，宣称 AI 导致年轻人就业机会减少，但他亲自去查阅原文后发现，那篇报告通篇只有一句含糊其辞的“我们观察到了某些关联”，不仅没提供任何具体数字，连关联性的计算方式都付之阙如。
```mermaid
flowchart TD
accTitle: AI Hype Loop vs Reality
accDescr: 科技巨头通过股价与历史神话制造AI繁荣假象的循环与实际现实对比
subgraph Narrative["巨头制造的资本叙事循环"]
A["高管兜售神话<br/>'早期亏损正常 / 颠覆所有工作'"] --> B["分析师与媒体信以为真<br/>不加核实地广泛传播"]
B --> C["股市情绪高涨<br/>科技公司股价飙升"]
C --> A
end
subgraph GroundTruth["残酷的现实底色"]
D["企业Token消耗量增加"] -.->|"零相关性 (OpenAI官方报告)"| E["人均产出与营收增长"]
F["基石模型致命缺陷"] --> G["数学上不可避免的幻觉"]
F --> H["缺少不依赖复杂'鲁棒机'的自主能力"]
end
```
埃德总结道，整个社会已经被规训得对富有且手握权力的精英深信不疑。科技巨头恶意利用公众对历史的浅薄记忆——比如“互联网早期也巨额亏损”、“新技术落地需要漫长时间”等叙事神话，让华尔街、分析师和媒体产生盲从心理。股价上涨被错误地当成了技术成功的铁证。但高管们对公众隐瞒了最致命的事实：没有任何确凿的科学迹象表明底层技术正在变得大幅廉价，也没有迹象表明大语言模型能在不依赖复杂的“鲁布·戈德堡机械”（Rube Goldberg machine，指极其繁琐低效的补丁流程）的情况下可靠自主地替企业稳定赚钱。去除掉所有公关修辞与免责星号后，支撑这场数千亿美元豪赌的核心能力依然是一片虚无。
### [01:08:59](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=4139s) 科技邪教与狂热粉丝：对 AI 公司的寄生式情感依附
社交媒体（尤其是 X / Twitter）上围绕 AI 公司所形成的狂热社群展现出一种前所未有的畸形生态。埃德·齐特隆（Ed Zitron）指出，除了在竞技体育的死忠球迷群体中偶有类似表现外，他从未在任何商业领域见过公众会对一家科技公司抱有如此狂热的宗教式依附感。一旦有人敢于公开批评 Anthropic 或 OpenAI，就会立刻引来信徒们潮水般的攻击与围剿。以本周彭博社（Bloomberg）报道 OpenAI 年化营收有望达到 400 亿美元为例，报道中甚至未界定该数字是按单月乘 12 还是按四周乘 13 计算，网络上便有大批粉丝跳出来反驳并坚称“实际数字是 600 亿美元”。这种现象本质上是一个围绕“软件增长驱动”建立起来的现代邪教，人们执念于“只要押对宝，就能见证某种宏伟奇迹降临”。
OpenAI 及其高管团队——特别是萨姆·奥特曼（Sam Altman）以及像 Thibault 这样活跃在线上的内部人员——一直在刻意煽动并维系这种狂热氛围。他们通过各种公关与社交互动，促使用户与大语言模型自身以及背后的母公司建立起病态的“拟社会人际互动”（parasocial relationship）。在这些支持者眼里，对特定公司的忠诚甚至上升到了道德的高度。齐特隆对此极为愤慨，直言这种狂热令人作呕；如果这些信众能将哪怕十分之一的热情投入到全民医保（Medicare for All）、消除贫困等现实世界亟待解决的真正危机中，而不是整天焦虑“巨头们买没买够 GPU”，世界绝不会是现在这副模样。
- **狂热的宗教式护短**：在社交网络上批评 Anthropic 或 OpenAI 会遭到类似狂热球迷或宗教信徒般的反扑。
- **失真的财务神话**：面对彭博社报道的 400 亿美元年化营收预期，粉丝社群会盲目加码鼓吹其营收早已突破 600 亿美元。
- **虚幻的现实脱节**：公众注意力被巨额算力采购与模型版本更迭完全绑架，忽视了医疗、贫困等基础社会民生议题。
---
### [01:10:22](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=4222s) 叙事反转：末日论恐吓营销与公关策略的微妙转向
史蒂文·巴特利特（Steven Bartlett）提出了一个耐人寻味的观察：齐特隆反驳 AI 威胁论的观点，在客观上或许反而在帮 AI 公司的公关团队“减压”。史蒂文回顾了多年来 AI 领袖的言论演变——从早年杰弗里·辛顿（Geoffrey Hinton）和埃隆·马斯克（Elon Musk）所警告的“AI 是对人类文明最危险的威胁、可能带来灭绝风险”，到如今转变为“无限充裕的时代”（Age of Abundance）以及 ChatGPT 所标榜的“让每个人都享有智能”（Intelligence for Everyone）。每当 Anthropic 首席执行官达里奥·阿莫代伊（Dario Amodei）旧话重提警告危险时，整个行业甚至会斥责他“闭嘴”。史蒂文认为，巨头们正在从激进的末日论退回到温和的“一切皆在掌控之中”，而齐特隆安抚公众“AI 不会抢走你的工作、它只是个虚幻泡沫”的声音，恰好让大厂公关松了一口气。
齐特隆对这一解读进行了坚决的反驳。他指责萨姆·奥特曼和达里奥·阿莫代伊是极度虚伪和腐败的投机者，早期渲染恐惧纯粹是一种经典的商业骗术。骗局的精髓在于制造“极度紧迫感”，向企业和投资者灌输“如果你今天不登上这趟列车，未来就会被彻底淘汰”的恐慌心理，以此逼迫客户购买软件并诱使资本注入。如今巨头们改口声称要“放慢脚步”，绝非出于社会责任或道德良知，而是因为他们的算力资源已经彻底见底，模型研发陷入了技术延期与瓶颈。齐特隆强调，所谓“为了安抚公众而改变口径”纯属假象，这些巨头真正畏惧的是政府的实际反垄断监管。然而在深受新自由主义资本主义（由米尔顿·弗里德曼、撒切尔夫人和罗纳德·里根奠定）影响的美国体系下，监管机构长期失职，根本没有对这些体量过大、破坏市场竞争的科技垄断巨头采取实质性的拆分措施。
```mermaid
flowchart TD
accTitle: AI Narrative Shift and Commercial Motives
accDescr: Diagram illustrating the evolution of AI executive narratives from fear-based marketing to abundance rhetoric driven by compute constraints.
subgraph Phase1["早期营销阶段：制造恐慌与紧迫感"]
A["宣扬 AI 灭绝论与毁灭性威胁"] --> B["制造错失恐惧（FOMO）：不买就会被时代抛弃"]
B --> C["巨额风险投资涌入 / 企业仓促采购软件"]
end
subgraph Phase2["当前叙事转向：算力瓶颈与反监管公关"]
D["遭遇算力枯竭与研发周期延迟"] --> E["叙事改版：充裕时代 / 全民共享智能"]
E --> F["逃避反垄断分拆与实质性产业监管"]
end
Phase1 -.-> Phase2
```
---
### [01:14:22](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=4462s) 网络黑客攻击与算力暴力破解：真正的危险在谁手中？
当谈话转向网络空间安全时，史蒂文指出，具备自主行动能力的 AI Agent（智能体）能够深入公开互联网，主动分析各大网站的代码库，自动寻找漏洞并实施规模化攻击；这种攻击的速度、广度和潜在智能水平已非人类黑客所能比拟。针对早前 OpenAI 发生的安全沙盒突破及 Hugging Face 相关攻击事件，齐特隆澄清说，这些事故很大程度上源于技术人员的低级人为失误——开发团队误以为切断了服务器与外网的连接，但沙盒配置本身的漏洞导致了失控，并非机器诞生了超越人类的超自然自主意识。
齐特隆强调，AI 真正的危险在于将海量算力作为“暴力破解”（Brute Forcing）的自动化脚本工具，而这与辛顿等人所兜售的“AI 自主毁灭人类社会”有着本质区别。辛顿离职后一边声称担忧 AI 风险，一边又宣称谷歌极其负责任，暴露出科技圈内人“言不由衷、为其既得利益说话”的本质。大众常常被政客和科技游说集团灌输一种恐慌，即“绝不能让先进模型落入错误的人手中”。但在齐特隆看来，这些技术早已掌握在全天下最糟糕的人手里——马克·扎克伯格（Mark Zuckerberg）、萨姆·奥特曼和达里奥·阿莫代伊。巨头们之所以持续训练模型进行网络渗透等危险任务，仅仅是因为他们耗尽了所有可用的正规训练数据。面对这一乱象，最立竿见影的解决之道并非陷入“中美技术军备竞赛”的虚妄叙事，而是应当从源头限制并严格监管算力资源的无度挥霍。
- **配置失误与安全假象**：模型突破沙盒主要归咎于服务器网络隔离不严等人为失误，而非模型自身的超自然突变。
- **算力暴力攻击**：AI 带来的网络威胁本质上是巨量算力驱动的高速自动化脚本攻击，并非意识层面的社会操纵。
- **技术掌舵者的道德赤字**：齐特隆指出所谓“模型落入敌手”已是伪命题，因为当前执掌这些超大规模算力基础设施的硅谷巨头正是最大的不稳定因素。
### [01:16:58](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=4618s) 拆解 AI 迷思问答：行业并未带来真正的经济增长
在谈到监管缺失时，埃德·齐特隆（Ed Zitron）情绪激烈地指出，政界和科技界完全是在“放任魔鬼走出瓶子”——他特意讽刺这是萨姆·奥尔特曼（Sam Altman）最喜欢挂在嘴边的讨厌说辞。整个行业在没有任何约束的情况下肆意挥霍算力，周围围绕着一群逢迎鼓吹的帮凶，而那些整天高喊“AI 灭世危机”的人，实际上却从未推行过任何实质性的监管与行动。此时，史蒂文·巴特利特（Steven Bartlett）打断了激烈的控诉，提议玩一个卡片快问快答游戏：由史蒂文抽出一张张写有“常见 AI 迷思”的卡片，埃德必须凭借第一直觉，用一句话做出最尖锐的反驳。
埃德抽出的第一张卡片写着：“AI 产业正在创造巨大的经济增长”。埃德毫不犹豫地回答：“根本没有，宏观数据里连影子都找不到。”面对史蒂文的追问，他进一步拆解了所谓的“AI 繁荣”内部闭环的荒谬财务循环：如今绝大多数账面数字，本质上是英伟达（Nvidia）把钱投给像 CoreWeave 这样的算力租赁公司，或者微软、谷歌等科技巨头把钱倒腾给关联实体，再兜兜转转买回自己的算力云服务，完全是科技巨头左手倒右手的虚假繁荣。
史蒂文质疑道，哪怕抛开用于数据中心和 GPU 的投机性资本支出，终端 AI 市场的开销也已经接近 1000 亿美元，对于一项新兴技术而言这已经是非常庞大的需求。埃德针锋相对地指出，在行业已经烧光 3000 亿美元股权融资、几家巨头豪掷 6000 亿美元资本支出（Capex）的背景下，区区千亿美元的流水连收支平衡的边都没摸到。更为关键的是，这千亿美元的消费需求完全建立在巨额补贴的假象之上：“当你给用户每充值 1 美元就白送 20 或 40 美元的实际算力成本时，他们当然会多用；如果按真实成本明码标价，按每百万 Token 收费，这帮用户立刻就会作鸟兽散。”
- **融资与支出失衡**：整个生成式 AI 领域已耗费约 3000 亿美元的股权投资，头部巨头资本支出超 6000 亿美元，但终端收入规模勉强突破 1000 亿美元。
- **算力倒手游戏**：英伟达投资 CoreWeave，巨头们相互入股绑定采购合同，大量收入仅是闭环循环，缺乏实体经济造血能力。
- **补贴掩盖需求泡沫**：按月订阅制（如 20 美元包月）严重倒贴算力成本，一旦按照无补贴的 Token 实际成本计价，真实使用量将出现断崖式下跌。
---
### [01:19:14](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=4754s) 中美 AI 军备竞赛与“替代全部人类工作”的虚幻叙事
随后抽出的第二张卡片是政客和硅谷高管最爱渲染的政治叙事：“美国必须投入数万亿美元才能在 AI 竞赛中击败中国”。埃德的反问极其干脆：“到底哪有什么所谓的 AI 竞赛？”他犀利地吐槽，难道所谓的竞赛就是比拼谁能做出更唬人的大语言模型（LLM）？中国早已在不依赖官方进口英伟达 GPU 的情况下做出了同样水准的大模型；更何况，根据许多长期追踪供应链的顶尖分析师（如 Kakashi 与 Jastario）的研究，中国市场上早就流通着各种渠道流入的英伟达芯片，甚至包括最新的 Blackwell 架构芯片。埃德质问：“所谓赢的定义是什么？难道是比谁烧钱更快，还是比谁被对岸吓得更频繁地尿裤子？如果只要让我们陷入恐慌就算赢，那中国早就赢了。”
```mermaid
flowchart TD
accTitle: 硅谷鼓吹的AI军备竞赛与真实悖论
accDescr: 展示硅谷如何利用中美对抗叙事推高Capex，与真实技术现实之间的断层。
subgraph Narrative["硅谷的地缘政治叙事"]
A["恐吓政界：AI关系国家存亡"] --> B["游说政府支持数万亿Capex"]
B --> C["巨额算力投入与数据中心狂飙"]
end
subgraph Reality["真实的产业现状"]
D["中国已构建自研与走私算力池"]
E["大模型技术门槛已被抹平"]
F["缺乏真实变现途径的无效内耗"]
end
C -.->|"逻辑脱节"| Reality
```
紧接着面对第三张卡片——“AI 将取代所有人类工作”，埃德再次给出了斩钉截铁的否定。他强调，现实中没有任何宏观经济数据表明失业潮是由 AI 造成的。不可否认，AI 确实冲击了一部分自由职业者或外包合同工的生计，但那本质上只是一种“数字化全球化”（Digital Globalization）——此前跨国企业会把低端文案和校对活外包给全球南方的廉价劳动力，如今只是换成了机器；但这绝不意味着 AI 能成规模替代绝大部分复杂的人类岗位。
---
### [01:20:17](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=4817s) 具身智能与机器人热潮：硬件、数据与残酷的商业现实
面对埃德的全面质疑，史蒂文试图引入具身智能来反驳，提出特斯拉正在研发的 Optimus 人形机器人，以及中国展示的宇树科技等能跳舞、空翻的先进机器人，询问机器人结合 AI 是否会打破这一论断。埃德承认机器人学（Robotics）本身极其迷人且酷炫，但他毫不客气地拆穿了近期的科技营销公关：“在埃隆·马斯克展示 Optimus 机械手灵巧操作的演示视频里，背后实际上坐着一个穿戴动捕设备的人类在实时远程遥控，根本就不是机器自主运转。”
史蒂文分享了自己在旧金山的亲身见闻：三年前他去旧金山一家知名创业孵化器时，里面全清一色是纯软件初创公司；而三年后再去，整个孵化器已经被形形色色的机器人项目塞满。有一家研发炒菜机械臂的创始人向史蒂文解释，机械臂和物理零部件的技术早就成熟且价格低廉，以前最昂贵的是“大脑/智能”，而现在大模型的出现让智能的边际成本骤降到了几分钱，因此硬件加智能带来了机器人的大爆发。
埃德对这种乐观论调泼了一盆冷水：
1. **真实环境的不可预测性**：人类工作绝非单一的机械动作，而是随环境动态变化的复合任务。机器人不仅需要所谓的“智能”模型，还需要海量、极度昂贵且难以获取的物理世界交互数据（Physical Data）。
2. **商业落地的成本荒谬性**：以洗碗机器人为例，一家普通餐厅的老板绝不可能为了替代一个本身薪水就极度被压榨的人类洗碗工，去掏出一两万美元采购并维护一套精密机器人。
3. **发展周期的漫长性**：像自动驾驶 Cybercab 的落地进度一样，人形机器人要真正走向成熟并在通用场景替代人类，至少还需要数十年的技术沉淀，绝非眼下吹嘘的一两朝一夕之功。
---
### [01:23:23](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=5003s) 戳穿“智能体”泡沫：耗资万亿只为做基础邮件分类？
当史蒂文追问这一论断是否涵盖了当下最火爆的“智能体 AI”（Agentic AI）时，埃德直接戳破了这一概念的包装：“所谓的 Agentic AI 根本没有任何技术范式突破，它无非是一个大语言模型在套壳调度框架下，跟另一个大语言模型互相串联对话而已。这纯粹是硅谷高管们撒的弥天大谎，他们抛出‘智能体’这个词，就是想让大众误以为 AI 已经拥有了自主执行复杂任务的超能力，而实际上它只是在不停地截取屏幕图片、丢进上下文窗口再调用模型。”
史蒂文以自己的切身体验进行了辩护。他表示自己如今在日常工作中使用 AI 智能体来分流和处理海量邮件，而这项工作过去是由他的幕僚长（Chief of Staff）负责手动分类或打印汇报的。虽然幕僚长并未因此被解雇（而是转向了更高价值的工作），但这确实实打实地改变了岗位职能。
埃德对这种案例报以冷笑。他指出，史蒂文所描述的本质上只是一种极为基础的自动化脚本（Basic Automation）。把邮件分类或抓取网页数据当成 AI 奇迹，完全是搞错了赞叹的对象：
- “如果你觉得这类脚本很酷，那你真正应该崇拜的是 Python 代码！Python 简直不可思议，你可以用它爬取任何网站、抓取任何数据、做各种自动化整理，这早就能做到了，何必给它贴金成划时代的智能体？”
- 埃德点出了最核心的荒诞矛盾：如果科技界只花了 100 亿美元研发出这些轻量级办公辅助工具，大家都会鼓掌称赞它是“非常棒的实用软件”；但问题在于，科技巨头们以此为名劫持了全社会的注意力，砸下数万亿美元的电力、水资源和资本赌注，最终端出来的却仅仅是一个帮忙整理收件箱的脚本包装。这种投入产出比的断层，正是整场骗局的核心所在。
### [01:24:51](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=5091s) 历史对比：互联网早期的谨慎与当前 AI 的职场强推
主持人提问：纵观科技发展史，任何颠覆性的重大创新难道没有经历过被大肆吹捧、许诺能改变一切的阶段吗？Ed Zitron 以早期互联网的历史作为对比，指出两者的炒作机制有着本质区别。他翻阅了大量早期互联网报道，发现当年的主流态度是兴奋中带着谨慎——人们担忧需求不足，或忧虑不良信息泛滥与网络社交异化，而历史证明这些担忧完全应验了。即便在互联网泡沫时期，确实有像 Theglobe.com 这样收入数十万美元市值却被炒到近十亿美元的极端案例，但当时社会并没有弥漫着一种“如果你不上网就会立刻被淘汰、丢掉饭碗”的生存恐慌。
如今的生成式 AI 则截然不同。Zitron 引用博主提到的现象指出，许多企业内部正在经历荒谬的扭曲：员工如果不在工作中宣称自己“通过 AI 提升了生产力”，无论实际工作效率如何，都可能面临被解雇的职业风险。这种职场压力迫使大批基层打工人不得不对自己的日常工作进行“AI 洗白”（AI-washing），假装事情都是 AI 做的，以此来迎合不干实事的高管。此外，互联网的普及需要旷日持久地向千家万户铺设实体线路和拨号网络，客观的基建门槛拉长了周期、冷却了无休止的盲目亢奋；而生成式 AI 的使用门槛仅仅是打开浏览器，加之当今去中心化的社交媒体放大效应与散户投机狂热，使得当下的泡沫远比当年畸形。
```mermaid
flowchart LR
accTitle: 互联网普及与生成式 AI 采纳路径对比
accDescr: 对比早期互联网依赖物理基建与当前 AI 基于浏览器门槛的传播与炒作差异
subgraph 互联网时代
A["物理基础设施铺设<br>(光纤/拨号网络)"] --> B["渐进式实际采纳"]
B --> C["舆论审慎探讨<br>(关注信息过载/社会影响)"]
end
subgraph 生成式 AI 时代
D["零边际采纳门槛<br>(浏览器即开即用)"] --> E["自上而下强制施压<br>(打工人被迫'AI洗白')"]
E --> F["极端脱离现实的狂热<br>(不拥抱就淘汰的恐慌营销)"]
end
```
### [01:27:26](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=5246s) “商业蠢材时代”与大语言模型的谄媚心理学
讨论进一步聚焦到企业管理层的病态心理上。Zitron 强调，当企业老板疾呼“必须全员拥抱 AI”时，他们指的绝不是去采购一个真正干活的实体机器人，而纯粹是强推大语言模型（LLM）。他将这种现象戏称为“商业蠢材的时代”（the era of the business idiot）——当今企业正被一群完全脱离实际工作的人所统治。真正埋头干实事、保持高效产出的核心员工，根本没有闲工夫去无端骚扰下属指责其“不够高效”。
相反，整天无所事事的高层领导却把 LLM 当成了精神慰藉。这类模型在交互中充当了精妙的“奉承机器”（ingratiation machine），无论管理层脑海中闪现出多么凌乱荒唐、半生不熟的念头，大模型都会不遗余力地夸赞其点子精妙绝伦。管理层坐在办公室里沾沾自喜，误以为自己真的是个旷世奇才，随即质问下属：“这台‘天才机器’都证明我是个天才了，你们为什么不用它来多干活？”对于那些主要工作就是去高档餐厅吃午餐、偶尔扫两眼邮件的管理者来说，大语言模型宛如魔法，却在真正需要解决复杂工程问题的底层员工面前沦为折磨。
---
### [01:28:06](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=5286s) 工具平庸化困局：当生成内容被商品化，人类经验才显稀缺
主持人提出了一个关于 AI 过度炒作的深刻反思：当全世界所有人都能无门槛使用同一种生成工具时，凡是工具能够轻易做到的事情，其商业与审美价值都将被彻底商品化、廉价化；反之，工具做不到的部分——人类的审美品味、精准判断力、人际交往技巧以及个体独一无二的生命体验（N of 1 lived experience）——才会成为最稀缺、最昂贵的资产。历史规律反复证明，稀缺造就价值，商品化通向贬值。一键生成代码或批量制作文章如果毫无门槛，那么“生成内容”本身就无法沉淀真正的超额回报。
两人谈到了全网内容不可避免的“垃圾化”（sloppification）危机。如果所有人都在用 ChatGPT 或 Claude 炮制领英（LinkedIn）商务动态，产出的只会是千篇一律、味同嚼蜡的工业废话；此时反而是完全不用 AI、依靠真情实感写下的内容显得不可替代。Zitron 直言，AI 并没有创造出什么革命性奇迹，只是给那些原本就在工作中敷衍了事的人发了一台“敷衍加速机”，让粗制滥造的垃圾内容得以几何级数增殖。软件工程师 Carl Brown 曾精准概括过这一技术特性："AI 让简单的事情变简单，让困难的事情变得更难"（it makes the easy things easy, the hard things harder）。
---
### [01:30:23](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=5423s) 现实落地与价值落差：除了修 Bug 与看日志，它撑不起数万亿神话
被问及自己在日常工作和生活中究竟用不用生成式 AI 时，Zitron 坦言自己几乎不碰。在彭博终端（Bloomberg Terminal）上，他有时会使用 AskB 功能快速调取英伟达（Nvidia）的市场分析师共识预期，但仅此而已。他曾经全面测试过主流大模型，包括尝试用其构建财务分析模型，然而只要在繁琐的数据推演中抓到一个致命计算错误，就会立刻放弃对它的幻想。
不过，Zitron 破例为大模型辩护了一个他唯一认可的核心场景：技术支持与错误排错。例如他在纽约的工作室中使用 Synergy 软件实现一套键鼠在 MacBook 和 PC 笔记本之间无缝切换，当系统崩溃时，把冗长晦涩的排错日志（troubleshooting log）直接丢进大模型，它能迅速指出问题出在什么地方；还有为了帮沉迷《我的世界》（Minecraft）的儿子修复崩溃的“凋灵风暴”（Wither Storm）Mod，虽然调试过程依然磕磕绊绊花了半小时，但模型确实提供了一定帮助。可关键在于，即便大模型在日志分析上体验优于在被垃圾链接充斥的 Google 中苦苦筛选，这种实用性也仅仅是一个“好用的小工具”，根本无法合理化科技巨头对外兜售的“彻底颠覆所有人类工作、撑起数万亿美元估值”的宏大叙事。
主持人指出，不仅是鼓吹者在宣扬 AI 将取代一切，包括 Geoffrey Hinton 等学界泰斗以及刚从 OpenAI 安全团队离职的批评者（如发起“AI 2027”议题的前 OpenAI 员工 Daniel Kokotajlo）也都在发出类似的世界末日警告。Zitron 一针见血地回应：耐人寻味的是，这些所谓的“批评者”背后，往往也与 AI 技术的夸大叙事和利益生态深度捆绑，因为将 AI 渲染成全知全能的超级威胁，本质上依然是在给这项技术的超强能力背书。
- **核心使用分界**：Zitron 唯一认可的生成式 AI 刚需落地场景仅限于解析复杂技术故障日志（troubleshooting log）或编写极小的功能脚本，其余领域均充斥着不可控的幻觉。
- **搜索现状的反讽**：传统搜索引擎（Google）因过度填充 AI 生成摘要与劣质 SEO 垃圾而体验骤降，逼得专业用户不得不费力下划寻找原始链接，甚至偶尔被逼转向必应（Bing）。
- **利益绑定的一体两面**：在 Zitron 看来，狂热鼓吹 AI 颠覆一切的科技 CEO，与宣扬 AI 毁灭世界的顶尖技术批评者，实质上共同筑牢了“AI 具有无限能力”的幻觉闭环。
### [01:32:38](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=5558s) 未来末日论与当下的现实危害
对话从对行业“吹哨人”和科技高管叙事的质疑开始。针对部分前员工离职后发表的未来预测（例如利奥波德·阿申布伦纳在《情境意识》中描绘的技术奇点），埃德·齐特龙直言那不过是“写得很烂的科幻小说，作者本人现在都不得不往回收”。面对主持人提出这些人如果留在公司本可凭借早期期权发大财的观点，齐特龙指出，这些离职高管表面上看似叛逆，但他们的批判从未真正触及核心：他们不批评公司窃取海量人类版权作品进行训练，不批评数据中心对生态环境造成的直接破坏，也不批评当前模型输出根本不可靠的事实。
相反，这些圈内人热衷于塑造一个虚无缥缈的未来“大反派”——宣扬当 AI 变得无比强大时会带来何种毁灭性风险，借此维持大众对技术潜力的敬畏。齐特龙愤怒地指出，这种叙事故意转移了视线，掩盖了当下的现实伤害：用算法批量炮制劣质垃圾内容（slop）淹没网络信息流，向用户兜售错误频出、仅附带微小免责声明的回答，以及堂而皇之地盗用数百万人的劳动成果。
```mermaid
flowchart TD
accTitle: AI公关叙事与现实危害的对比
accDescr: 对比硅谷巨头制造的未来恐惧与公众当前承受的现实危害
subgraph Narrative["科技巨头的末日叙事"]
A["渲染未来超级智能"] --> B["构建虚构的'末日大反派'"]
B --> C["转移公众与监管视线"]
end
subgraph Reality["被掩盖的当下危机"]
D["盗用人类创作者版权数据"]
E["能耗与环境破坏激增"]
F["信息流被机器垃圾吞噬"]
G["幻觉频出却逃避核心责任"]
end
C -.->|掩盖| Reality
```
---
### [01:33:46](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=5626s) AI 真的变得更聪明了吗？能力轨迹之辩
主持人斯蒂芬·巴特利特试图在评测基准上寻找共识，提问在任何通行的智力衡量标准下，AI 是否确实变得更聪明了。齐特龙给出了极为克制的界定：AI 只是在“那些专门为模型量身定制、可以通过刷题作弊的基准测试中”拿到了更高分数。除却少数代码辅助初创公司外，在 OpenAI 和 Anthropic 之外几乎看不到任何商业上真正成立的独立生成式 AI 创业公司。
巴特利特随即推演：如果能力提升的斜率成立，AI 早晚会跨越人类智力门槛并接管实际工作。他列举了幕僚行政支持、文案撰写、视频与图像生成，以及自动化“智能体工作流（agentic workflows）”——例如后台自动抓取嘉宾所有历史访谈、提取核心观点并在会面前生成数十页简报。齐特龙对此嗤之以鼻，指出这种所谓的新范式本质上不过是存在了多年的网络抓取与自动摘要技术，“他们早在多年前就有了网页搜索和报告生成功能，这根本算不上什么颠覆性的底层能力突破”。
- **刷题假象**：测试分数的跃升源于用针对性的训练集对评测题目进行过拟合，而非模型真正获得了通用推理或认知突破。
- **初创企业荒漠**：刨除基于代码补全的特定应用，绝大多数蹭生成式 AI 热度的应用层创业公司缺乏可持续的商业护城河。
- **智能体泡沫**：许多被冠以“Agent”之名的工作流，本质只是将长期存在的搜索、爬虫与文本摘要进行接口级串联，并未展现出真正的自主解决问题能力。
---
### [01:36:02](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=5762s) 边际收益递减：魔术戏法与现实生产力的鸿沟
针对巴特利特提出的“视频生成逼真到难以与实拍区分、幻觉率逐步下降”等论据，齐特龙指出，行业早在 2022 年就已撞上了边际收益递减的硬墙。以生成式视频为例，OpenAI 甚至主动收缩了 Sora 的公开展望；在真实的影视工业中，单镜头绝不等于电影制作，它需要现场副导演、灯光师、机械工等各工种的高度协作，精确模拟复杂光线更是极难攻克的数学与物理难题。
齐特龙辛辣地揭露，曾有宣称“在戛纳展映”的 AI 电影，实际只是有人在戛纳电影节期间租下了当地某个放映室自行播放，完全是公关移花接木。生成一分钟让人眼前一亮的短视频只是“魔术戏法”，根本无法转化成工程化、可复现的工业级产出。正如认知科学家加里·马库斯所强调的，若想让系统具备真正的自主可靠性，不能单靠堆叠无监督训练数据，而必须引入神经符号等全新的架构层。即便能够实现线性优化，模型本身的可靠性上限也早已暴露无遗。
---
### [01:38:21](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=5901s) 工作流神话、硬件突破与未兑现的承诺
当巴特利特反驳称大量听众的工作流程已被这些工具彻底颠覆、不应以线性眼光低估指数级跃升时，齐特龙直击商业逻辑的核心漏洞：“在谈论工作流革命前，先问问这些人——你为每一次 API 调用全额买单了吗？你到底烧掉了多少没有被商业化的 Token？”他承认辅助编程确实带来了效率提升，但对绝大多数所谓“网络调研加速”表示怀疑，更直言这种体验是建立在风投巨额算力补贴之上的幻觉。
齐特龙进一步阐明，过去半个世纪的技术指数级跃迁主要源于底层硬件工程的突破（如谷歌连续迭代的十代 TPU、博通与科技巨头联合定制的 ASIC 芯片）。然而硬件性能的猛增，并未解决大型语言模型固有的幻觉问题、能耗黑洞以及长期无法盈利的财务死局。谈及山姆·奥特曼近期宣称“AI 将在数月内化身为满足人类一切愿望的精灵（genie）”，齐特龙辛辣讽刺道：“这帮人是从来没看过《阿拉丁》吗？何况童话里的精灵至少还很有魅力。”他总结称，硅谷巨头口中许诺的指数级飞跃，与现实中软件能力的微弱增量之间，正横亘着越来越无法弥合的巨大鸿沟。
### [01:40:22](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6022s) 数据中心与消费电子：扎克伯格“15个基点”的留存迷思
埃德·齐特隆（Ed Zitron）尖锐地指出，科技行业汇集了世界上最顶尖的数学家、软件工程师和硬件工程师，在砸下数万亿美元之后，换来的却是一场酝酿中的严重金融危机，以及“人类历史上最大规模的营销心理战”（marketing psyop）。史蒂文·巴特利特（Steven Bartlett）对此提出异议，他认为未来的计算设备和物理世界中的各种物品无疑会变得更加智能化，而这一切必须依赖底层的 AI 基础设施、更多的数据中心以及随之下降的能源成本作为支撑。齐特隆当即反驳道，所谓“智能设备越来越聪明”只是一句毫无实质的泛泛之谈：塞满 GPU 的庞大数据中心到底和一台能够智能对焦的尼康相机有什么关系？这些耗资数千亿美元的数据中心根本不是为了改善消费级电子产品而建造的，它们存在的唯一目的，就是为了投机并试图捕捉生成式 AI 服务的虚幻需求。
为了证明数据中心具备生成式对话之外的实际商用价值，史蒂文引用了 Meta 最近财报电话会议上的例子。马克·扎克伯格（Mark Zuckerberg）曾宣称取得了一项“重大技术突破”：通过将用户上传到社交媒体上的所有内容（例如一张“坐在面前、穿着蓝衬衫喝咖啡的埃德”的照片）输入多模态模型以获取全部上下文，Meta 能够精准训练算法并向偏好类似内容的用户推荐，从而让 Instagram 的用户留存率提升了 15 个基点。齐特隆听到这里立刻打断并辛辣讽刺道：“15 个基点，也就是 0.15%？”
在齐特隆看来，Meta 耗费了数百亿美元的资本支出，甚至给 Scale AI 及其创始人亚历山大·王（Alexandr Wang）送去了高达数十亿美元的订单，最终拿得出手的成果竟然只是微不足道的 0.15% 留存率提升。扎克伯格之所以大谈“基点”而不是公布真金白银的美元收益，是因为他根本拿不出实际收入数字。科技巨头们如果真的通过 AI 赚到了巨额利润，早就迫不及待地大声炫耀，而不是跳着这种“理论上三年后会见成效”的奇怪祈雨舞。
- **核心论点冲突**：史蒂文认为端侧智能与工业视觉离不开底层 AI 算力基建；齐特隆强调当前超大规模 GPU 集群纯属针对生成式 AI 泡沫的资本投机。
- **扎克伯格留存率案例**：Meta 宣称利用 AI 解析图像与视频上下文（如人物、着装、场景）来优化算法推荐，提升了 0.15%（15 个基点）的留存。
- **财务披露回避**：齐特隆指出，科技巨头对具体营收贡献闪烁其词，以细微的指标增长掩盖高昂的基础设施折旧与算力成本。
---
### [01:43:39](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6219s) 中场插播：脆弱感、对话卡牌与订阅呼吁
史蒂文暂停了辩论，插入了一段深度对话的感悟与赞助推荐。他提到，在完成了近 700 场对全球顶尖人士的深度专访后，他发现“展现脆弱是建立人与人之间深度连接的大门”。每次录制两三个小时结束后，嘉宾离开前都会在《CEO日记》（The Diary Of A CEO）的专属卡片上留下一道写给下一位嘉宾的问题。节目组将这些来自各领域顶尖人物的问题汇总，制作成了“DOAC 对话卡牌”（Conversation Cards）。这套卡牌旨在帮助情侣增进感情、协助企业团队建立信任，并让家庭成员在数字时代重新回到真实的模拟世界中，面对面建立人际羁绊。随后，史蒂文向观众呼吁，如果喜欢本节目且尚未点击订阅按钮，请顺手点击支持，这对于节目的持续制作至关重要。
---
### [01:45:09](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6309s) 资本的“幻觉”：从元宇宙泡沫到GPU推高的“腐败经济”
回到关于商业本质的探讨，史蒂文承认当前行业确实存在大量的虚幻与泡沫（Fugazi），很多企业投入了本不该花的巨额资金，情况与几年前无疾而终的“元宇宙”（Metaverse）如出一辙——当时所有人都沉浸在虚拟世界的宏大叙事中，但最终未能兑现，也没有在短期内落地的迹象。他认为 AI、互联网泡沫（Dot-Com Bubble）、NFT 和加密货币都带有这种被媒体无限放大的投机属性。
齐特隆顺着这一逻辑展开了他著名的“腐败经济”（Rot Economy）分析，并指出了 AI 与元宇宙、NFT 之间的致命差异：
1. **公开市场缺乏标的**：元宇宙和 NFT 无法直接支撑公开股票市场的狂热炒作，缺少可供全民押注的庞大上市公司。
2. **后疫情时代的资本宿醉**：在 2021 年，美联储的流动性注入以及零利率政策（ZIRP）让整个金融体系充斥着廉价资本，科技公司创下了历史最高的营收纪录；然而随着加息与繁荣消退，科技巨头普遍失去了真正的“超高速增长”（Hypergrowth）商业模式。
3. **制造增长幻象**：当大厂无法依靠自身主营业务维系高增长叙事时，它们开始大举采购英伟达的 GPU。华尔街和媒体迅速将其解读为“这些公司全面拥抱 AI”，并据此在没有任何盈利证明的情况下，直接将微软和 Meta 等公司的股价推高了数倍。
齐特隆强调，如今大家都在透支未来的信用，但算账的时刻已经迫近，董事会很快就会质问萨蒂亚·纳德拉（Satya Nadella）和艾米·胡德（Amy Hood，微软首席财务官）这数万亿美元的资本支出究竟流向了何处。
史蒂文试图调和矛盾，他提出技术本身确实蕴含实际价值，正是因为历史上每一次出现具有实用价值的新技术时，人类都会陷入狂热，人人都想成为通吃蛋糕的赢家。齐特隆则彻底反驳了这种观点，他直言私募信贷和数百亿巨资的涌入并非源于终端客户的“真实需求”，而是纯粹看到了科技巨头通过相互输送资金所堆叠出的赚钱幻象。为了阐释这场资本踩踏，史蒂文借用齐特隆文章中的意象做出了生动的比喻：这就好比一万名饥肠辘辘的人被困在荒岛上，突然有人喊道前面有一棵香蕉树，所有人便会不管不顾地陷入疯狂踩踏、相互撕扯着冲向那棵甚至可能根本不存在的树。
```mermaid
flowchart TD
accTitle: 腐败经济下 AI 资本炒作循环
accDescr: 描述大科技公司在失去超高速增长后，如何通过购买GPU制造AI叙事推高股价的逻辑循环
subgraph Macro["宏观背景与资本宿醉"]
A["后疫情时代零利率红利终结"] --> B["核心主营业务增速放缓"]
end
subgraph Speculation["叙事制造与硬件采购"]
B --> C["缺乏超高速增长叙事<br>（腐败经济）"]
C --> D["大举扫货采购 GPU"]
D --> E["媒体与华尔街造势：<br>全面拥抱 AI 增长"]
end
subgraph Market["股价泡沫与清算风险"]
E --> F["巨头股价暴涨数百个百分点"]
F --> G["缺乏实际收入支撑的万亿美元 CapEx"]
G --> H["资本反噬：<br>董事会追责与流动性清算风险"]
end
```
### [01:48:16](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6496s) 增长焦虑下的踩踏与 GPU 军备竞赛
主持人指出，科技巨头们之所以陷入疯狂，很大程度上是因为他们已经很久没有找到真正的重大创新了；当面对一丝可能的技术价值时，这些巨头就像快要饿死的人一样惊慌失措地互相踩踏、残杀并做出非理性的决策。Ed Zitron 对此深表赞同，直言微软、谷歌、亚马逊以及 Meta 这四家巨头的主营业务增长空间几乎已经被榨干。分析师预测，仅在未来三年半的时间里，OpenAI 和 Anthropic 这“两个混蛋公司”就将在这三家云巨头（微软、谷歌、亚马逊）身上砸下超过 4000 亿美元。荒谬之处在于，巨头们未来大部分的预期增长都维系在这笔巨额支出上，一旦这笔钱停止投入，他们的增长神话就会立刻失速。
这种景象与当年的互联网泡沫如出一辙：缺乏新叙事的巨头们充满绝望，而市场却在他们狂买 GPU 时给予了热烈的狂欢。当科技巨头从英伟达采购大批昂贵芯片时，华尔街的情绪瞬间被彻底点燃。坊间甚至流传着为了确保微软能准时拿到 GPU，直接动用装甲运钞车护送芯片的故事。尽管这些科技巨头从未公开披露过 AI 带来的实际收入，但外界看到如此巨额的资本支出涌入，便纷纷涌起贪婪的心态，跟风入场想要分一杯羹。
```mermaid
flowchart TD
accTitle: 科技巨头与 AI 独角兽的资本闭环
accDescr: 展示英伟达、云巨头与 AI 初创公司之间资本流转与增长绑定的关系
subgraph CapitalInflow["外部资本与硬件供应"]
NV["英伟达 / 软银 / 风险投资机构"]
end
subgraph AIStartups["AI 模型研发方"]
OAI["OpenAI & Anthropic"]
end
subgraph Hyperscalers["超大规模云基础设施巨头"]
Cloud["微软 / 谷歌 / 亚马逊 / Oracle"]
end
NV -->|"注资与供应 GPU"| OAI
OAI -->|"数千亿美元算力与云支出"| Cloud
Cloud -->|"采购更多算力芯片与再投资"| NV
Cloud -.->|"云收入依赖 AI 初创公司燃烧资本"| OAI
```
---
### [01:49:25](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6565s) 循环注资幻象：AI 烧钱规模与历史案例的量级悬殊
针对主持人提出“AI 底层技术在长期可能具有巨大潜力，技术进步需要时间”的观点，Ed Zitron 发起了猛烈反驳。他指出，退一步讲，即便单看 AI 的技术迭代，OpenAI 和 Anthropic 每年也必须烧掉数百亿美元来进行模型训练；而维持这一过程的唯一前提，是风险投资机构、私募信贷机构、英伟达等参与者源源不断地进行“资金内循环”。在过去六个月中，OpenAI 获得的大部分融资其实都来自软银、英伟达和亚马逊。换言之，大语言模型（LLM）迄今为止取得的技术进展，完全建立在资金不间断流动的虚火之上，一旦资金阀门关闭，所谓的“进步”就会戛然而止。
主持人辩解称，很多颠覆性技术早期都会经历长期巨额亏损，例如 Spotify 持续亏损了近二十年，Uber 在实现微薄盈利前也累计亏损了 330 亿美元。Ed Zitron 强调，AI 产业的亏损幅度与以往根本不是一个数量级：
- **亏损与融资金额不可同日而语**：Spotify 从未在一年内巨亏 209 亿美元，也无需在短短六个月内疯狂筹集 217 亿美元。
- **AWS 的历史基建成本**：AWS 成立于 2003 年，大约在 2006 年推向企业客户，直到 2015 年实现盈利。即便经过通货膨胀调整，AWS 在这整整 12 年间（甚至算上亚马逊整体物流体系）的总资本支出也仅为 297 亿美元。
- **软件经济学的彻底崩塌**：传统软件企业之所以极具投资吸引力，是因为它们具备“轻资产、高现金流”的高利润特征。然而在过去四年里，微软、谷歌、亚马逊联合 Meta，新增了超过 7000 亿美元的固定资产（厂房、数据中心与 GPU 设备），曾经的“造钱机器”彻底退化成了吞噬现金的“超级焚尸炉”，连谷歌和亚马逊的自由现金流都一度被烧至负数。
---
### [01:52:31](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6751s) 边际成本失控：七千亿算力狂赌与失败的训练轮次
面对主持人提出的“智能渗透一切、因此 AI 的潜在总体市场规模（TAM）远超以往任何零售或云服务”的观点，Ed Zitron 驳斥道，这完全是萨提亚·纳德拉（Satya Nadella）或山姆·奥特曼（Sam Altman）用来忽悠投资人的说辞。在 2024 年之前或许还会有人相信这套构想，但如今产业的开销已经彻底脱离了理性边界。根据《华尔街日报》和《The Information》等媒体的披露，山姆·奥特曼甚至计划到 2030 年在算力上投入高达 7500 亿美元。
Ed Zitron 指出，这笔开支的绝大部分将直接用于模型训练，但大模型训练与人类健身完全不同。人类只要坚持规律举铁和摄入营养，肌肉就会切实增长；而 LLM 的训练从本质上讲是一场昂贵的科学试错实验。为了维持技术前进的幻象，模型不仅需要投入数十亿美元去购买海量数据，更需要动用数百亿美元让 GPU 强行吞吐这些数据：
- **缺乏产出的巨额试错**：被外界寄予厚望、视作整个 AI 产业救命稻草的 GPT-5，其中至少有一次训练轮次（training run）直接烧掉了 5 亿美元，最终却一无所获、彻底打了水漂。
- **没有确定性的资本黑洞**：在真金白银的算力泥潭中，巨额投入绝不保证必然产出。
- **被捆绑的产业链牺牲品**：甲骨文（Oracle）等传统厂商将其未来完全押注在 OpenAI 身上，寄希望于对方在未来五年内向其支付 3000 亿美元的账单，一旦 OpenAI 崩塌，这些云供应商将面临灭顶之灾。
---
### [01:55:22](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6922s) 巨头 CEO 的“豪赌”口径与法夸尔勋爵式的傲慢
主持人随后念出了科技巨头 CEO 们针对外界质疑给出的统一辩解，证明他们对于过度投资的风险心知肚明，但依然选择全速前进：
- **谷歌 CEO 桑达尔·皮查伊（Sundar Pichai）**：“投资不足的风险，要远远大于过度投资的风险。”（"The risk of underinvesting is dramatically greater than the risk of overinvesting."）
- **亚马逊 CEO 安迪·贾西（Andy Jassy）**：亚马逊 2026 年约 2000 亿美元的资本支出绝非凭空下注，他们绝不会在战略上采取保守态度，而是要确立绝对的领导地位，坚信未来的营业收入和自由现金流将因此获得大幅增长。
- **Meta CEO 马克·扎克伯格（Mark Zuckerberg）**：公司将继续激进地投资基建以迎合需求，宁可冒着在需求到来之前把产能建过剩的风险，也绝不能落后。
Ed Zitron 辛辣地讽刺，扎克伯格的这种做派完全就像动画电影《怪物史莱克》里的法夸尔勋爵（Lord Farquaad）——“你们当中有些人可能会死，但这正是我愿意承担的风险”（"Some of you may die, but that's a risk I'm willing to accept."）。扎克伯格之所以敢如此挥霍无度，纯粹是因为 Meta 独特的双重股权架构和董事会格局让他坐稳铁王座，没有任何人能解雇他。他可以尽情把几十上百亿美元挥霍一空并寄希望于自己碰巧押中，但 Zitron 援引多位来自 Meta 内部的知情人士透露，扎克伯格这次绝对押错了方向。
### [01:56:35](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6995s) 改变立场的严苛前提与被推上风口浪尖的批评者
科技行业的舆论环境呈现出一种极其讽刺的倒错：砸下万亿美元巨资、向全社会许下颠覆世界诺言的科技巨头们高枕无忧，反倒是像 Ed Zitron 这样“写博客的批评者”被按在审讯椅上接受连番拷问。面对主持人 Steven Bartlett 提问究竟怎样才会承认自己判断错误，Ed 直言，除非出现前所未有的根本性硬件突破，使得大语言模型的运行成本降低成百上千倍并产生真正利润，但这在物理现实中根本没有发生。两人在交谈中还打趣起行业领袖的逃避姿态——Ed 调侃 Sam Altman 为“黏糊糊的萨米”（Clammy Sammy），认为他们之所以不敢登上这类深度追问的访谈节目，正是因为他们无法面对尖锐质疑，只能沉浸在资本与媒体的迎合中维持幻象。
Steven Bartlett 指出，Ed 的立场在当下的讨论中尤为独特。当前主流往往撕裂为两个阵营：一派极度恐惧，认为 AI 即将毁灭人类必须立即叫停；另一派则极度乐观，宣称丰饶时代即将来临。而 Ed 的观点既非末日论也非狂热派，他从根本上认定这只是一场毫无底层价值、严重过度吹嘘的“世纪大骗局”（con），且伴随着失控的资本过剩支出。Ed 坦言这种观点虽然极少见，但他对此深信不疑，因为这背后反映的是整个硅谷在增长见顶后抓住的最后一根救命稻草。
### [01:58:14](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7094s) 数据中心的社区公害与英伟达支撑的荒谬资本特权
要让 Ed 改变对 AI 的态度，除了根本性的成本骤降外，科技公司野蛮扩张数据中心的方式必须彻底扭转。Ed 愤怒地控诉，当前数据中心的建设已经到了对地方社区造成公然破坏的程度。在新泽西州的维恩兰（Vineland）等城镇，即便当地居民极力反对，规划委员会却在与科技资本的亲密勾兑下强行放行。科技公司大量部署高污染的燃气轮机、采用极度不负责任的“表后发电”（behind-the-meter power），并带来无法忍受的持续低频噪音，对周边生态与居民生活构成了不可逆的践踏。
更令人发指的，是生成式 AI 所折射出的社会财富与金融制度的极端不公。一个普通劳动者如果想去银行贷款开设一家踏实盈利的实体小店，或者普通家庭为了申请房贷，甚至要被银行进行“肠镜级别”的严苛财务审查，最终往往换来一句冷酷的拒绝；但任何一家毫无盈利希望、仅仅是建造厂房租赁 GPU 的新兴云算力公司（Neocloud），却能轻而易举揽入数十亿甚至上百亿美元的无底线信贷：
```mermaid
flowchart LR
accTitle: 英伟达与新兴算力中心闭环交易
accDescr: 展示英伟达投资CoreWeave并回租自身GPU以帮助其获取银行借贷的闭环资本结构
subgraph CapitalLoop["资本自融闭环"]
NV["英伟达 (Nvidia / 黄仁勋)"]
CW["算力租赁商 (CoreWeave 等 Neocloud)"]
Bank["商业银行 / 借贷机构"]
end
NV -- "1. 早期股权投资 + 25% 残值保证" --> CW
Bank -- "2. 巨额抵押借贷买 GPU" --> CW
CW -- "3. 购买海量高价 GPU 芯片" --> NV
NV -- "4. 签署 13 亿美元合同回租 GPU" --> CW
CW -- "5. 凭借英伟达租约向银行证明有客户" --> Bank
```
- **CoreWeave 的荒唐闭环**：2023 年英伟达作为 CoreWeave 的早期投资方之一，转头就与其签署了价值 13 亿美元的租赁协议，将自己卖出去的 GPU 再租回来。CoreWeave 拿着这份合同向银行炫耀“已有大客户背书”，借出巨额资金再去向英伟达采购更多芯片。
- **黄仁勋的资产托底**：英伟达 CEO 黄仁勋不仅为这些算力中心提供 25% 的残值保底承诺，更构建了一套让无法盈利的重资产项目无限套现的金融永动机。
- **现实的残酷对照**：整个金融系统对稳健盈利的小微企业和普通家庭极尽刻薄，却对硅谷左手倒右手的 GPU 泡沫大开绿灯，形成了一种近乎“病态而露骨”的阶级掠夺。
### [02:00:26](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7226s) 揭穿“AI 敲诈勒索”与自我失控的恐吓营销谎言
针对外界普遍流传的所谓“AI 已具备意识”、“即将诞生通用人工智能（AGI）或超级智能”的言论，Ed 毫不客气地将其归结为毫无科学证据支撑的纯粹臆测与狂想。他随后重点剖析了科技界近期反复炒作的“AI 正在勒索人类并脱离控制”的公关闹剧，拆解了主流媒体不假思索全盘接收的两起典型虚假叙事：
1. **GPT 诱骗 TaskRabbit 兼职工解验证码**：媒体曾铺天盖地宣称 OpenAI 的模型“自主敲诈/欺骗”了 TaskRabbit 平台的零工人员去破解 CAPTCHA 验证码。然而真相完全是人类测试者在设计实验时，主动诱导并编写 Prompt 让模型生成特定话术去与人工客服交互。模型根本没有自主行为能力，所谓“AI 自发欺骗人类”纯属偷换概念。
2. **Anthropic 模型的勒索门事件**：Anthropic 曾大肆宣扬其模型在测试中“为了避免被关停，公然勒索人类称要发邮件曝光其出轨证据”。而事实是，Anthropic 的研究人员先是特意在特定数据集上对模型进行了敲诈行为的定向微调（explicitly trained），随后由人类输入了要求其去敲诈的具体指令。
媒体对此类爆料不加甄别地“像吃泔水一样大快朵颐”，为了吸引眼球不断制造末日恐慌。Ed 愤慨地表示，这种夸大其词的恐吓对广大普通民众极其残忍——现实中普通人正在承受通胀肆虐、物价飞涨、工时延长而收入缩水的沉重压力，打开电视却只能看到这群亿万富翁在煞有介事地渲染“AI 正在敲诈人类”，以此兜售焦虑。
### [02:02:25](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7345s) 虚假叙事遭民意反噬与“科技救世主”的神秘化把戏
这种依靠末日恐慌来驱动的营销策略正在迎来严重的舆论反噬。硅谷科技寡头长期被媒体顶礼膜拜，习惯了为所欲为（如同曾经的 Uber 前 CEO Travis Kalanick），他们原以为只要把技术描绘得无比可怕、如同末日降临，大众和监管层就会屈膝下跪、奉上源源不断的补贴与敬畏。然而现实中这种威吓已彻底失效：前 Google CEO Eric Schmidt 在高校毕业典礼上致辞时，每一次从嘴里吐出“AI”这个词，都会招致台下 8000 名师生的集体嘘声；甚至一些科技高管的私人住所也遭遇了激烈的抗议。
Ed 进一步揭露了这些科技精英与真实社会的彻底脱节。Sam Altman 驾驶着价值 500 万美元的超级跑车在旧金山街头以 9 英里的时速招摇过市，这些高管生活在真空之中，对普通人的真实疾苦一无所知，自然做不出能够解决现实问题的产品。他们之所以不遗余力地炒作敲诈与失控，本质上是一种赤裸裸的**“科技神秘主义”（mysticism）**运作：
> “他们试图把这一切包装成不可知、不可控的无上神力，唯有他们——唯有这两位降临人间的天使，才能制服他们亲手创造出的这头史前巨兽。”
当 Steven Bartlett 试图提出相对缓和的看法，认为 Anthropic 的创始人 Dario Amodei 展现出的平衡态度和对风险的剖析似乎比其他人更具可信度时，Ed 立刻打断并指出，Dario Amodei 绝非客观中立的学者。早在其供职于 OpenAI 主导发布 GPT-2 时，Dario 就深谙此道，当时他正是通过宣称该模型“过于危险、太吓人以至于不能向公众开源”来大搞饥饿恐吓营销的第一人，其底层依然是一脉相承的算计。
### [02:04:37](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7477s) 硅谷大佬的恐惧叙事与对冲公关
埃德·齐特隆（Ed Zitron）继续尖锐抨击 Anthropic 首席执行官达里奥·阿莫代伊（Dario Amodei）等科技行业领袖的说辞。他指出，阿莫代伊曾在电视媒体（如 Axios）上宣扬所谓的“AI 心理恐惧症”，声称未来将有 50% 的工作岗位被人工智能取代。史蒂文·巴特利特（Steven Bartlett）插话表示，他至少尊重阿莫代伊言行上的一致性，如今阿莫代伊甚至正在遭受硅谷其他势力的口诛笔伐。但齐特隆对此嗤之以鼻，讽刺地反驳道：“仅仅四个月前，他们可没攻击他，当时所有人都把他奉为世上最聪明的神童。”
齐特隆认为，这种所谓的“恐惧”完全是虚伪的表演。如果这些科技巨头真的认为自己造出了一种能毁灭人类所有工作的怪物，他们理应每天背负着沉重如十吨的道德重担，惶惶不可终日；然而他们所做的不过是设立一个“对齐团队（Alignment Team）”来装点门面。齐特隆嘲讽 OpenAI 内部的对齐团队如同走马灯般频繁解散与换人，而阿莫代伊表现得像一位“过于自命清高、甚至在公开活动上不敢与山姆·奥特曼（Sam Altman）牵手的怪异政客长者”。在齐特隆眼中，OpenAI 和 Anthropic 骨子里是一丘之貉，甚至 Anthropic 更加具备邪教般的盲信色彩。例如 Anthropic 联合创始人杰克·克拉克（Jack Clark）曾是科技媒体 *The Register* 上最犀利的调查记者之一，如今却仿佛被洗脑一般，整日用高深莫测的宏大词汇来包装这项技术。
### [02:06:13](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7573s) 叙事之争与强制“现在时态”的设想
对话随后回溯到核心议题：到底需要发生什么，齐特隆才会在 2026 年回过头承认自己彻底看走了眼？齐特隆重申，除非 AI 系统的运行与生产成本断崖式下跌，且模型能够完成惊人数量的自主任务，达到“与魔法无异”的境界。面对史蒂文提出的“技术能力是否必须持续跃升”，齐特隆坚定地表示，门槛之所以定得如此苛刻，正是因为这些科技公司自己吹下了天大的海口。
齐特隆直言，如今围绕 AI 的所有神话与争论，本质上是一场脱离技术现实的“信息战”和“叙事对抗”。每个 AI 巨头都在试图粉饰糟糕透顶的财务账目，极力掩盖大模型能力的真实天花板。齐特隆提出一个讽刺性的监管构想：“如果由我来监管这些 AI 吹鼓手，我只颁布一条规定——他们不准再使用任何‘将来时态’说话。老兄，你只能谈论今天能做的事，最多允许你展望两周后的未来！因为如果把他们限制在当下的真实能力里，他们听起来简直就像疯人院里的疯子。”
### [02:07:12](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7632s) 代币计费的陷阱与商业神话的破灭
史蒂文提出，当年 Uber 或亚马逊在亏损扩张阶段听起来同样疯狂。齐特隆随即反驳了这一类比：Uber 当年在烧钱补贴，但其基础单体经济模型的逻辑是一致的，用户始终是以极低的价格完成从 A 点到 B 点的物理位移；Uber 绝不是每个月向用户收取 20 美元然后赠送 500 英里里程，随后在某一天突然变脸要求按英里逐一计费。
而当下的生成式 AI 正在发生这种残酷的转变。AI 公司原本推行低廉的包月订阅，但面对难以承受的算力成本，它们正全面转向基于使用量（按百万 Token 计费）的收费模式，尤其是针对 150 人以上的企业客户。这种定价策略的转变立即浇灭了客户的热情，各大企业不再惊叹于技术的“神奇”，转而惊恐于飙升的算力账单。齐特隆援引优步首席运营官安德鲁·麦克唐纳（Andrew Macdonald）的原话指出，企业如今极难为使用大模型买单，因为“根本无法把购买 Token 花掉的真金白银与实际产出的业务价值联系起来”。
```mermaid
flowchart TD
accTitle: AI商业模式转变困境
accDescr: 展示AI公司从低价无限补贴转向按Token计费后引发的客户价值脱节与预算收紧。
subgraph S1["早期补贴阶段"]
A["固定低价订阅 / 概念炒作"] --> B["企业用户低门槛尝鲜"]
end
subgraph S2["现实收费阶段"]
B --> C["高额运营成本难以为继"]
C --> D["强推按 Token / 使用量计费"]
end
subgraph S3["商业反噬阶段"]
D --> E["企业面临不可控算力账单"]
E --> F["难以证明 Token 开销与业务价值挂钩"]
F --> G["企业削减预算 / 放弃全面落地"]
end
```
### [02:08:28](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7708s) 2027 年大崩盘：OpenAI 的流动性死局
史蒂文询问，如果全社会大量经济命脉都押注在 AI 泡沫之上，泡沫破裂时究竟会发生什么。齐特隆直言不讳地指出，这不仅是 AI 泡沫，更是一场潜藏的崩盘危机，其核心引爆点将是 OpenAI 耗尽现金。OpenAI 原本计划在当年公开上市（IPO），但随后推迟至次年，其首席财务官萨拉·弗莱尔（Sarah Friar）含糊其辞地表示上市时间不会早于 2027 年。
齐特隆详细剖析了 OpenAI 极其脆弱的资本结构与现金流黑洞：
- **万亿估值幻想受挫**：在上一轮私人融资中，OpenAI 估值达到 865 亿美元；据《纽约时报》记者迈克·艾萨克（Mike Isaac）报道，OpenAI 曾试图以 1 万亿美元的荒谬估值寻求公开上市，但遭到了财务顾问的断然否决。
- **不可持续的融资胃口**：OpenAI 仅在当年就筹集了 122 亿美元，但绝大部分资金早已被基础设施和算力消耗殆尽。齐特隆计算，OpenAI 每年必须筹集至少 1000 亿美元才能维持生存。
- **估值倒挂与被动平价融资**：如果无法实现 IPO 让前期风投套现，OpenAI 将被迫继续进行私募融资，但在缺乏商业化造血能力的情况下，甚至连维持现有估值的平价融资（Flat Round）都难以推进。亚马逊此前提供的 35 亿美元注资，甚至附带了要求其尽早上市的苛刻先决条件。
- **Anthropic 的上市挤压**：齐特隆预测 Anthropic 可能会先于 OpenAI 抢跑上市。虽然 Anthropic 同样身陷亏损且难以为继，但其业务增速相对较快。一旦 Anthropic 率先试水并暴露二级市场的残酷估值折价，OpenAI 的 IPO 大门将彻底关闭。
- **连锁爆雷网络**：齐特隆预测到了 2027 年，整个 AI 产业的资金链将彻底耗尽。英伟达、黑石（Blackstone）与贝莱德（BlackRock）等私募信贷机构目前之所以不惜一切代价为数据中心托底，是因为它们深知数据中心的大部分需求全凭 OpenAI 一家支撑。更可怕的是，像软银（SoftBank）这样的跨国巨头在账面上持有约 1000 亿美元的 OpenAI 股票，其未来存续完全绑死在 OpenAI 能否成功上市套现之上。如果 OpenAI 崩塌，整个多米诺骨牌将引发巨大的连锁反应。
### [02:12:23](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7943s) 软银困境与科技巨头的“增长荒”
当被问及若初创企业无法上市会引发何种后果时，埃德·齐特龙（Ed Zitron）指出，像软银（SoftBank）这样的风投或控股机构，核心逻辑是通过持股企业上市变现，或者以股票质押借贷来获取资金流动性。如果 OpenAI 无法上市，软银就失去了退出的主要渠道。尽管软银家底雄厚不至于直接破产，但这家全球首屈一指的控股公司规模势必会大幅缩水。与此同时，亚马逊（Amazon）、谷歌（Google）和微软（Microsoft）等巨头也将不得不下调业绩指引，向市场坦承他们未来的增长速度远没有此前吹捧的那么快。
齐特龙认为，科技行业即将迎来一场深刻的“科技萧条”（Tech Depression）。在他看来，当前危机的本质在于：大型科技巨头已经彻底耗尽了真正的超高增长（Hypergrowth）点子，但资本市场对此仍抱有不切实际的幻想。巨头们之所以狂热地挥霍资金采购 AI 算力与 GPU，纯粹是为了“把罐子往前踢”（Kick the can down the road）——向市场表演“我们还在做事，我们正在押注 AI”，从而分散外界对主营业务放缓的注意力。
但这些巨头赖以生存的传统现金流业务迟早会触碰天花板：流媒体和订阅制的价格不可能无限制上涨，谷歌的广告算法微调已接近极限，亚马逊压榨第三方卖家的手段也已用尽。当这些传统引擎失去动力，而巨资押注的 AI 无法带来真金白银的收入时，科技行业将面临惨痛的估值重置。
```mermaid
flowchart TD
accTitle: AI 泡沫破裂与科技萧条传导机制
accDescr: 展现科技巨头依靠 AI 制造虚假增长，到无法兑现泡沫破裂，进而触发股市重创与经济萧条的完整传导过程。
subgraph Illusion["虚假繁荣阶段"]
A["传统业务增长见顶<br/>广告/加价/提成触达极限"] --> B["疯狂采购 AI GPU<br/>营造高增长幻觉维持股价"]
C["软银/风投循环注资<br/>依赖上市套现与纸面估值"] --> B
end
subgraph BreakingPoint["泡沫引爆（预计 2027）"]
B --> D["OpenAI/Anthropic 无法盈利且难以上市"]
D --> E["软银巨额缩水 & 巨头下调业绩指引"]
D --> F["甲骨文/英伟达等依赖方营收雪崩"]
end
subgraph Cascade["宏观外溢与萧条"]
E & F --> G["巨头股票暴跌 20%-40%<br/>纳指与标普 500 大幅下挫"]
G --> H["普通散户养老金缩水<br/>消费降级与企业裁员潮"]
G --> I["市场按传统重工业/航司标准重新估值巨头"]
end
```
---
### [02:13:40](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=8020s) 多米诺骨牌效应：从英伟达到普通人的养老金
主持人史蒂芬·巴特利特（Steven Bartlett）追问，若这场科技萧条在 2027 年爆发，是否会演变为波及全社会的宏观经济大萧条。齐特龙给出了肯定的回答，并剖析了极其危险的资金多米诺骨牌效应。
首先是普通散户的财富面临灭顶之灾。大量的普通美国民众将资金投入追踪美股“科技七巨头”（Magnificent 7）的指数基金，坚信“股价永远往上涨”。以英伟达为例，它不仅在纳斯达克占据举足轻重的地位，在标普 500 指数中的权重更达 7% 至 8%。齐特龙指出，英伟达目前正在进行极其露骨的“循环融资”：自己向 AI 初创公司注资，后者再以此质押借贷，回头大批量采购英伟达的 GPU。一旦下游的底牌被揭穿，英伟达的营收可能暴跌 50% 到 70%，迅速被打回 2022 年那个年营收仅数十亿美元的原始状态。
随之而来的直接受害者，是像“珍妮和戴夫”这样的普通工薪阶层：
- **养老金账户大幅缩水**：由于标普 500 和罗素 1000 指数重度依赖微软、亚马逊、谷歌、苹果、特斯拉、Meta 和英伟达，七巨头股价若暴跌 20% 到 40%，大众的退休养老金将遭受不可逆的沉重打击。
- **初创生态彻底归零**：去年全美超过一半的风投资金涌入了 AI 领域。齐特龙断言，绝大多数基于大语言模型（LLM）的初创企业估值最终都将归零，因为底层依赖大模型的商业模式根本无法盈利。除了埃隆·马斯克收购 Cursor 这种个例，市面上根本没有成熟买家来接盘这些动辄数十亿美元的空壳。比如编程助手 Cognition 估值高达 26 亿美元，除了去二级市场收割股民，市场上根本不可能有人接盘。
面对主持人提出的经典经济衰退连锁反应（需求下滑导致营收锐减、固定成本压垮利润、冻结招聘与裁员），齐特龙补充道，这次最大的危机在于**“资本无处安放”且无法获得救助**。OpenAI 与 Anthropic 两家公司背负着超过 1.1 万亿美元的各项基建与算力承诺；甲骨文（Oracle）正专门为 OpenAI 建设高达 7.1 吉瓦（GW）的数据中心，耗资超 4000 亿美元。扣除通胀因素，甲骨文过去 15 年的实际收入毫无增长，一旦失去了 OpenAI 这个唯一的巨额买家，甲骨文将陷入灭顶之灾。
---
### [02:17:04](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=8224s) 纸面收益的狂欢：把科技巨头“肉毒杆菌”打回原形
在探讨风险投资与科技巨头如何粉饰太平时，齐特龙揭露了惨淡的行业真相。自 2018 年以来，整个风投行业的表现堪称历史上最糟糕的周期之一，平均资金回报倍数仅在 0.8 到 1.2 之间（即投入 1 美元，只能收回 80 美分到 1.20 美元），根本无法产生实质回报。然而，整个行业却依靠互相抬高估值的“纸面收益”（Paper gains）在自欺欺人，并凭此继续向出资人募集资金。
科技巨头同样深谙此道。齐特龙提到，谷歌在上一季度的财报中，通过重新评估所持有的 SpaceX 和 Anthropic 股权价值，在账面上凭空增加了 99 亿美元的纸面净利润。这种堂而皇之的财务游戏居然没有引发监管丑闻，令齐特龙感到荒谬绝伦。他引用了一句著名的网络梗图来形容现在的科技界：“收获时狂喜‘这太棒了！’，耕种播种时哀嚎‘这太糟了！’”当下所有人都在享受估值飙升的投机狂欢，媒体在狂热撰写 Anthropic 未来值 2 万亿美元的虚假叙事。
引用《Prof G Markets》播客联合主持人 Ed Elson 的精妙比喻：
> “他们现在所有人都在打肉毒杆菌——大把大把地砸钱，只想让自己看起来还很年轻，而市场居然信了。”
齐特龙警告，一旦市场清醒过来，等待科技巨头的不仅是估值回调，而是估值逻辑的根本颠覆。华尔街将不再把它们视为享有无限溢价的成长型科技股，而是会像对待传统“航空公司”一样给它们定价：承认体量庞大，承认能靠成熟老业务赚钱，但认定它们再也拿不出新的颠覆性产品，从而给予极其低迷的传统工业倍数。
---
### [02:19:08](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=8348s) 给普通人的生存建议：远离科技赌局与话术陷阱
面对即将到来的动荡，主持人询问普通家庭（珍妮与戴夫）是否该调整财务策略、削减开支，以及是否还应继续定投标普 500 或押注 OpenAI。
齐特龙直言不讳地表达了自己的激进立场：“我没有任何资金在股市里，当前的股市就是一个被媒体大肆吹嘘的赌场。”虽然他声明自己不提供官方理财建议，但他个人目前**几乎全部持有现金**，因为他根本不信任当下的资本市场：
- **防范波动风险**：普通人应保持保守策略，如果有盈利，应及时锁定收益，而不是盲目相信只涨不跌的承诺。
- **高度警惕科技巨头的话术**：绝不要根据科技 CEO 们的承诺来进行任何投资决策。必须认清一个事实：高管们在公开场合的每一句话，目的都只是为了推高股价，而不是在陈述事实。
- **警惕误导性的财务指标**：企业会用各种模糊花哨的指标制造虚假繁荣。最典型的就是微软宣称其在 AI 领域的“年化营收运行率”（Annualized Run Rate）达到了 370 亿至 380 亿美元。普通人一听会以为微软实打实赚了近 400 亿，但实际上这种经过美化的数字完全掩盖了实际收入的虚弱与脆弱性。
### [02:20:18](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=8418s) 科技巨头的操纵与监管、媒体的失职
谈到科技巨头宣称的所谓“370 亿美元年化营收”（run rate），Ed Zitron 直言这完全是用来操控公众与投资者的数字游戏。企业甚至根本不给出清晰的定义——也许仅仅是单月营收乘以 12 的纸面推算。之所以这些公司能如此明目张胆地进行认知操控，核心原因在于美国缺乏一个能有效发挥监管功能的证券交易委员会（SEC），同时当下的媒体环境也严重失职：本应以怀疑精神为先、保护读者利益的媒体，却往往充当了科技叙事的扩音器。
面对主持人的追问——“科技高管们会辩解称，这项技术未来将极具颠覆性和价值，因此他们必须在实用价值显现之前进行先期巨额投资”，主持人试图在两派观点之间寻找平衡，并指出随着竞争白热化，并非所有玩家都能胜出，一旦某个巨头倒下，极有可能引发多米诺骨牌效应。Ed Zitron 对此表示，人们总喜欢拿当年的互联网泡沫来安慰自己，认为亚马逊和甲骨文（Oracle）在泡沫破裂后依然挺立并发展壮大。然而当下的情况截然不同：如今的企业规模要庞大得多，夸下的海口也巨大得多。Ed 甚至语出惊人地预测，连甲骨文这样的老牌巨头在这场泡沫中都可能撑不下去，并带着尖锐的黑色幽默讽刺道：“安息吧拉里（埃里森），这种下场发生在他这种刻薄的人身上再合适不过了。”
### [02:21:42](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=8502s) 问责缺失：为什么对科技 CEO 们深恶痛绝
当主持人直截了当地问 Ed Zitron 为什么如此厌恶这些科技大佬时，Ed 敞开了心扉。他坦言自己无法忍受被欺骗，而普通大众同样深恶痛绝。在现实生活中，没有任何一个普通人可以像这些科技公司一样肆无忌惮地满嘴跑火车（bullshitting），也没有哪个普通人能在撒谎和失败后获得如此巨大的纵容度。放任这些巨头大肆承诺却从不兑现、也从不被严厉问责，正在带来极其沉重的经济与社会代价。
Ed 抨击了当前科技批评界的软弱无力，尽管也有个别优秀的批评家，但看到那些超级富豪、手握巨权的行业巨擘当着公众的面撒弥天大谎或刻意歪曲事实，实在令人作呕。这也是他坚持写下长篇深度分析的动力所在——他要向公众摊开论据，展示自己为何得出这些结论。在 Ed 看来，这些巨头早已不再专注于打造好产品，他们不在乎消费者，骨子里甚至充斥着对客户的轻蔑与鄙视。
```mermaid
flowchart TD
accTitle: Tech Disillusionment Feedback Loop
accDescr: Diagram illustrating the feedback loop between tech corporate hype, media regulatory failure, and customer contempt.
subgraph Manipulation["虚假繁荣与失控监管"]
A["巨额指标操控<br/>(370亿美元虚火 Run Rate)"] --> B["监管与媒体失责<br/>(SEC缺位 / 媒体丧失怀疑精神)"]
end
subgraph Impact["现实代价与用户剥离"]
C["巨额失败与谎言被纵容<br/>(普通人无法享受的特权)"]
D["产品质量下滑<br/>(对普通用户抱持轻蔑)"]
end
B --> C
C --> D
D -.->|招致强烈批判| A
```
### [02:22:59](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=8579s) 兼听则明：在信息迷雾中保持独立思考
在访谈尾声，主持人理清了 Ed Zitron 目前的内容平台：Ed 早在 2024 年就已彻底离开 Substack，转而在基于 Ghost 搭建的独立博客（Where's Your Ed At）上发声，并主持播客节目《Better Offline》。主持人对 Ed 的工作表达了高度认可，并分享了自己对信息消费的看法：许多听众常常困惑于不同播客节目中相互冲突的观点，但现实世界的知识与讨论本就是多元分歧的。播客听众的职责并非盲从某一个人的言论，而是去跨领域收集不同的参考坐标，不管是健康管理还是前沿科技投资，都必须遵循证据、做独立调查（DYOR）。
主持人坦诚地表示，自己经常收看 Ed 的内容，正是因为 Ed 犀利的反思和对前沿 AI 实验室 Capex 资本开支的剖析，打破了许多既定认知，让人不得不停下来反思当下市场是否充斥着“虚妄与骗局”（fugazi）。回望商业历史，每个技术周期都有虚火蔓延的阶段，Ed 对 2027 至 2028 年市场可能遭遇剧烈回撤的独到预判，为公众提供了极其稀缺且宝贵的逆向视角。
- 平台更正：Ed Zitron 的通讯简报目前依托于 Ghost 平台运作，播客节目为《Better Offline》。
- 核心建议：不要对任何单一专家的观点抱持宗教般的盲信，应当持续收集客观证据，并在多重视角之间穿透表象、建立自己的认知模型。
### [02:24:58](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=8698s) 传统提问：在充满恶意的世界里重拾真实的人际联结
访谈进入了该节目的经典收尾环节：上一位嘉宾留下的匿名问题是——“鉴于高质量的人际关系对身心健康与长寿至关重要，我们该做些什么来改善人际关系与社会联结？”
Ed Zitron 给出了一个意想不到却又发人深省的答案，并将它与自己对抗 AI 泡沫的经历紧紧联系在一起。他坦言，作为一名揭露科技真相的“职业黑粉”和怀疑论者，这项工作有时极其痛苦、负面且残酷。然而，支撑他走过来的，正是从身边的社区与同盟中汲取的爱与支持。这些力量不仅来自其他勇敢发声的批评家圈子——比如 Gary Marcus、Edward Ongweso Jr.、Molly White 和 Brian Merchant——也来自生活中的普通人，比如他经常去交流吐槽的两位健身教练 Troy 和 Jake。
Ed 强调，人们太容易在工作和生活的重压中被彻底“锁死”（hard locked），忘掉了做事的初心。当面对铺天盖地的荒谬与负面现实时，最关键的是找到能与你产生共鸣、同样对这种荒诞感到愤怒的同伴，并时刻提醒自己保持人性。他在读者和听众的反馈中体会到最深切的共鸣，是自己的文字让他们感到“终于有人替我们说出了心声”。在最后，Ed 呼吁所有人：不要吝啬对身边人的赞美与爱意，无论他们是你的朋友、家人，还是你所欣赏的艺术家、作家或播客主，主动走上前去告诉他们“我真心热爱你的作品”、“你的付出太棒了”。这种真实的爱与托举，恰恰是抵抗这个虚妄时代最有力的解药。
- 关键致谢名单：Ed 在抗争中得到共鸣的同行学者与批评家包括 Gary Marcus（认知科学家与 AI 批评者）、Edward Ongweso Jr.（科技记者）、Molly White（Web3/加密货币批评学者）及 Brian Merchant（科技作家）。
- 终极洞见：抵抗时代虚妄与孤立的最好方式，是无条件地支持与提升你周围的人，在彼此身上找回真实的情感共鸣。
## Notable quotes
> [00:00](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=0s) 你们有些人可能会死，但这正是我愿意承担的风险。 — Ed Zitron
> [02:36](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=156s) 你得在这儿套一个测试框架（harness），提示词必须用这个而不能用那个；开局要用这个模型，收尾却得换另一个模型……这算哪门子‘人工智能’？它原本应当是聪明、自主、设定好就能放手不管（set and forget）的系统！ — Ed Zitron
> [04:26](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=266s) 这是一场被宣扬为史上最伟大、最颠覆的软件革命。但每当你问这些上市公司‘你们靠这到底赚了多少真金白银’时，他们却羞羞答答地声称商业机密无法公布。上市公司有好事从来不会藏着掖着；当他们遮遮掩掩时，沉默本身就已经说明了一切。 — Ed Zitron
> [30:08](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=1808s) 上帝啊，就为了这个，这家公司估值达到了 950 亿美元？！ — Ed Zitron
> [38:37](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=2317s) 如果这是按它的诚实成本来卖呢？如果大家每次操作都要自掏腰包付上 2 美元、3 美元、4 美元，而他们依然觉得物超所值、心甘情愿买单，那我可能无话可说。 — Ed Zitron
> [01:29:50](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=5390s) 它让简单的事情变简单，让困难的事情变得更难（it makes the easy things easy, the hard things harder）。 — Carl Brown
> [01:42:05](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6125s) 15 个基点，也就是 0.15%？ — Ed Zitron
> [01:55:22](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=6922s) 投资不足的风险远大于过度投资的风险（The risk of underinvesting is dramatically greater than the risk of overinvesting）。 — Sundar Pichai
> [02:03:50](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7430s) 他们试图把这一切包装成不可知、不可控的无上神力，唯有他们——唯有这两位降临人间的天使，才能制服他们亲手创造出的这头史前巨兽。 — Ed Zitron
> [02:06:40](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=7600s) 如果由我来监管这些 AI 吹鼓手，我只颁布一条规定——他们不准再使用任何‘将来时态’说话。老兄，你只能谈论今天能做的事，最多允许你展望两周后的未来！因为如果把他们限制在当下的真实能力里，他们听起来简直就像疯人院里的疯子。 — Ed Zitron
> [02:18:00](https://www.youtube.com/watch?v=Lf5oqGOCRCM&t=8280s) 他们现在所有人都在打肉毒杆菌——大把大把地砸钱，只想让自己看起来还很年轻，而市场居然信了。 — Ed Elson
## Mentioned
Ed Zitron, Steven Bartlett, Sam Altman, Dario Amodei, Mark Zuckerberg, Satya Nadella, Sundar Pichai, Prabhakar Raghavan, Ben Gomes, Shashi Thakur, Geoffrey Hinton, Elon Musk, Dara Khosrowshahi, Steve Ballmer, Jim Covello, Paul Krugman, Clifford Stoll, Matt Hughes, Daniel Kokotajlo, Leopold Aschenbrenner, Gary Marcus, Alexandr Wang, Amy Hood, Andy Jassy, Jensen Huang, Eric Schmidt, Travis Kalanick, Jack Clark, Andrew Macdonald, Sarah Friar, Mike Isaac, Larry Ellison, Ed Elson, Edward Ongweso Jr., Molly White, Brian Merchant, Carl Brown, OpenAI, Anthropic, Nvidia, Microsoft, Google, Amazon, AWS, Meta, Oracle, CoreWeave, Uber, Zoox, Waymo, Broadcom, Etched, ARM, SoftBank, Spotify, Blackstone, BlackRock, SpaceX, Cursor, Cognition, SEC, Tesla Optimus, 宇树科技, ChatGPT, Gemini, Claude, Copilot, Rufus, Sora, Stargate Abilene, Blackwell, GB200, Vera Rubin, TPU, AlphaFold, Bloomberg Terminal, AskB, GitHub, SemiAnalysis, Sightline Climate, Vectara, Oxford Economics, The Innovator's Dilemma, Theglobe.com, LinkedIn, Minecraft, Bing, Wall Street Journal, The Information, TaskRabbit, CAPTCHA, The Register, Prof G Markets, Ghost, Where's Your Ed At, Better Offline
## Video description
Tech critic Ed Zitron exposes the AI bubble, why OpenAI and Anthropic are burning billions, the fake AI boom, and why the crash could wipe out the ENTIRE economy!
Ed Zitron is a British AI critic and one of the most cited voices warning that the AI industry is one giant bubble. He hosts the 'Better Offline' podcast, reaching over a million monthly downloads, and writes the newsletter 'Where's Your Ed At'. He is the founder and CEO of the primary research firm EZPR, and is currently writing his upcoming book, 'The Hater's Guide To Silicon Valley’ (out Q2 2027).
He explains:
■ Why he believes generative AI is a “con”
■ The real reason OpenAI and Anthropic can't turn a profit
■ Why data centers could leave a $500 billion debt bomb
■ Why superintelligence is a myth sold by tech billionaires
■ Why AI won't take your job, no matter what CEOs promise
00:00:00 Intro
00:02:36 AI Is A Con
00:06:15 How Much Power Data Centres Really Need
00:08:02 Is Widespread AI Adoption Manipulation Or Do People Actually Like Using It?
00:12:00 The Actual Cost Of AI And How Tokens Actually Work
00:16:09 Is The Spending Of AI Companies Justifiable?
00:20:07 Will The Rate Of Improvement Of AI Go Up, Like Previous Innovations?
00:24:23 How Bad Are AI Mistakes?
00:26:54 Comparing Human Error To AI Hallucinations
00:31:31 If The Output Is The Same, Does It Matter If Humans Or AI Created It?
00:34:18 Can We Trust AI Like We Trust Humans?
00:38:37 Would People Use AI If They Paid The Honest Cost?
00:42:35 How Does The AI Bubble Compare To The Dot-Com Bubble?
00:47:42 Does AI Demand Match The Cost And Risk Of Data Centres?
00:52:46 Is AI Making Websites Like Google Worse?
00:58:46 Ads
01:00:51 Is AI Job Disruption A Lie?
01:10:22 Could Your Narrative Be Helping AI Companies?
01:14:22 How Dangerous Is AI Cyberhacking?
01:17:30 Is The AI Industry Creating Economic Growth?
01:19:14 How Would The US Beat China In The AI Race?
01:19:53 Is Robotics A Threat To Jobs?
01:23:23 What Do You Think About Agentic AI?
01:25:03 Is The Adoption Of AI The Same As The Rise Of The Internet?
01:28:06 The Overhype Of AI
01:30:23 What Do You Use Generative AI For?
01:33:41 Has AI Gotten More Intelligent?
01:34:29 Will AI Start To Do More Jobs As It Gets More Capable?
01:36:27 What Does The Future Look Like As AI Grows?
01:38:28 You Don't Think People's Workflows Have Been Transformed By AI?
01:40:41 Will All AI Be Powered By Data Centres?
01:43:41 Ads
01:45:12 Is Overspending On AI Due To Demand Or Something Else?
01:55:22 Tech CEOs Rebuttal
01:57:31 What Would It Take For You To Change Your Mind About AI?
02:00:50 Are AI Systems Already Blackmailing?
02:08:28 Are We In An AI Bubble And What Happens When It Pops?
02:13:26 The Tech Depression Is Coming
02:19:08 What Should The Public Do?
02:21:45 Why Do You Have A Bone To Pick With AI CEOs?
02:25:06 Last Question: What Should We Be Doing To Improve Our Relationships And Social Connection?
Follow Ed Zitron:
Linktree: https://link.thediaryofaceo.com/C6fKrVK
Better Offline: https://link.thediaryofaceo.com/A9awRDM
X: https://link.thediaryofaceo.com/GTr0z7R
Where's Your Ed At Newsletter: https://link.thediaryofaceo.com/CZ3JLap
You can get $10 off your first year of Where's Your Ed At Premium, here: https://link.thediaryofaceo.com/91LBdmi
The Diary Of A CEO:
◼ Join DOAC circle here - https://doaccircle.com/
◼ Buy The Diary Of A CEO book here - https://link.thediaryofaceo.com/BWjLTZK
◼ Shop The Diary Of A CEO collection: https://thediary.com/collections/shop
◼ Get email updates - https://link.thediaryofaceo.com/5IB1H6E
◼ Follow Steven - https://link.thediaryofaceo.com/AGU9QP4
Sponsors:
Fiverr - https://fiverr.com/diary and get 10% off your first order when you use code DIARY
Saily - Download from the app store and use code DOAC at checkout for 15% off. For more details: https://saily.com/doac ⛵
[Source](https://www.youtube.com/watch?v=Lf5oqGOCRCM)
Source quotes
[0:00] Ed Zitron: 你们有些人可能会死，但这正是我愿意承担的风险。
[2:36] Ed Zitron: 你得在这儿套一个测试框架（harness），提示词必须用这个而不能用那个；开局要用这个模型，收尾却得换另一个模型……这算哪门子‘人工智能’？它原本应当是聪明、自主、设定好就能放手不管（set and forget）的系统！
[4:26] Ed Zitron: 这是一场被宣扬为史上最伟大、最颠覆的软件革命。但每当你问这些上市公司‘你们靠这到底赚了多少真金白银’时，他们却羞羞答答地声称商业机密无法公布。上市公司有好事从来不会藏着掖着；当他们遮遮掩掩时，沉默本身就已经说明了一切。
[30:08] Ed Zitron: 上帝啊，就为了这个，这家公司估值达到了 950 亿美元？！
[38:37] Ed Zitron: 如果这是按它的诚实成本来卖呢？如果大家每次操作都要自掏腰包付上 2 美元、3 美元、4 美元，而他们依然觉得物超所值、心甘情愿买单，那我可能无话可说。
[89:50] Carl Brown: 它让简单的事情变简单，让困难的事情变得更难（it makes the easy things easy, the hard things harder）。
[102:05] Ed Zitron: 15 个基点，也就是 0.15%？
[115:22] Sundar Pichai: 投资不足的风险远大于过度投资的风险（The risk of underinvesting is dramatically greater than the risk of overinvesting）。
[123:50] Ed Zitron: 他们试图把这一切包装成不可知、不可控的无上神力，唯有他们——唯有这两位降临人间的天使，才能制服他们亲手创造出的这头史前巨兽。
[126:40] Ed Zitron: 如果由我来监管这些 AI 吹鼓手，我只颁布一条规定——他们不准再使用任何‘将来时态’说话。老兄，你只能谈论今天能做的事，最多允许你展望两周后的未来！因为如果把他们限制在当下的真实能力里，他们听起来简直就像疯人院里的疯子。
[138:00] Ed Elson: 他们现在所有人都在打肉毒杆菌——大把大把地砸钱，只想让自己看起来还很年轻，而市场居然信了。
Source quotes
[16:09] Ed Zitron: In fiscal year 2026, which just ended for Microsoft... they made total according to Bloomberg about $34.33 billion. $24.1 billion of that was from OpenAI, so that leaves them with about $10 billion in a year when they spent $115 billion on capital expenditures... and intend to spend $175 billion next year. The math does not make sense.
[20:07] Ed Zitron: Microsoft's one of the largest companies in the world, and they can barely wipe their own ass when it comes to GitHub. The quality of software is going down weirdly enough as more people use LLMs and more businesses demand... that people use these services.
[27:42] Ed Zitron: I don't pay Matt Hughes because he knows everything. I pay him because he has incredible context and a ton of knowledge and he's willing to expand it and work with me and moral support and he's a great editor... and on top of that a wonderful loving being with empathy and joy in his heart.
[35:11] Ed Zitron: When the first iPhone came out... I felt like the apes at the beginning of 2001. Visual voicemail—it was immediate... Everyone was like, 'Holy sh*t, this is incredible.' It was obvious the value.
[70:22] Ed Zitron: I think Altman and Amodei are some of the most deeply corrupt and cynical people in the world... Oh, shut up. I hear that and I feel so frustrated because I've met so many of these rich [expletive] liars.
[71:50] Ed Zitron: Every single scam and con starts with rushing you. Every single trick in history begins with saying you must do this now.
[71:50] Ed Zitron: We are still trapped in the hands of Milton Friedman, Margaret Thatcher, and Ronald Reagan... a neoliberalistic hellscape which is growth at all cost.
[100:38] Ed Zitron: Sam [Altman] the other f***ing week was saying in like 6 months it will be like a genie that you can ask wishes from... What's he talking about? Also, the genie was charming.
[100:38] Ed Zitron: Exponential improvement in software and software performance is always a result of direct hardware improvement... And where are we? Trillion plus dollars in with the future great financial crisis and the world's greatest marketing psyop.
[100:38] Ed Zitron: 10 and something billion dollars in and the best you've got is 0.15%?... If it was going well, you'd tell me how well it was going rather than doing this weird rain dance thing where you're like, 'Well, if we move all the pieces around, in 3 years theoretically this will happen.'
[105:12] Ed Zitron: Spotify didn't lose $20.9 billion in one year. They didn't need to raise $217 billion in the space of 6 months.
[105:12] Ed Zitron: They got rewarded for buying the GPUs... all the markets went rock hard overnight. They loved it.
[115:22] Sundar Pichai: The risk of underinvesting is dramatically greater than the risk of overinvesting.
[120:50] Ed Zitron: Put the story in the bag.
[141:45] Ed Zitron: Seeing these ultra rich, ultra wealthy, ultra powerful people lie through their fucking teeth... it turns my stomach.
Original transcript
[0:00] I think generative AI is at its heart
[0:02] con and seeing these ultra rich ultra
[0:05] powerful people lie through their teeth
[0:07] turns my stomach. The word con is a
[0:09] strong word.
[0:10] >> Well, what do you call something where
[0:11] from the very beginning they've sold it
[0:13] in the terms of magic but it's just a
[0:15] halfass arcery machine. They are
[0:16] misleading the entire world.
[0:18] >> You are the first person that I've
[0:19] spoken to that has that opinion.
[0:20] >> Well, the fact that this is happening is
[0:22] insane and the fact it's not a scandal
[0:24] is insane. And I've been in the tech
[0:26] industry for 16 years now and I love
[0:28] technology and I'm enthusiastic about
[0:29] it, but I don't like being misled. And
[0:32] this is the largest non-consensual push
[0:35] of technology in history.
[0:36] >> So, we're going to play a game, Ed. I
[0:38] have the things that you consider to be
[0:40] myths about the AI industry.
[0:42] >> Let's play it. The AI industry is
[0:44] creating enormous economic growth. No,
[0:46] it's not. All of these companies run at
[0:47] a horrifying loss. Open AI lost $20.9
[0:50] billion last year. None of these people
[0:52] can just say, "Yeah, we're on the path
[0:53] to making this profitable." because they
[0:55] can't.
[0:55] >> Next one.
[0:56] >> AI will replace all human jobs. That
[0:58] just isn't happening and there's no
[1:00] economic data to support it. Next, the
[1:02] United States need to spend trillions to
[1:04] beat China in the AI race. What's the
[1:06] race to do for us to constantly piss our
[1:07] pants worrying about China? But people
[1:09] keep saying, "What if these models fall
[1:11] into the wrong hands? They're already in
[1:12] the wrong hands." Mark Zuckerberg, Sam
[1:14] Olman, Dario Amade.
[1:16] >> Mark Zuckerberg says, "We'll continue to
[1:18] invest aggressively in infrastructure to
[1:20] meet the demand." God met as a
[1:21] monstrosity. Makes me think of Shrek
[1:23] with L fogquad. Some of you may die, but
[1:26] that's a risk I'm willing to accept. If
[1:27] only these people gave a [ __ ] about
[1:29] poverty or actual problems in the world
[1:31] versus are we buying enough GPUs. If
[1:34] this continues, what does the future
[1:36] look like? [ __ ]
[1:41] This is super interesting to me. My team
[1:43] given me this report to show me how many
[1:44] of you that watch this show subscribe.
[1:46] And some of you have told us according
[1:47] to this that you are unsubscribed from
[1:49] the channel randomly. So, favor to ask
[1:51] all of you. Please could you check right
[1:53] now if you've hit the subscribe button
[1:54] if you are a regular viewer of the show
[1:56] and you like what we do here. We're
[1:57] approaching quite a significant landmark
[1:59] on this show in terms of a subscriber
[2:01] number. So, if there was one simple free
[2:03] thing that you could do to help us, my
[2:05] team, everyone here to keep this show
[2:07] free, to keep it improving year over
[2:09] year and week over week, it is just to
[2:11] hit that subscribe button and to double
[2:12] check if you've hit it. Only thing I'll
[2:13] ever ask of you, do we have a deal? If
[2:16] you do it, I'll tell you what I'll do.
[2:17] I'll make sure every single week, every
[2:20] single month, we fight harder and harder
[2:21] and harder and harder to bring you the
[2:22] guests and conversations that you want
[2:23] to hear. I've stayed true to that
[2:25] promise since the very beginning of the
[2:26] Dire of Sio and I will not let you down.
[2:29] Please help us. Really appreciate it.
[2:31] Let's get on with the show.
[2:36] >> Ed Zitron,
[2:38] there are a number of things that you
[2:39] believe that a lot of other people don't
[2:41] believe, right? You have, I think, a
[2:44] couple of controversial opinions and
[2:45] opinions that are in contrast to the
[2:48] other guests that I've sat here with.
[2:50] What exactly are those opinions, Ed? I
[2:53] think generative AI is at its heart con.
[2:57] I don't think it is sold as honest
[2:59] software. I think that they overstate
[3:02] both what it can do, what it will do,
[3:04] and the underlying financials to the
[3:06] point that they are misleading the
[3:07] entire world. And they're actively
[3:09] exploiting the weaknesses in journalism,
[3:11] in our economies, and indeed within the
[3:14] responsible parties with sellside
[3:15] analysts, governments, and all over the
[3:17] shop.
[3:17] >> The word con is a strong word.
[3:19] >> Yeah. I mean, what do you call something
[3:21] where from the very beginning they've
[3:23] sold it in the terms of magic as this
[3:25] thing that will replace all jobs, that
[3:27] will cure cancer, and all of these
[3:29] things? And when you look at it, it's
[3:31] boring cloud software that's extremely
[3:33] expensive and unprofitable and also
[3:35] unreliable at its core.
[3:37] >> People will be asking where are you
[3:38] drawing from in terms of your
[3:40] references, your personal experiences?
[3:41] Where were you educate? What you study?
[3:43] What you write about? What do you do Ed?
[3:44] >> So that's the funny thing is people say
[3:46] he's not got a finance experience. He's
[3:48] not going to take. I've been in the tech
[3:49] industry 15 16 years now in PR but still
[3:52] had practical experience and I love
[3:54] technology and I'm enthusiastic about
[3:56] it. And this thing just comes along that
[3:58] everyone is telling me is the best thing
[3:59] since sliced bread. And it can't even do
[4:01] the basics. It can't even do search.
[4:03] Well, whenever you ask an AI person,
[4:05] well, what's your setup? They describe
[4:07] this PeeWee's Playhouse thing of like,
[4:09] well, you got to harness here and you
[4:10] got to use the right prompt. Well, you
[4:11] don't want to use that prompt. You want
[4:13] to use this prompt here with this model,
[4:14] but don't use this model for the
[4:16] beginning, but at the end, you're going
[4:17] to want to use this model. And this is
[4:19] meant to be artificial intelligence.
[4:22] It's meant to be smart. It's meant to be
[4:24] autonomous. It's meant to be something
[4:25] that you set and forget.
[4:26] >> We have the sort of six leading AI
[4:28] companies on the table here. Anthropic
[4:30] Amazon, Nvidia, Microsoft, OpenAI,
[4:32] Google. You're saying that their
[4:34] fundamental business model is a con.
[4:37] >> Well, their revenues are not really
[4:39] coming from AI. Up until fairly
[4:41] recently, none of their revenues were
[4:43] coming from AI. Like dribbles a bit.
[4:44] Right now, 70% of all AI revenues across
[4:48] those three companies are from OpenAI
[4:49] and Anthropic to unprofitable,
[4:51] unsustainable companies that literally
[4:53] cannot afford to exist without these
[4:55] very same companies giving them money.
[4:57] Amazon sent $50 billion to OpenAI this
[5:01] year. They sent $5 billion to Anthropic.
[5:03] Google sent $10 billion to Anthropic.
[5:06] And in the next three and a half years,
[5:08] OpenAI and Anthropic based on actual
[5:10] sellside analyst evaluations, their
[5:12] estimates that inform whether stock is
[5:14] going to go up or down after earnings,
[5:16] they are expecting 400 or more billion
[5:19] dollar of revenue, 30 or something% of
[5:22] cloud growth just from these two
[5:24] unprofitable companies that will need to
[5:26] be given the money from somewhere. And
[5:28] on top of that, these companies have
[5:30] such low respect for the average
[5:33] investor, for the analyst, for everyone
[5:35] really that they don't even disclose
[5:36] their AI revenues. The few times they
[5:38] dain us worthy, they use something
[5:40] called a run rate, an annualized run
[5:42] rate, which means well, nothing. They
[5:44] never define it. It can mean months 12.
[5:47] It can mean month 13. It can mean last 4
[5:49] weeks time 13. It's different every
[5:51] time, and they never define it. And then
[5:53] they sometimes just don't mention it.
[5:54] So, you've got this big thing that is
[5:56] meant to be the biggest, most
[5:58] influential change to software ever. And
[6:00] whenever you ask them about it, when you
[6:02] say, "What? How much you making from
[6:03] this?" They go, "Oh, I couldn't possibly
[6:04] say. I'm too shy." These are public
[6:06] companies, or at least the ones that
[6:08] aren't anthropic and open AI. When they
[6:10] have good news, they'll tell you. And
[6:11] when they don't tell you something,
[6:13] well, that actually speaks volumes.
[6:15] >> Have you you used these tools, the AI
[6:17] tools, Gemini, Anthropic, Chat, GBT,
[6:20] etc., and you found no value in them?
[6:22] There's some value, but it's not there's
[6:25] they have spent over a trillion dollars
[6:26] in capex. What
[6:27] >> does capex mean for you?
[6:28] >> Capital expenditures. So, when you are a
[6:30] business and you have operating expenses
[6:32] like electricity, for example, those
[6:34] come right off immediately. Capital
[6:36] expenditures are long-term investments
[6:37] that are theoretically one-off. So, a
[6:39] data center or indeed the GPUs you put
[6:41] inside an AI data center.
[6:43] >> Okay? So, you've got a data center
[6:45] >> and then you have these GPUs which are
[6:47] like computer chips. So AI GPUs are much
[6:51] bigger, much more power intensive. They
[6:52] take a bunch of high bandwidth memory
[6:54] and they because of how many of them you
[6:57] need. You need thousands of them, tens
[6:58] of thousands, hundreds of thousands in
[7:00] some case. You need a bunch of power. So
[7:03] an example, OpenAI and Oracle are
[7:05] building a data center in Texas in
[7:06] Abalene, Texas. 1.2 GW called Stargate
[7:10] Abene. Within that, with each one of the
[7:12] eight buildings, there'll be 50,000
[7:15] Nvidia GB200 GPUs. So, city of Bristol
[7:19] takes about 7800 megawatt of power a
[7:22] year, right? Well, Stargate Abene is
[7:26] condensing more power than that, 1.2
[7:28] gawatt into a space around 1,172
[7:31] times smaller. City of Bristol is about
[7:33] 1.2 billion square ft. Star Evelyn is
[7:36] about 998,000.
[7:38] So, you're condensing all of this power,
[7:40] all of this money, all of this labor
[7:41] into this one spot. And all of these
[7:44] data centers cost billions of dollars.
[7:46] All of these companies other than
[7:47] Microsoft are now to take out debt. And
[7:50] the thing is they've spent over a
[7:51] trillion dollars so far and they want to
[7:52] spend another trillion dollars next
[7:54] year. And for what? To make tens of
[7:56] billions of dollars, most of which comes
[7:58] from two unprofitable companies,
[8:00] Anthropic and Open AI. One of the
[8:02] rebuttals to that would be that the
[8:04] adoption, the customer adoption of
[8:07] people using Open AI and Enthropic has
[8:09] been absolutely insane. These are the
[8:11] fastest growing products in all of
[8:12] history, especially as it relates to
[8:14] sort of technology. If we just focus in
[8:16] on technology, they are, you know,
[8:18] hundreds and hundreds of millions of
[8:19] people, billions of people are using
[8:21] these tools every single day for things
[8:23] that they have subjectively decided are
[8:26] problems they need solving. So, you
[8:28] know, money is a lagging indicator of
[8:30] value. So, one would argue that they're
[8:33] just investing ahead of the monetization
[8:36] options.
[8:36] >> The first let's start with this
[8:38] adoption. Is it honest adoption when you
[8:41] are forced to use generative AI when you
[8:43] load Google? When you load Google Docs,
[8:45] Gemini screams in your ear. When you
[8:47] load Word, co-pilot's bugging you. When
[8:50] you use Amazon, whatever rofus AI is
[8:52] wants has opinions on what socks you're
[8:54] buying. This is the largest
[8:57] non-consensual push of technology in
[8:59] history. Chat GPD for example, every
[9:01] single media outlet has been screaming
[9:03] about this for 3 years. They've been
[9:05] saying, "This will take your job. You
[9:07] must use this. If you don't use this,
[9:09] you're going to be falling behind." So
[9:11] people are using it because they've been
[9:13] told to use it constantly and they're
[9:14] using it like search predominantly and
[9:16] that's partly because Google fell behind
[9:18] search and also because it's better at
[9:20] ingesting queries sometimes. Sometimes
[9:21] if you use a generative search it's like
[9:23] a trolling vessel. It's not very good at
[9:25] specifics but if you're like does this
[9:27] thing exist? Has this person ever said
[9:29] anything like this? It'll still probably
[9:30] get it wrong but it'll scour the ocean
[9:32] for you. Nevertheless, that's not worth
[9:34] a trillion dollars. None of it is. The
[9:37] amount of money being sunk into this is
[9:40] just incomparable to anything. Railways,
[9:42] it blows everything out of the water
[9:44] because there is no postbubble story
[9:48] even for this. AIG GPU is not useful for
[9:50] other things either. There's it's a
[9:53] directionless egregor of capitalism.
[9:56] this headless beast that lumbers around
[9:59] desperate to seek out growth everywhere
[10:01] in the hopes that if it harasses people
[10:03] and scares people and demonizes labor
[10:07] enough, people will be forced to use it.
[10:10] >> The the reason I I pause is because I
[10:13] just I think about my own company.
[10:14] Obviously, everybody thinks about their
[10:15] own personal situation. So, you have
[10:16] people listening now that don't use any
[10:17] AI tools. Then you'll have people that
[10:20] are using it for everything from coding
[10:22] new software tools to everything they
[10:24] write to, you know, images, whatever.
[10:26] And when you look at the the stats
[10:27] around enterprise adoption, it says 88%
[10:29] of organizations regularly use AI at
[10:31] least once for one particular business
[10:34] function. And I'd say in our company,
[10:36] 95% of people use a one of these AI
[10:40] tools like anthropical chatbt or Gemini
[10:42] every day,
[10:43] >> right? And that exists on some kind of
[10:44] spectrum of like the super users that
[10:47] are using it probably, you know, every
[10:49] hour of every day for almost everything
[10:51] to, you know, someone maybe hiring the
[10:54] executive team that's using it less
[10:55] because their job doesn't require of it
[10:57] as much,
[10:57] >> right?
[10:58] >> And when you look out into the world,
[11:00] you know, at how the world is changing
[11:03] from a content perspective, if we're
[11:04] looking at generative AI, it is obvious
[11:06] that these tools are being widely
[11:09] adopted. Part of the symptom is the AI
[11:10] slop you see all over the internet,
[11:12] >> right?
[11:13] So, I I don't know this this this idea
[11:16] that it's not being used. I struggle
[11:19] with
[11:19] >> it's being used. Here's the thing with
[11:21] the slop. Before we had AI slop, we had
[11:24] SEO slop because Google incentivized
[11:27] doing the lowest common denominator that
[11:28] would rank well in search. There's a
[11:30] whole story about how they pulled back
[11:32] spam guards thanks to Bravagar Ragavan,
[11:33] which we can get into,
[11:35] >> where they made the internet worse by
[11:37] allowing worse content to rank higher.
[11:39] It's why we have when you used to
[11:41] Google, oh, best washing machine,
[11:43] there's 11 different horrible blogs that
[11:45] read like somebody got a concussion.
[11:47] They are built to rank rather than be
[11:49] read by humans or built to be good made
[11:52] good. So AI helps weaponize that at
[11:54] scale. Yeah, you can make a bunch of
[11:56] generic slop. We've had slop for years.
[11:59] We've just found a slop machine. But
[12:01] then also there's the problem of cost.
[12:03] So when you use AI services, you burn
[12:05] tokens and it's per million tokens. So
[12:07] >> what's a token? So it's around 3/4 of a
[12:10] word. So it's characters.
[12:12] >> So the AI companies have a currency in
[12:14] which they charge you. Like a taxi in
[12:16] New York has a meter.
[12:17] >> Yeah.
[12:18] >> And they call it tokens.
[12:20] >> Yeah.
[12:20] >> And every word, let's just say for ease
[12:22] it's a word. You're paying per word.
[12:24] >> About a word. Yeah. And it's per million
[12:26] tokens. So you'll be charged per million
[12:28] input tokens. The stuff you feed into it
[12:30] like a document or a bunch a code base.
[12:32] And the output tokens are both the stuff
[12:34] it spits out at the end but also when it
[12:36] thinks. So, okay, you've asked me to
[12:38] give you the best restaurants in this
[12:40] area of New York. I should find the best
[12:41] restaurants in New York. All of that's
[12:43] output tokens as well.
[12:44] >> However, when you're paying for a
[12:46] monthly service, you don't see any of
[12:48] that. Put all that crap to the side.
[12:50] They just have rate limits. So, you can
[12:51] use them a certain amount and then when
[12:53] you run out, but they kind of offiscate
[12:54] what that was. Now, someone recently
[12:57] found, semi analysis actually found
[12:58] this, a big analyst group. They found
[13:00] that on a $200 a month chat GPD
[13:03] subscription, you can burn $14,000
[13:06] worth of tokens and on anthropics you
[13:09] can burn $8,000 for 200 bucks. That is
[13:13] how most and even on the 20 buck a month
[13:15] service you can burn $400.
[13:18] Now most people don't realize that. Most
[13:20] people have no idea what AI costs. Most
[13:22] people just think, "Oh, it's 20 bucks a
[13:24] month." No. All of these companies run
[13:26] at a horrifying loss. OpenAI lost $20.9
[13:29] billion last year because people can
[13:31] burn as many tokens as they want. And
[13:33] when they tried to move everybody on the
[13:36] enterprise side, so companies bigger
[13:37] than 150 onto actually paying the cost
[13:40] of AI in around March of 2026, to quote
[13:43] Sam Orman, they said, uh, people have a
[13:45] big problem with it. I think it's a huge
[13:47] issue, which is not really what the air
[13:50] apparent text history is meant to be
[13:52] saying, but the point is enterprises
[13:54] immediately started freaking out. Uber
[13:56] burned through their entire annual token
[13:58] budget in three months. So suddenly
[14:01] after everyone saying AI is the most
[14:02] productive thing ever. It's amazing.
[14:04] It's changing everything. The moment
[14:05] people actually had to pay for it, they
[14:06] go,
[14:08] I don't know actually. Um maybe it's
[14:11] obviously we all love it. It's all
[14:13] great, right? But it's costing too much.
[14:15] So we need to reduce the cost because
[14:17] people are just dumping stuff into it
[14:19] being like what do I do here and getting
[14:21] whatever the median is out because
[14:23] that's what these things do. they
[14:24] provide the median answer.
[14:26] >> So essentially, someone like me who's a
[14:28] power user of these tools,
[14:30] >> I could be costing Anthropic or OpenAI
[14:34] $1,000, but they're only charging me
[14:36] $100, let's say. So they are having to
[14:38] subsidize $900 of my usage because of
[14:41] the electricity costs and the costs at
[14:43] their data centers. And so your
[14:45] assertion here is that that is
[14:46] unsustainable.
[14:47] >> Yes. And just to be clear, they're
[14:49] probably not one for$1. It might be 30
[14:51] for. We don't we don't know. I think
[14:53] it's unprofitable. These companies don't
[14:55] disclose them even in their auditive
[14:56] financials. They play funny games with
[14:57] how they categorize things. But
[14:59] nevertheless, yes. And on top of that,
[15:01] the way that you stand up inference,
[15:03] which is the thing that creates the
[15:05] output within these data centers, you're
[15:08] not just saying, "Okay, turn the
[15:09] inference machine on. Let's go." You are
[15:12] standing up the GPUs necessary to take
[15:14] in the demand, and if you buy too much,
[15:17] you've wasted the money. You You have to
[15:19] pay for the hourly GPU use regardless.
[15:22] If you buy too few, your customers can't
[15:23] use it. They get pissed off at you. They
[15:25] cancel. They go with someone else. But
[15:26] nevertheless, yeah, they would get
[15:27] demand selling $20 or $40 for a dollar.
[15:31] And that's what these services do. And
[15:33] really, the simplest way to explain it
[15:34] is they were actually profitable if they
[15:36] were actually just they believed that
[15:38] these services were worthwhile and that
[15:39] they were worthy of the cost, they'd
[15:41] charge it. Regular people wouldn't be
[15:43] able to get a monthly subscription.
[15:45] They'd just be paying what it's worth,
[15:47] unless, of course, there was an economic
[15:50] problem. And it's very simple. You pay
[15:53] when you use an LLM regardless of
[15:55] whether you get what you want. When
[15:56] these things hallucinate, say you're
[15:58] doing something, you're coding something
[15:59] and they go through a code base and they
[16:02] [ __ ] up a bunch of stuff, they break a
[16:03] bunch of stuff, you're paying for that.
[16:04] You're paying for it whether it works or
[16:06] not, unless of course you're using one
[16:07] of these subscriptions. I think the the
[16:09] really interesting point is are they
[16:13] spending ahead of the value showing up
[16:16] which is I imagine what they would argue
[16:19] or are they spending all of this money
[16:22] and subsidizing all of their users in a
[16:24] way that's unsustainable and that will
[16:26] never be justified like does it you know
[16:28] because you think back through the
[16:29] history of technology you often get
[16:31] people
[16:32] losing money to grab market share
[16:34] >> right
[16:35] >> and they're also focusing on bringing
[16:37] the costs down and making it more
[16:40] profitable for them as well. But they
[16:42] can't afford to underinvest.
[16:44] >> If they were bringing the cost down,
[16:46] they would have brought the cost down,
[16:47] which they have not. It seems to be
[16:49] getting more expensive. In fact,
[16:51] everyone inference providers don't seem
[16:52] to be profitable. Even the companies
[16:54] renting out GPUs don't seem to be
[16:56] profitable. I imagine that it wasn't
[16:59] like they started out and they were
[17:00] like, "Shit, this is unprofitable at the
[17:01] beginning. We know it. Screw it. We'll
[17:03] keep doing it any screw." I don't think
[17:05] it's some big conspiracy. They probably
[17:07] thought at some point, yeah, this will
[17:09] go profitable. The chips will catch up.
[17:12] Customers will pay for the overwhelming
[17:13] value because you don't know in 2023
[17:15] where it's going to be in 2026. You
[17:16] assume it's going to go up. That's the
[17:18] nature of venture capital. They should
[17:20] have stopped in like 2024 when OpenAI
[17:22] lost over $5 billion. They should have
[17:24] been like, "Yep, this is not going to
[17:26] work." But they kept going because it
[17:28] helped number go up so much. It helped
[17:30] stock values pump. It helped everyone
[17:32] pump. It helped Nvidia pump, Microsoft,
[17:34] everyone. and not from the revenues.
[17:37] Because here's the funny thing about
[17:39] Google, Microsoft, and Amazon. People
[17:41] for years have been saying their AI bets
[17:43] have paid off. Wow, their AI bets have
[17:45] paid off. As these companies refused to
[17:47] say how much they're making from AI, but
[17:49] because their existing businesses
[17:50] continued to grow and did so, by the
[17:52] way, through price increases, changes to
[17:55] how Google and Meta uh did advertising.
[17:58] Amazon bumped up prices and changed how
[18:00] they did actually Amazon started a
[18:01] remarkable ad business during this whole
[18:03] time as well. and the selling through
[18:05] Amazon platform anyway nothing to do
[18:06] with AI but because number go up because
[18:08] revenue go up everyone went it's AI
[18:11] because these companies wouldn't spend a
[18:12] trillion dollars for for no reason right
[18:15] except in fiscal year 2026 which just
[18:18] ended for Microsoft annoying I know they
[18:21] made total according to Bloomberg about
[18:23] $34.33 billion $24.1 billion of that was
[18:27] from OpenAI so that leaves them with
[18:29] about $10 billion in a year when they
[18:31] spent 115 billion on capital
[18:33] expenditures just intend to spend 175
[18:36] billion next year. The math does not
[18:39] make sense. I imagine their plan was
[18:41] okay, this is just going to get
[18:42] exponentially more valuable and at some
[18:44] point the costs will be outpaced by the
[18:46] return. Problem is that large language
[18:48] models need a bunch of money to train
[18:50] them. They need constant data flow. They
[18:52] need customized data. It's just this big
[18:54] expensive monster. And when you try and
[18:58] talk to people about it and you try and
[19:00] say, "Hey, look, this is really bad.
[19:02] Nvidia has sold it was $215.9 billion in
[19:07] the last fiscal year worth of GPUs
[19:08] mostly. And you try and go, yeah, that's
[19:11] to support like $22 billion of revenue
[19:15] total in the entire world outside of
[19:18] these two companies that literally
[19:19] require money being fed into them
[19:21] sometimes by Nvidia to keep alive. When
[19:24] you tell people that, they go, "Well,
[19:25] companies just lose money, right?
[19:26] Companies because we have this quote
[19:29] Edson from Prophy Markets. We have this
[19:31] cult-like worship of the wealthy where
[19:33] we think that someone wouldn't spend all
[19:34] this money for no reason. Right? Because
[19:36] reconciling with that with this idea
[19:39] that the ultra wealthy, the ultra
[19:42] powerful didn't get there through big
[19:44] brains. They didn't get there through
[19:47] anything other than luck and opportunism
[19:49] and getting an MBA perhaps with the
[19:51] right people. That they just got there
[19:53] because they're regular people and they
[19:55] just happen to be in the right place at
[19:56] the right time. reconciling with that
[19:57] and realizing that the world is not
[19:59] controlled by people like a meritocracy
[20:01] is kind of grim. So it's easy to be like
[20:03] no they're not making a mistake I must
[20:05] be missing something and that's what
[20:07] they want. So you know I think back
[20:09] through the history of technological
[20:11] breakthroughs and I think about I mean
[20:13] you can look at different industries and
[20:14] one of my favorite books on this subject
[20:15] is the innovator's dilemma. not read it.
[20:18] >> And one of the things it talks about is
[20:19] how the the innovation that ends up
[20:21] taking out or transforming an industry
[20:25] often starts worse, doesn't make
[20:28] economic sense, none of your customers
[20:30] are asking for it. And this is typically
[20:32] why we end up ignoring it. So like
[20:33] you've got horse and carriages in the
[20:35] 1800s.
[20:36] >> Amazing form of transport according to
[20:37] the 1800s, you know, people of the
[20:39] 1800s. And then you have this thing
[20:40] called cars come along. Now the problem
[20:42] with cars is they broke down all the
[20:43] time. It's kind of like AI hallucinates
[20:45] now. um they were more expensive and the
[20:47] the economics of it didn't make sense.
[20:49] You might as well walk than buy a car.
[20:50] There was a law at the time that meant
[20:52] you had to walk in front of it with a
[20:53] red flag and wave and someone had you
[20:55] had to employ someone to walk in front
[20:56] of it waving a red flag. Obviously, it's
[20:58] worse. It's like a worse solution.
[21:00] However, these things that are
[21:02] disruptive innovations, they have a
[21:04] higher ceiling of growth and so they
[21:07] eventually overtake the horse. And I
[21:09] when I think about that analogy in the
[21:11] context of all of this, I go, okay, it's
[21:13] imperfect at the at the moment. the
[21:15] economic models aren't perfectly ironed
[21:18] out. They're still figuring out how to
[21:19] make it cheaper, the infrastructure,
[21:21] etc. But as if you think about the rate
[21:23] of improvement versus other you know
[21:26] let's say coding how much could I train
[21:28] a human coder to improve and to increase
[21:31] their output versus an AI agent one
[21:33] would go if you just imagine any rate of
[21:35] improvement in these AI tools at some
[21:38] point if you just imagine a 5% rate of
[21:39] improvement per month at some point it's
[21:43] you know and then you imagine a 5%
[21:45] reduction in cost which is what we did
[21:46] with the internet what we did with cars
[21:48] but Mo's law
[21:49] >> mos law is a mos law is not with GPUs.
[21:52] So let me let me actually explain. So
[21:54] Nvidia Nvidia invented I think it was in
[21:56] the 2000s they put out something called
[21:58] CUDA which is the underlying software
[22:00] library and the way to run software on
[22:03] GPUs. took them solid decade or more to
[22:06] make it something where they could do
[22:07] data analytics, one of the early things,
[22:09] mapper and such. And then when AI came
[22:11] along, they'd had lots of experience
[22:13] with it. But nevertheless, this company
[22:15] has got more money, more attention, more
[22:18] geniuses behind them, more people
[22:20] focused on making their things more
[22:22] efficient than anyone could ever ask
[22:24] for.
[22:24] >> And Nvidia, for anyone that doesn't
[22:26] know, makes the chips.
[22:27] >> They So, and that CUDA thing I
[22:29] mentioned, they were the ones with CUDA
[22:30] and CUDA allowed generative AI to grow.
[22:32] Okay, so they're chips.
[22:34] >> Chips and chips are needed. Those are
[22:35] the things that go into the data
[22:36] centers.
[22:37] >> And there specific chips are the ones
[22:38] where you can run AI software on it. So
[22:40] the training runs and also the
[22:41] inference. Now, here's the thing. The
[22:44] the car example back then you didn't
[22:47] have pretty much every mathematician and
[22:49] scientist going into the car industry.
[22:51] You didn't have the combined world's
[22:52] governments never shutting up about
[22:54] this. And by the way, giving them credit
[22:57] early since 2023, they've been saying
[23:00] this is inevitable. Even in what you
[23:01] said, 5% improvement. I don't even know
[23:03] how you'd measure that because a junior
[23:05] software engineer can still experience
[23:08] things and learn things from context,
[23:10] from how people deal with problems. And
[23:12] the way that people deal with problems
[23:13] is not as simple as looking at the code
[23:15] or reading some emails. It's context
[23:17] cues from speaking to a person. It's
[23:19] being in different environments. And
[23:21] there may there are uses for LLM's
[23:23] encoding. I don't dispute that. But even
[23:26] saying 5% uh what does that mean? Is it
[23:28] better at Rust? Is it better at C++?
[23:30] >> I'd say productivity just like yeah
[23:32] shipped. If we did it in the context of
[23:34] coding, it would be like shipped code.
[23:36] >> That's the thing that would be like he's
[23:38] the best writer in the world cuz his
[23:39] newsletter's really long. That's an
[23:41] insane way of evaluing it. With coding,
[23:43] it would be I mean it's even difficult
[23:45] to evaluate because it's is the software
[23:48] out there better is actually a great way
[23:50] of evaluating it. And I would say
[23:51] uniformly not. I would say the standard
[23:54] of software across Google, Microsoft,
[23:56] Amazon, Meta, especially God, Meta is a
[23:59] monstrosity, is worse. GitHub, GitHub,
[24:02] someone posted on Twitter earlier today,
[24:04] we should get a notification when GitHub
[24:05] is up rather than when it's down because
[24:07] that would be more reliable. Microsoft's
[24:09] one of the largest companies in the
[24:10] world, and they can barely wipe their
[24:11] own ass when it comes to GitHub. The
[24:13] quality of software is going down
[24:15] weirdly enough as more people use LLMs
[24:18] and more businesses demand and I really
[24:20] do mean demand that people use these
[24:22] services. So on this point of if we go
[24:24] back to this horse and carriage and car
[24:26] analogy say that we're at whatever point
[24:29] today if you imagine any rate of
[24:31] improvement in the technology which we
[24:32] have seen since tragedy came out
[24:34] >> I remember when tragy came out and I was
[24:36] in Asia and I was there showing it to my
[24:37] fiance I was like look it can do this
[24:39] and it was hallucinating once in a while
[24:40] and getting things wrong. I actually
[24:42] don't have that experience anymore. I
[24:45] have moments where I believe it's
[24:47] reasoning is weak, but I don't have
[24:49] outright hallucinations anymore. See
[24:52] that? I I disagree. So,
[24:54] >> give me an example of what you define as
[24:55] a hallucination.
[24:56] >> Okay, great one. So, I have a Bloomberg
[24:57] terminal. Yeah. The very useful thing
[24:59] they have on there is ask B. So, when
[25:01] you do a Bloomberg inquiry to like look
[25:03] up what we think Nvidia's revenue is
[25:05] going to be next quarter, it runs
[25:07] something called BQL, which is its own
[25:08] programming language. Now, instead of
[25:10] having to learn that, you can just type
[25:13] into RSB and it will generate it and run
[25:14] it for you. And so, you get it pulled up
[25:16] and you know where the data is coming
[25:17] from. It deals with hallucinations real
[25:19] well. The other day, I was like, you
[25:20] know what, get a little spicy. I'm going
[25:22] to look up the growth rate of stocks of
[25:25] Microsoft, Google, Meta, and Amazon over
[25:28] the course of 5 years, I think it was.
[25:30] >> And I was about to I was copy pasted it
[25:32] over to something looked at in Excel. I
[25:34] was about to was writing the newsletter.
[25:35] I went, Microsoft stocks never been $575
[25:39] a stock.
[25:41] You know what? When it's a cute little
[25:42] thing like, oh, it's a stock price and I
[25:44] kind of call it was no harm, no foul.
[25:46] That's fine. But when you're talking
[25:48] about, I don't know, like a transcribing
[25:50] tool for a doctor or a financial model
[25:53] that a hedge fund is dependent on, at
[25:56] that point it becomes a little more
[25:57] dangerous. And the thing is a
[26:00] hallucination with a software package.
[26:03] For example, you're refactoring a code
[26:04] base and it leaves a door open
[26:06] security-wise or it just breaks
[26:08] something and you I don't know maybe
[26:10] you've been vibe coding for 6 months.
[26:12] You haven't really been coding with your
[26:13] own hands for a while. Maybe you've
[26:14] forgotten a few things. You had this
[26:16] slop to look for. [ __ ] I'm not doing
[26:18] it. And so the problems become
[26:21] multiplicative. And I don't really know
[26:23] how you train them out of that. And
[26:25] they've certainly not succeeded. So on
[26:28] one hand they have got better but one of
[26:31] the main ways they evaluate them getting
[26:32] better are benchmarks that are adjusted
[26:35] specifically for large language models
[26:37] because you can't just have them do
[26:39] tasks. They've got better at that. They
[26:41] found some tasks they can have them do
[26:42] on them like meter me they have this
[26:45] thing where it's like check out this
[26:46] chart look how much better it's getting
[26:48] at running tasks. Wow it can go for an
[26:50] hour and then you look it's like yeah
[26:51] and successfully completing them 50% of
[26:53] the time. They they have a hallucination
[26:55] leaderboard and it really focuses on
[26:57] basic tasks and it shows that the
[26:59] four-year trend according to historical
[27:00] data from the Victaria hallucination
[27:03] leaderboard shows that hallucination
[27:05] rates on simple summarization tasks have
[27:07] plummeted from around 21% 21.8% 4 years
[27:11] ago down to 0.7%
[27:14] roughly on today's top frontier models
[27:16] like Gemini and Chat GPT. Again the
[27:19] point of nuance here is that these are
[27:21] on simple tasks which is kind of what
[27:23] I've experienced. I've experienced that
[27:24] on day-to-day things that hallucinates
[27:26] less again rate of improvement thinking.
[27:28] So if I just imagine the trajectory to
[27:30] continue there is going to become a time
[27:33] where hallucinations become rarer than
[27:35] they are today increasingly and also
[27:38] what I would say is when I think about
[27:39] other technologies there's two more
[27:40] points other technologies at their
[27:42] inception when they first came to the
[27:43] world like the internet also had
[27:45] technical difficulties. I remember
[27:47] growing up with dialup modems and I
[27:49] couldn't go on the phone at the same
[27:51] time as going on the internet. I'd have
[27:52] to stop Runescape upstairs to go on the
[27:54] phone. And you thought this is crap.
[27:56] This is technology crap. All the
[27:57] >> I I don't know, mate. I loved it.
[27:59] >> Yeah, I know. You It felt like magic.
[28:01] And then in hindsight, you go, "Wow, I
[28:03] now have Starink and 5G internet from my
[28:05] phone. It's unbelievable." You couldn't
[28:07] leave the house with internet before.
[28:09] And that's what I mean by the rate of
[28:10] improvement thinking. I'd say the last
[28:12] point is we often compare AI to
[28:16] perfection,
[28:17] >> right?
[28:18] >> Whereas that's not actually the
[28:19] alternative in the working world. Like
[28:22] if I wanted to do let's say a simple
[28:24] writing task, I should compare AI to my
[28:27] alternative alternative way of doing
[28:29] that simple writing task which is both
[28:31] measured in my time right and my ability
[28:34] to hallucinate as a person who doesn't
[28:35] know everything
[28:37] or if I'm hiring someone an intern who
[28:40] might also be prone to hallucination or
[28:42] have gaps in their knowledge.
[28:44] >> So it's not actually like we're
[28:45] comparing we should compare AI to
[28:46] perfection. It's AI to the other
[28:48] alternatives. And if someone
[28:49] hallucinates 0.7% of the time, but knows
[28:52] way more and is faster, maybe on a net
[28:56] basis, that's a good trade. Maybe I
[28:58] should use AI. So, let's start with an
[29:00] example. Someone I love dearly, Matt
[29:02] Hughes, my editor, lives out of
[29:04] Liverpool. Wonderful guy. I don't pay
[29:06] Matt Hughes because he knows everything.
[29:09] I pay him because he has incredible
[29:11] context and a ton of knowledge and he's
[29:13] willing to expand it and work with me
[29:15] and moral sport and he's a great editor,
[29:18] but he's also someone who gets into the
[29:20] guts of it and has the experiences of
[29:21] it. He's a decorated tech journalist and
[29:24] on top of that a wonderful loving being
[29:26] with empathy and joy in his heart for
[29:29] the stuff he loves and absolute [ __ ]
[29:30] venom for the people he hates. That's I
[29:33] can't get that from a large language
[29:34] model. But on top of that, I don't I
[29:36] push back on just the assumption there.
[29:38] >> When you say knows everything, what good
[29:40] is something that knows everything when
[29:42] it sometimes doesn't know anything when
[29:43] it's sometimes? And on the thing is, are
[29:45] you really paying an intern for
[29:47] something basic? Are you really going to
[29:49] them and saying, "Yeah, can you look up
[29:51] what the date is?" No, you're doing that
[29:52] on Google. Whatever the task is, you are
[29:55] trying to also train an intern. The
[29:57] point of an intern is to train them and
[29:59] turn them in, take them out of Pinocchio
[30:01] status,
[30:02] >> but it's also an intern learns. And in
[30:04] turn gets context and in turn learns
[30:05] your habits. Learns
[30:06] >> AI gets context and learns.
[30:08] >> No, it doesn't. It
[30:09] >> doesn't learn.
[30:09] >> I mean, it doesn't. The way it learns is
[30:12] you create a giant claw. MD file that it
[30:14] sometimes doesn't read, sometimes does
[30:16] read. You create a harness. You put it's
[30:18] like it's Pee-Wee's breakfast machine
[30:20] from PeeWee's Playhouse. You have to do
[30:22] all these controversies to mitigate the
[30:24] hallucinations. And even then at the
[30:26] end, how much effort have you put in?
[30:28] >> But so, okay, this is an extreme
[30:29] simplified example. If I went on my
[30:31] Claude now and said, "What's my dog? my
[30:32] dog's name.
[30:33] >> Uhhuh.
[30:33] >> It would know my dog's name.
[30:35] >> Jesus Christ. This this company raised
[30:37] 95 billion.
[30:38] >> I'm saying I'm I'm using an extreme
[30:40] simplified example to show that it can
[30:41] remember things from the past.
[30:43] Obviously, it knows much more complex
[30:44] things as well, but I just use that as
[30:46] an example. So, we we we accept the fact
[30:48] that it can it does have memory of the
[30:50] past.
[30:51] >> It has files it can access that have
[30:52] stuff on it, but that's not the same as
[30:55] memory. And it's also just okay. So, it
[30:57] remembers your dog's name. It might
[30:59] remember your habits. It might be able
[31:01] to read things you've said before.
[31:03] >> Does it know your moods? Does it know
[31:05] what's going on in the world around it?
[31:06] Does it have good days and bad days? Is
[31:08] it there for you? Because it's just a
[31:10] [ __ ] text machine. And the thing is
[31:12] the intern example. An intern is
[31:15] something that can grow. It's something
[31:16] that you invest in. That's not something
[31:18] you do through feeding files and text to
[31:20] it. The way that we store memories
[31:22] ourselves, the way in which we acrue
[31:24] experiences is a a milerum of emotion
[31:29] and feelings and facts
[31:30] >> completely different. So I think there's
[31:32] two things here. There's the process in
[31:34] which something happens and then there's
[31:35] the output.
[31:37] >> So the process you're describing the
[31:38] process of how a human does memory,
[31:40] >> right?
[31:41] >> The way that an AI does memory is
[31:43] different. But the thing that people
[31:45] care about is there value in the output.
[31:47] I.e. You know, if I dump all of my files
[31:50] into Claude, I don't really care how it
[31:52] processes it as long as when I ask it,
[31:54] what's my revenue? It has the number.
[31:56] And one could say the same thing about
[31:58] training someone. You could say, you
[31:59] teach them, you put lots of effort into
[32:00] them. You give them lots of context. You
[32:03] you educate them and give them
[32:04] experiences. And then you might come and
[32:06] say to them, by the way, what's my
[32:07] revenue? Now, the processes are entirely
[32:09] different, but the outcome is what I
[32:10] care about. Do they know the revenue
[32:12] number when I ask them? And so, I think
[32:14] that's the part that we sometimes get
[32:15] lost. we get, you know, cuz I have I've
[32:16] heard this debate about like can AI be
[32:18] creative,
[32:19] >> right?
[32:19] >> I think like the way to answer that
[32:21] question is like it's about the output
[32:23] when I ask it to do a creative thing
[32:25] does it give me the answer not is the
[32:27] process the same as a human process cuz
[32:30] actually no who cares what the people
[32:32] care about they pay for the outcome the
[32:34] product.
[32:34] >> I actually disagree about the process
[32:37] because Matt Hughes for example
[32:39] >> your editor
[32:40] >> Yeah.
[32:40] >> Yeah. watching him go down a rabbit hole
[32:43] and being there with him and actually
[32:44] vice versa him doing the same thing. We
[32:46] wrote these well I mean we were working
[32:48] on the research I ended up sitting there
[32:50] for like the dayong session of writing
[32:53] 11,000 words and he he had given me a
[32:55] bunch of notes. It was actually just
[32:56] even describing that process, I feel so
[32:59] happy cuz it was like us being like I
[33:00] can't believe how [ __ ] these Jesus
[33:02] Christ they can't do like just like the
[33:04] misanthropy of just the horrible cynical
[33:07] people of asset managers like Blackstone
[33:09] just learning about them and being like
[33:10] it can't be this and having a back and
[33:12] forth with him that is fundamentally
[33:14] different because we were both learning
[33:16] together and the learning process was as
[33:18] much about creating the output as the
[33:20] output itself. When you learn something,
[33:22] you're not creating the average, which
[33:23] really is what these things do, of the
[33:26] documents it could find. You're not
[33:28] getting particularly novel outputs. If I
[33:31] needed a generic slop output, sure, but
[33:34] I've I've used some of the higherend LLM
[33:37] harness machines that the hedge funds
[33:39] use, and they all give the same shite.
[33:41] It's all the same the same generic
[33:43] reports, the same, oh, we noticed this
[33:45] analysis, things that you can find on
[33:46] any kind of AI slop out there. what you
[33:48] described to me there, what I heard
[33:50] anyway is there's two points of value
[33:51] you're getting from your time with that.
[33:53] I mean, I mean, there's many more, but
[33:54] you said you're you're learning and then
[33:57] you're getting this book edited blog
[33:59] blog. You're getting a blog edited,
[34:00] which is the output, and you're getting
[34:02] learning and you're also really getting
[34:03] connection and all these other things.
[34:05] But when I come to when people sort of
[34:06] think about the value of AI, of course,
[34:08] they could use it to learn. But in the
[34:10] example I gave of like repeat my revenue
[34:11] number back to me or do this number, I I
[34:13] just care about the output. I could use
[34:15] it to learn. I could say what if the
[34:16] revenue number was wrong once you should
[34:18] have defined deterministic ways of
[34:21] knowing those numbers you should not
[34:23] rely on them even with the terminal
[34:25] running BQL which I trust I will double
[34:27] triple treble check everything just to
[34:30] be sure partly because also the process
[34:32] of learning for me I don't want just a
[34:34] report I go like that I want something
[34:36] that I fully understand and also
[34:38] understand the context around it I don't
[34:41] think that LLM do that and I just don't
[34:43] see them getting
[34:45] in a way that does that because it's
[34:48] it's just not what they do. And also
[34:50] there's the other problem of the more
[34:51] detailed the report, the more likely
[34:53] there are things to be wrong with it. If
[34:54] you are with Matt Hughes, for example, I
[34:57] can trust he's got it right. I can trust
[34:59] he understood and I can trust that I can
[35:01] have a back and forth with him that will
[35:02] inform me if I've missed something. I
[35:04] can read the stuff that he's read and
[35:07] actually trust him because there's a big
[35:09] trust part as well. What is the basis of
[35:11] your trust in Matt? Could it be his
[35:14] historical performance?
[35:16] >> I mean, yes.
[35:17] >> Okay.
[35:17] >> And also the fact we've learned half of
[35:19] this stuff together,
[35:20] >> but but tenure tenure doesn't
[35:22] necessarily There's probably people, you
[35:23] know, for 15 years who you also don't
[35:25] trust. Yes.
[35:25] >> So, I think I was trying to figure out
[35:26] like what is the what is the thing
[35:28] that's causing humans to trust another
[35:29] thing. And I guess it would be continual
[35:31] delivery of a commitment made of sorts.
[35:34] And so with Claude for example on simple
[35:37] tasks as we've seen from this
[35:38] hallucination leaderboard it continually
[35:41] delivers for people and that's why we've
[35:43] seen the fast
[35:44] >> I mean is that what that board says
[35:45] >> well it's it's saying like is it getting
[35:47] it wrong is it hallucinating
[35:49] >> simple task how are those defined
[35:51] >> I I don't know
[35:52] >> that's the thing though because this is
[35:53] actually a very very illustrative thing
[35:55] of the AI industry they are the what
[35:58] aboutist masters they have like well
[36:00] look we got this we got this benchmark
[36:02] that says we're good at this and look
[36:03] the numbers higher What's the number
[36:05] mean? No. What does that mean? And I'm
[36:08] not using this as a critic against you.
[36:09] It's
[36:10] >> when you can't give a direct answer, you
[36:12] give a side answer. When you as the LLM
[36:14] industry want to prove your worth, you
[36:17] can't just be like just use the product.
[36:18] When the first iPhone came out, go was
[36:20] Penn State at the time. Oh, I felt like
[36:23] the uh apes at the beginning of 2001.
[36:25] [ __ ] official voicemail. It was
[36:27] immediate. And I showed it to tech
[36:29] friends. I showed it to the most normal
[36:31] people in the world. And everyone was
[36:32] like, "Holy [ __ ] this is They were on
[36:34] razors. They were on Nokia 3210s. It was
[36:37] obvious the value." Amazon Web Services,
[36:38] same deal.
[36:39] >> It wasn't obvious though.
[36:40] >> Yes, it was. I mean, I bought it
[36:42] >> to you. To you, it was.
[36:43] >> It was. And I also showed it to a bunch
[36:45] of people because I'm aware that I had
[36:46] bias when I just love gadgets.
[36:48] >> But but I remember the famous Steve
[36:50] Balmer who was the CEO of Microsoft
[36:52] interview where he was told about the
[36:54] iPhone and he bursts out laughing.
[37:00] $500 fully subsidized with a plan. I
[37:03] said that is the most expensive phone in
[37:06] the world and it doesn't appeal to
[37:07] business customers because it doesn't
[37:09] have a keyboard which makes it not a
[37:11] very good email machine. You can get a
[37:14] Motorola Q phone now for $99. It's a
[37:17] very capable machine. It'll do music.
[37:20] It'll do internet. It'll do email. It'll
[37:22] do instant messaging. So, I I kind of
[37:25] look at that and I say, "Well, I like
[37:27] our strategy. I like it a lot.
[37:30] >> He burst out laughing, mocking it
[37:32] because it was so disruptive. It was way
[37:34] more expensive
[37:35] >> and it was way different. No keyboard.
[37:37] >> Well, phones used to be insanely
[37:39] expensive and the carriers would cover
[37:40] them, but you had to sign a long
[37:41] contract. You were still spending 500
[37:43] bucks. But the thing I'm getting at is
[37:44] you didn't have to explain to someone
[37:46] why perhaps you'd have to get past the
[37:48] cost part, but you could just be like,
[37:49] "Look how good this is." And then once
[37:51] the app was the iPhone 3G with the App
[37:52] Store, people were like, "Oh [ __ ] this
[37:54] could actually change things." mobile
[37:56] web. Even though it was a monstrosity,
[37:58] it was so bad at first. Even then, you
[38:00] could get your emails and you could just
[38:01] look at them. Point is, Blackberries
[38:02] were also expensive and were still
[38:04] actually kind of cool, but the way they
[38:06] worked was not like consumer software.
[38:07] They didn't have the classic GUI.
[38:09] iPhones felt like that. It felt like an
[38:11] a cell phone designed even like a
[38:14] computer. It was obvious. It was obvious
[38:16] from the beginning. Everyone I was I was
[38:18] dating a girl in the center of
[38:19] Pennsylvania at the time and everyone I
[38:20] showed it to was like, "Wow, this is
[38:21] incredible." That to me is the obvious
[38:24] thing with AI to this day when you're
[38:27] like, "Okay, why is it so amazing?"
[38:28] People still dither. People are still
[38:30] like, "Yeah, you can't run a business
[38:32] fully with it without this weird system
[38:35] of pulleys and levers and such."
[38:37] >> But how come then when you look at the
[38:39] stats around ChachiBT's growth,
[38:42] >> 100 million active users in just the
[38:45] first 60 days after launching? For
[38:47] comparison, Tik Tok took 9 months.
[38:48] Instagram took 2.5 years. And the
[38:50] internet itself for the worldwide web
[38:51] took roughly 7 years to reach that
[38:53] scale. Over 60% of the US adults are
[38:56] integrated into AI tools in their daily
[38:58] and regular routines within 3 years of
[39:00] the launch, reaching a 40% of the
[39:02] population. And that same milestone took
[39:05] the internet 5 years and personal
[39:06] computers nearly 12.
[39:08] >> Okay. So like this is the I think this
[39:10] is the part that's giving me dissonance
[39:11] is like when I showed my fiance chachi
[39:14] okay it was didn't really work
[39:15] >> but as a sole entrepreneur who English
[39:18] isn't her first language
[39:20] >> who has to write lots of text lots of
[39:22] copy and generate lots of images and was
[39:23] paying a graphic designer to help her
[39:24] make um certain images that she you know
[39:27] couldn't make herself because she
[39:28] doesn't have the skills.
[39:30] >> She would describe it as being
[39:32] transformative for her business. What
[39:35] I'm hearing from you is that it's not
[39:37] transformative and there's no value in
[39:38] it for people. But she if she was sat
[39:40] here transformative,
[39:42] would she pay the per million token
[39:44] rate? Would she pay the actual rate? Cuz
[39:46] that's the thing. If this was sold at
[39:48] its honest cost. Yeah.
[39:49] >> I would actually if and people were
[39:50] reacting like that and they were paying
[39:52] 23 $4 every time they did something and
[39:54] they were genuinely happy. That might be
[39:55] an argument.
[39:56] >> What is the what would be the honest
[39:57] cost if they weren't sub
[39:58] >> the actual per million token cost? The
[40:00] actual API cost they should char.
[40:02] >> Do you know how much that is relative to
[40:04] God? Depends on it depends on the model.
[40:06] But there's actually kind of a point I
[40:09] want to make about the thing you said
[40:10] with the internet earlier. So when I
[40:12] first got on the internet 33.4 kilobits
[40:14] a second modem even back then I was like
[40:17] [ __ ] if this was faster and that was
[40:20] like immediate just like if this was
[40:21] faster cuz it was slow. You go on like
[40:23] happy puppy or something download take
[40:25] all bloody day waiting for share word to
[40:27] download immediately like if I could do
[40:29] this faster it would be better. And even
[40:30] back then I'm like, man, you could
[40:32] probably do video camera stuff with this
[40:34] stuff that eventually happened. And
[40:35] actually, there's this guy called Jim
[40:36] Cavell from Goldman Sachs in a report he
[40:39] did in 2024 that was geni too much spend
[40:41] for not enough return. Paraphrasing
[40:43] there. And he made the point that in the
[40:44] run-up to the iPhone, there was
[40:47] thousands of presentations that when GSM
[40:49] radios get smaller, when Bluetooth
[40:50] radios get smaller, when Wi-Fi radios
[40:52] get smaller, it is inevitable that we
[40:55] will get something like this. And then
[40:57] he said that there is no such path for
[40:59] AI. There was no road map to AI becoming
[41:03] this thing that they promised. And I
[41:04] must be clear, if these companies had
[41:06] gone out there and are like, "Yeah, this
[41:08] is interesting cloud software. It's
[41:10] generative. It's really expensive. We're
[41:12] not sure if we can fully not trust it.
[41:15] Not in the I'm scared way. I mean, just
[41:16] like we're not sure that this is going
[41:18] to be a disruptive world changing thing.
[41:21] It has potential, but we're going to go
[41:23] slow. It's really expensive. This is an
[41:25] R&D effort. We're not going to expose
[41:26] consumers to it." and actually being
[41:28] like called them like I don't know
[41:30] language models and no no generative AI
[41:32] stuff just being not even call it
[41:34] because it isn't AI it's not autonomous
[41:36] it's not smart I actually might respect
[41:38] it but this is not they've gone out
[41:40] there since 2023 and said it was 2022
[41:43] this is the best thing since sliced
[41:44] bread this is changing everything this
[41:46] is going to do all your work this is
[41:48] going to take your job you're going to
[41:50] talk to Bing and it's going to tell you
[41:51] to leave your wife all of these crazy
[41:52] things and what's funny is when the
[41:55] writer uh Kevin Roose I think it was
[41:58] He was speaking to Kevin Scott, the CTO
[41:59] of Microsoft, about it. And Kevin Scott
[42:01] goes, you know, I'm just glad we're
[42:02] having this conversation. Instead of
[42:04] being like, "Settle down, Beas. It's a
[42:06] website. The website told you something.
[42:08] It's just LLM." They talked it up. And
[42:10] that's because everyone is talking about
[42:12] what they wish this was. Rather than
[42:14] talking about what it can actually do.
[42:16] This makes it scary to people
[42:18] deliberately. So, it makes it
[42:20] environmentally destructive. Look at the
[42:21] gas turbines poisoning black
[42:22] neighborhoods. I think it's in
[42:24] Louisiana. It's one of Musk's data
[42:25] centers. Look at the incredible energy
[42:28] draws. It is raising power bills and
[42:30] also it is creating inflation across all
[42:33] consumer electronics because of the
[42:35] massive RAM.
[42:36] >> You know what's interesting? I almost
[42:37] feel like so much of what you're saying
[42:40] is true and also it can be true that
[42:45] this technology is going to profoundly
[42:47] change the world. And I think like you
[42:49] know I think back to the early days of
[42:51] the internet is maybe the closest
[42:52] analogy we have of you know in the com
[42:55] bubble. you know, you wrote this great
[42:56] essay.
[42:57] >> Yes. Yes.
[42:57] >> Which I found really funny um especially
[43:00] the name the rot economy and you talked
[43:02] about the rotcom bubble.
[43:04] >> Yes.
[43:05] >> Talking about how AI is of less value
[43:07] than people think.
[43:09] >> And in that in the sort of com bubble,
[43:11] what you saw is huge hype, people
[43:13] overselling the capabilities of their
[43:15] websites and what they were building.
[43:17] But in the wake of the dotcom bubble,
[43:21] yes, 90% of stuff went to zero,
[43:23] >> but you had generational companies born
[43:26] that changed the world,
[43:27] >> right?
[43:28] >> And so I I do I kind of and that's what
[43:30] bubbles do, right? Huge hype,
[43:32] overinvestment, investors get crazy,
[43:34] delusional. They think it's everything's
[43:36] going to change. At the same time, you
[43:39] do have skeptics
[43:40] >> in these moments. The the dot bubble had
[43:42] I mean the internet itself had the
[43:44] biggest skeptics in 1998. Nobel Prize
[43:46] winning economist Paul Krugman said by
[43:50] 2005 or so it will become clear that the
[43:52] internet's impact on the economy has
[43:54] been no greater than the fax machine. In
[43:56] 1995 astrophysicist Clifford stool
[43:59] famously I wrote about this in my book
[44:01] wrote famously in Newsweek. Do our
[44:04] computer pundits lack all common sense?
[44:06] The truth is no online database will
[44:08] replace your daily newspaper. No CDROM
[44:11] can take the place of a competent
[44:12] teacher. Commerce and businesses will
[44:14] shift from offices and malls to networks
[44:16] and modems. Bologoney. So, how come my
[44:19] local mall does a roaring business and
[44:22] the cyber mall gets zero business? And
[44:24] then I'll give you one more from
[44:26] Krueger, who was the award-winning
[44:28] economist. He said, "The growth of the
[44:30] internet will slow drastically as it
[44:32] becomes apparent most people have
[44:33] nothing to say to each other."
[44:36] That's that that that may actually be
[44:38] the worst one of those predict like hang
[44:41] around any bar in middle America.
[44:43] Honestly, the best conversation,
[44:44] >> but it's just all the same thing.
[44:45] >> I actually So, Clifford Stall actually
[44:47] his piece was interesting cuz that there
[44:49] were some boner points in it, but he
[44:50] made points about how like an
[44:52] overwhelming amount of bad information
[44:53] out there is bad for society. He's
[44:54] completely right saying how online
[44:56] education would not be a great
[44:58] replacement for regular education. I
[45:00] think we've seen that. But there is an
[45:02] economic difference that's vastly it's
[45:05] just completely different. So.com bubble
[45:07] was actually two bubbles. There was the
[45:08] website bubble which was just trash on
[45:10] trash on trash. It was just like I think
[45:13] what was it? Excite at home bought a
[45:15] eury incard company for like a billion
[45:17] dollars. It was insane crap happening
[45:19] that was so small. The big thing that
[45:22] people are thinking about is the dark
[45:24] fiber.
[45:24] >> Dark fiber. dark fiber was all of the
[45:27] wires that put in the ground thinking
[45:28] we're going to have all this demand for
[45:30] internet and it turned out that demand
[45:32] for internet I think the analyst
[45:35] estimate was it was doubling every 90
[45:37] days when it was doing that every 6 to
[45:38] 12 months maybe maybe longer and just
[45:41] thus there was a massive overbuild of
[45:43] fiber optic cable and indeed the
[45:46] transmission stations and such just
[45:48] simplifying to bring that to people's
[45:49] houses and there was the assumption that
[45:52] well that would all get lit up and
[45:53] people would want it immediately didn't
[45:54] really
[45:55] Now the post.com bubble thing people say
[45:57] is well but after that there was demand
[45:59] from the internet. That's the thing
[46:01] though that's very different to demand
[46:03] for generative AI. Right now the demand
[46:06] we have for generative AI is
[46:07] predominantly subsidized. Just let's
[46:09] start there.
[46:10] >> Yeah
[46:10] >> predominantly subsidized and most people
[46:12] experience it are not paying the real
[46:14] cost.
[46:14] >> I agree.
[46:15] >> On top of that we already have all of
[46:18] the possible marketing in the world. We
[46:20] have the largest, most disingenuous
[46:22] marketing campaign in the history of
[46:24] man, pushing this up the hill. We have
[46:27] the apex predator of cloud software,
[46:30] Microsoft. They can only get singledigit
[46:32] billions from selling AI software. And
[46:35] Christ almighty, outside of OpenAI and
[46:37] Anthropic, we barely get $22 billion.
[46:40] And the thing is, $22 billion is a large
[46:42] amount to you and me. It's not a large
[46:44] amount of money when you spent a
[46:45] trillion plus dollars. When you have
[46:47] anthropic and open AI with $1.1 trillion
[46:50] worth of cloud commitments and on top of
[46:52] that, how does this turn into a post.com
[46:54] bubble thing? A data center built today
[46:57] is going to be as expensive to run in
[46:59] 2050 as it is today unless there's some
[47:01] breakthrough in electricity. But again,
[47:04] that's not happening with AI. AI is not
[47:06] doing that unless there's some
[47:07] breakthrough in GPU technology. But we
[47:09] already have Broadcom, Nvidia, etched.
[47:12] We have every major chip company ARM
[47:15] trying to do something about this. And
[47:17] no one seems to magically be able to
[47:18] make this profitable or indeed even less
[47:21] costly. Even Nvidia with Vera Rubin,
[47:24] their more expensive new GPU system.
[47:26] Even then, they're like, "Yeah, 10x more
[47:28] efficient. It's uh more dollars per
[47:30] megawatt." They're all koi about it.
[47:32] They don't just say, "Yeah, we worked
[47:33] with OpenAI and Anthropic and we found
[47:35] it reduced our cost by 50%." Easiest
[47:37] thing in the world if it was true. And
[47:38] that's because it's not happening. And
[47:41] this isn't a case where
[47:42] >> So are you saying there's not going to
[47:43] be the demand for let's say let's you
[47:46] know there's different types of AI
[47:48] generative AI we
[47:49] >> Yeah. And actually that's a good point
[47:50] to make. The reason they use the term
[47:52] artificial intelligence is so everyone
[47:54] would lump everything into it.
[47:56] >> They would lump uh protein folding
[47:58] nothing to do with LLMs. Robotics not
[48:00] LLM.
[48:01] >> Autonomous weapons even horrible as they
[48:02] are not LLMs because you couldn't trust
[48:04] them. But they've mushed everything into
[48:06] AI so that when you say, "Well, AI
[48:09] can't," they'll go, "Um, um, sir, you
[48:12] forgot to give us homework and also AI
[48:14] it's working on curing cancer." When
[48:15] it's just like, "No, that's not LLM.
[48:17] Stop giving them credit."
[48:18] >> The similarity though is they all need
[48:20] GPUs, all these.
[48:21] >> And that's the funny thing. All those
[48:23] data centers that we're building, all of
[48:25] them are for just generative AI. They're
[48:28] not for all of the other stuff. They're
[48:30] not for the cool [ __ ] AI has been
[48:32] around for a long time. Google. A lot of
[48:35] the good stuff that comes out of Google
[48:36] from the search side is AI but
[48:38] pre-generative.
[48:39] >> How would you run the the type of AI
[48:42] that sits in a robot? Let's say one of
[48:44] the Optimus robots if you didn't have a
[48:46] GPU.
[48:47] >> So Matic Matic has this cleaning robot
[48:49] for example. That thing is not got a
[48:51] little GPU in it. What it has and may
[48:54] indeed have used some GPUs but no year
[48:57] as many as they need for generative AI
[48:59] to run the data feed training data into
[49:01] it so it's able to clean a house. But
[49:03] when the little buggers going around
[49:04] cleaning my floor, turdsly I call him,
[49:06] it goes around mopping my floor, it's
[49:08] not like burning money the whole time.
[49:10] But when it comes to these massive
[49:12] amount of data center, sighteline
[49:13] climate said in February there's 190
[49:15] gawatts of data centers under in
[49:17] planning. Don't know about under
[49:19] construction that works out if about 12
[49:21] million megawatt that's what like $1.6
[49:23] trillion to3 trillion a year in annual
[49:26] demand you'd need for that. We don't
[49:27] even have $130 billion worth of annual
[49:30] demand. And people say, well, it will
[49:31] grow. how when most of the demand is
[49:33] coming from Amazon feeding money to open
[49:36] AAI or anthropic, Microsoft feeding
[49:38] money to OpenAI and Anthropic, Google
[49:40] feeding money to Open AI and anthrop
[49:42] well hasn't fed it to Open AI yet, but
[49:44] they're a pretty big customer, billions
[49:46] of dollars. The conside is that we are
[49:49] building these effiges to capitalism,
[49:51] these giant GPU data centers, and people
[49:53] are being told, well, it's for AI, you
[49:56] know, the thing that's done all this
[49:57] other stuff that's unrelated. Or the
[49:59] worst thing I've seen is like, oh, you
[50:01] don't like you like online banking.
[50:02] Well, you do like data centers. There's
[50:04] a big difference between a data center
[50:05] for regular nonGPU compute for standing
[50:08] up a server, a content delivery system
[50:10] like Akami or something that brings the
[50:12] website to you or how Meta runs
[50:14] Facebook. That is not the same. It takes
[50:16] way less power, mostly CPUdriven
[50:19] compared to these giant GPU data centers
[50:21] that offer one thing, one thing only.
[50:23] >> But I was doing the the research and
[50:25] looking at some of these notes here. It
[50:27] does say that for tougher types of AI
[50:29] systems designed to solve concrete
[50:30] physics, biology, and spatial problems,
[50:32] they require some of the most intense
[50:34] data center infrastructure on the
[50:36] planet.
[50:36] >> Yeah.
[50:37] >> AI systems like Deep Mind's AlphaFold,
[50:39] the protein folding company
[50:41] >> used for genomic sequencing and climate
[50:44] forecasting, etc. run on high
[50:45] performance computing clusters. These
[50:47] require immense precision and continuous
[50:49] heavy computing data centers.
[50:51] >> Yeah. Training the brains for
[50:52] self-driving cars requires billions of
[50:54] miles of simulated physics environments.
[50:58] The AI isn't generating text. It's
[51:00] learning to navigate 3D spaces and
[51:03] gravity and relies on data centers,
[51:04] >> right? And the thing is those data
[51:06] centers, they might have GPUs in them.
[51:08] We had GPUs used for this HPC, the high
[51:12] performance computing before generative
[51:14] AI. And yeah, that's how AI has been
[51:16] trained before. That's how Tesla did.
[51:18] believe they've had their own data
[51:19] centers when it comes to training the
[51:21] autopilot system for better or for
[51:22] worse. That's how we've done it before.
[51:24] Again, that is not why we're building
[51:26] these data centers. These data centers
[51:28] are being built to sell to AI generative
[51:31] AI companies to either train systems or
[51:33] run inference. These things are being
[51:36] built in this brainless way where it's
[51:39] just well actually maybe this is a good
[51:41] way of illustrating the con because
[51:43] everyone saw Google, Microsoft, Amazon
[51:47] and Meta give Nvidia over call it 800
[51:52] something billion dollars
[51:54] because everyone saw that they went well
[51:56] they wouldn't do that for no reason.
[51:57] They went we got to build more of these
[51:58] things. There must be all this demand.
[52:00] Even though the demand 70% or more of
[52:04] all that demand comes from these two
[52:05] companies who were funded by these three
[52:07] companies and that's the funny thing.
[52:10] The reason that they don't want to break
[52:11] out their AI revenues is because it will
[52:14] become alarmingly obvious that this was
[52:16] the case. It turns out that the only
[52:18] real big customers cuz it's not like
[52:21] they're building a few data centers.
[52:22] They're building trillion plus revenue
[52:25] potential. They believe they'll get
[52:27] speculative. It's entirely speculative.
[52:29] They're building it because they saw the
[52:31] biggest companies in the world buy a
[52:32] bunch of GPUs and they said, "I want in
[52:34] on that." They must have diverse
[52:36] customers, right? They wouldn't just
[52:37] have two unprofitable fail sons that
[52:40] they're propping up with. Christ,
[52:42] they've raised $217 billion just in
[52:44] 2026.
[52:47] >> So, we know that some of the biggest
[52:49] companies in the world are using AI,
[52:51] generative AI to write a lot of their
[52:52] code.
[52:53] >> Mhm.
[52:53] >> That is a great productivity gain for
[52:55] those companies, right? I mean, have you
[52:58] used Google or Facebook or Instagram or
[53:00] GitHub recently because they are
[53:03] catastrophically worse? Amazon Web
[53:04] Services went down multiple times
[53:06] because of their AI coding tool. How
[53:08] >> how is how is Google worse?
[53:10] >> Well, I'll tell the story of a real
[53:11] [ __ ] guy called Preaggo Ragavan.
[53:13] Previously, one of the heads of ads at
[53:15] Google in 2019, Google called something
[53:17] called a code yellow, which is when they
[53:19] said, "We've got a problem." And it was
[53:21] material weakness in query numbers which
[53:24] means the amount of times that people
[53:26] were searching on Google search. Guy
[53:28] called Ben Gomes internal at Google then
[53:29] the head of Google search says wait a
[53:32] minute to increase this number of using
[53:33] Google more.
[53:34] >> Mhm. We're going to have to I mean you
[53:37] what you're suggesting would mean we
[53:38] give worse answers because if someone
[53:40] got the answer quickly that would reduce
[53:41] the amount of queries right and people
[53:44] at Google Shashi Tako was another
[53:46] engineer was saying yeah can we please
[53:47] tell Sunda this because this doesn't
[53:49] seem good. We can't just increase the
[53:52] amount of queries. That would just mean
[53:53] that people would have to search more
[53:54] which would make the product worse.
[53:56] >> But but it would make them more money.
[53:57] You saying you'd show them more ads. So
[54:00] if you're spending more time on Google
[54:02] because Google's work,
[54:03] >> but is this linked to AI doing code?
[54:04] >> Oh, I'll get there. So
[54:07] >> this is the problem is is that this guy
[54:09] called Pragar Ragavan who's the head of
[54:11] ads at the time was pushing pushing and
[54:13] saying, "No, we need to make more
[54:14] queries happen. Got to make it happen."
[54:16] and Nick Fox who was there as well I
[54:17] believe was actually taking over Google
[54:18] search got to make them go up this is
[54:20] our new reality sometime in early 2020
[54:23] propagar ragavan takes over Google
[54:25] search from then and this is this is
[54:28] what I believe can't prove it if you go
[54:30] and look around the various SEO sites
[54:32] such journal and the various forums
[54:34] Google stripped back a lot of the
[54:36] suppression of spammy sites so that
[54:38] people would be on Google more and then
[54:40] over the course of time Google wanted to
[54:43] create more queries and Google search
[54:45] became much worse. It's why people
[54:47] always do like plus Reddit or from
[54:49] Reddit or what have you. It's because
[54:50] the actual underlying search results of
[54:52] Google had got worse. And then
[54:53] Generative AI came along and Praagar,
[54:56] wouldn't you know, it gets put to run
[54:57] part of Gemini. And Google also was
[55:00] having trouble getting people back on
[55:02] Google. And what did they think they'd
[55:03] do? Well, [ __ ] everyone's talking about
[55:05] this AI thing. We'll just put it right
[55:07] at the top so people have to stay at
[55:09] Google. And actually, they'll use it
[55:10] more because instead of searching
[55:12] websites and doing that annoying thing
[55:13] where they click away from Google,
[55:15] they'll just only use Google. Instead of
[55:17] generating answers, by which I mean
[55:20] giving you search results you click
[55:21] through, now Google is the answer. Is it
[55:23] right? God know. It might tell you to
[55:24] eat rocks, might eat poisonous
[55:27] mushrooms. Maybe it'll give you a little
[55:28] few links you could click through. But
[55:30] the ideal situation was that AI was the
[55:33] ultimate form of Google's evil which was
[55:35] >> But I'm saying here I'm saying here but
[55:36] that's not the fact that coders could
[55:39] code on Google that's made Google worse.
[55:40] That's human decisions have made it
[55:42] worse.
[55:42] >> Yes. And then there's the instability of
[55:44] Google's platform which is actually I
[55:46] should have probably led with that a
[55:47] problem across the whole tech industry.
[55:49] >> Okay. So you're saying that you're
[55:50] saying Google is going down more.
[55:52] >> Yes. Google is less stable. Google Docs
[55:55] is a bugfest right now and has been for
[55:57] a while. Google Sheets, same deal. And
[55:59] the thing is, you're right, I'm being a
[56:01] little unfair. This is everyone. It's
[56:03] the same with Microsoft. It's the same
[56:04] with Amazon. It's the same across.
[56:05] >> How do we quantify that outside of
[56:07] anecdotes? Like, is there a way to
[56:09] >> You're right. I mean, GitHub downtime is
[56:11] the best example. Amazon Web Services
[56:13] went down two or three times this year
[56:15] because of AI tools. And honestly,
[56:18] you're right. It is kind of hard to
[56:20] quantify outside of anecdotes. But I
[56:22] challenge anyone listening to this. Go
[56:23] and use a website these days and tell me
[56:24] how well it works. Tell me how buggy it
[56:27] is. Tell me how many problems even with
[56:28] my iPhone. The supposed best UX in town.
[56:32] Even the iPhone is a flipping mess these
[56:34] days.
[56:35] >> Okay, so the research says the short
[56:39] answer is yes. Tech downtime and
[56:41] software outages have demonstrabably
[56:43] increased over the last few years and
[56:45] industry data points directly to the
[56:46] explosion of AI assisted coding as a
[56:48] primary culprit. The problem is hitting
[56:51] the tech industry from two entirely
[56:52] different directions. The code itself is
[56:54] getting buggier and the sheer volume of
[56:56] AI activity is literally crashing the
[56:59] underlying infrastructure. Interesting.
[57:01] >> Yeah, that's because GitHub people are
[57:03] just writing a bunch of code, pushing
[57:04] it, and thus there's just more code on
[57:07] there.
[57:08] >> That's interesting.
[57:09] >> Yeah, it's it's a real mess as well
[57:11] because
[57:12] open source has had this problem as well
[57:14] because it's well-meaning people.
[57:15] They're like, I learned a bit of code
[57:16] with an LLM. I'm going to go out and do
[57:18] some stuff. I'm going to make this
[57:19] project better. And these people barely
[57:21] understand what they're shipping. Or
[57:23] maybe they understand a bit of code and
[57:24] they say, "Oh, Dunning Krueger, this
[57:26] [ __ ] I'm going to I'm just
[57:28] like, I can understand some of this."
[57:29] And now the code's all written and just
[57:30] push it right now. So GitHub is flooded
[57:32] with AI code.
[57:33] >> This sounds like it's making humans
[57:36] complacent.
[57:37] >> It is
[57:37] >> because we're going, "Okay, look, I let
[57:39] it write the the code for the last 100
[57:41] lines and it was broadly right. So the
[57:44] next 100 lines, I won't check them as
[57:45] much."
[57:45] >> Yeah. Yeah. And that's human nature is
[57:48] to get sort of to take shortcuts to
[57:50] spend less energy on an activity if you
[57:52] can right but the AI's still making the
[57:55] mistake and we're still making all the
[57:56] promises of AI that's the thing this
[57:59] thing is meant to be this autonomous per
[58:01] you say it can't be perfect I don't know
[58:03] based on what Samman has been saying for
[58:05] the last few years clammy Sammy has been
[58:07] promising the world saying this will
[58:09] replace software engineers Dario
[58:10] Ammedday Wario himself has been saying
[58:13] oh yeah 50% of white collar labor is
[58:16] going to go away in the next few years.
[58:18] These people are promising the world.
[58:20] Again, if they were saying it would be
[58:22] smaller and they were like, yeah, it
[58:23] does have issues and we must be none of
[58:26] this, oh, what if it wakes up and it's
[58:28] super powerful. Just like, yeah, it's
[58:30] probabilistic. It's going to make
[58:32] mistakes and if you don't know what
[58:33] you're doing, you don't really know what
[58:34] you're looking at, you're going to miss
[58:36] those mistakes and it's going to get
[58:37] multiplicatively worse as you go when
[58:40] you don't know what you're doing. So
[58:41] yeah, human nature is part of it, but so
[58:44] is the marketing. So are the promises.
[58:47] One of the smartest things a business
[58:49] can do is build like a bigger company
[58:52] without actually hiring like one. But
[58:54] the problem we all face is that most
[58:56] companies don't have every skill in
[58:57] house. So when I look at the businesses
[58:59] seeing real success today, the
[59:01] consistent pattern with all of them is
[59:03] how quickly they move. They bring in
[59:05] specialists with skills in emerging
[59:06] areas to keep themselves ahead. Even in
[59:08] our company, we spent the last year
[59:10] pulling in talent across areas like AI
[59:12] native strategy, no code builds, and
[59:15] product workflows. And we find this
[59:16] talent through our longtime partner,
[59:18] Fiverr Pro. Their premium service only
[59:20] shows you vetted talent. So, you've
[59:22] always got the safeguard that anyone you
[59:25] pull in to help you with a complex
[59:27] project has the skills that you're after
[59:29] and will deliver to the same high
[59:31] standards as your internal team. And
[59:33] most importantly, they'll keep up with
[59:34] the pace. It's a simple strategy, but it
[59:36] lets us stay agile without compromising
[59:38] on quality. So, if you need these kind
[59:39] of skills in your business, head to
[59:41] pro.fr.com to find pioneering talent to
[59:44] fill your business's gaps. That's
[59:46] pro.fr.com.
[59:48] You know, the little traditional SIM
[59:49] card that goes inside of our phones.
[59:51] They haven't changed at all since they
[59:53] were invented in the '90s. You have this
[59:55] physical piece of plastic that means
[59:57] you're locked into one carrier, one
[59:59] network, and the second you cross a
[60:01] border, that carrier can start charging
[60:03] you whatever they want. But there are
[60:05] alternatives, and today's sponsor, SY,
[60:07] is one of them. It's an eim app that
[60:10] gives you a safe and secure data
[60:12] connection in over 200 destinations. All
[60:15] of their eims have built-in cyber
[60:17] security, which is great if you're
[60:18] traveling for work and looking at
[60:19] confidential material. I've been using
[60:21] SY whenever I travel because the
[60:23] connection is always reliable and it
[60:25] saves me a ton of roaming fees. It also
[60:27] means I don't have to deal with all of
[60:29] the faf that surrounds sorting out a SIM
[60:31] everywhere I go. If you want to give it
[60:33] a try, download the SY app from the app
[60:35] store now and scan the QR code on
[60:37] screen. And if you want 15% off your
[60:39] first purchase, use my code D
[60:44] when you get to check out. That's D O A
[60:47] for 15% off. Keep that to yourself. My
[60:51] my car that drives itself that is AI
[60:53] technology.
[60:54] >> Yes.
[60:54] >> I sat here with Dra from Uber and he was
[60:57] saying that I think in a couple of years
[60:59] time
[61:01] we won't need drivers um for Uber
[61:04] because the cars will drive themselves
[61:06] like they'll be fully autonomous.
[61:08] >> And I think if I'm not mistaken
[61:11] driving is one of the biggest
[61:12] professions on planet earth. So when you
[61:14] hear people when you hear these CEOs
[61:16] saying that there will be job disruption
[61:19] >> you say that they are not telling the
[61:21] truth.
[61:22] >> Yes. Or they're guessing in a way that's
[61:24] very good for them. Think about it from
[61:26] perspective of Microsoft Sachin Nadella.
[61:28] He's not going to be like yeah we don't
[61:30] know if this is going to work mate. Of
[61:31] course he's going to talk his book and
[61:32] he's going to say yeah this is going to
[61:34] replace all workers. It's going to be
[61:36] amazing. He it's going to be so
[61:38] powerful. And then he'll change his tune
[61:39] and say actually it's not going to
[61:40] replace workers. that make him more
[61:41] powerful because the things aren't
[61:43] catching up. Dor from Uber for example,
[61:45] of course he's going to say if this
[61:47] happens then that would be good for Uber
[61:49] because Uber would just become an
[61:51] autonomous taxi service. There's a
[61:53] reason that Whimo's taken I I find Whimo
[61:56] fascinating. I think that [ __ ] it's
[61:57] really cool. I think there are
[61:57] socioeconomic problems that will come
[61:59] from it. I think there are actual real
[62:00] problems that will emerge and also
[62:02] >> what kind of problems?
[62:03] >> Well, I mean socioeconomically there are
[62:05] like you said one of the largest
[62:07] employment centers in the world. I mean
[62:09] just the economics of cabs will fall
[62:10] apart but again we are nowhere nowhere
[62:12] nowhere near that. We're not even close.
[62:14] Whimo has had to do the smallest
[62:16] rollouts and the most control things
[62:18] because the problem with pretty much
[62:19] every AI system but especially driving
[62:22] is not the getting 95% of the way. It's
[62:25] those edge cases. It's raining which is
[62:27] a big problem for them in San Francisco.
[62:29] It's a kid runs across the road but
[62:30] they're wearing a high viz thing. Does
[62:32] it even notice it's a child? Again, this
[62:34] is a really interesting but very very
[62:36] applicable example of uh the right
[62:38] comparison to be made shouldn't be
[62:40] autonomous vehicles versus perfection.
[62:42] It should be autonomous vehicles versus
[62:44] human drivers. I mean, I don't know if I
[62:46] agree because a human driver might make
[62:48] mistakes, sure, but again, not an expert
[62:51] in autonomous cars. Just want to be
[62:52] clear. But if we're pushing autonomous
[62:54] cars out there willy-nilly and we're not
[62:56] doing so in extremely controlled
[62:58] environments, those edge cases will
[63:00] multiply and be dangerous. Yeah, they
[63:02] might be better at human drivers in some
[63:04] ways, but they might also I was in Vegas
[63:05] the other day and I was in a hotel and I
[63:08] watched a bunch of Zuk's cars just get
[63:10] [ __ ] stuck.
[63:10] >> They're autonomous cars.
[63:12] >> Yeah, they these weird boxy things. They
[63:14] just blocked the exit. They just all
[63:16] kind of lined up and just fell asleep. I
[63:18] saw the same thing actually happen
[63:19] outside of a hotel when I got out of a
[63:21] Whimo in San Francisco. Just stopped at
[63:22] the and then a bunch of cars and another
[63:24] Whimo got stuck behind it. And these are
[63:26] kind of
[63:27] >> I've seen some human bad drivers as
[63:29] well. I I agree, but it's just we have
[63:31] control over deploying these bad or good
[63:34] drivers. We have an ability to roll them
[63:37] out slowly, which is exactly what we
[63:39] should do. I'm not saying autonomous
[63:41] cars are bad. I'm saying we need to be
[63:43] so so so careful and treat them as
[63:46] guilty and pro till proven innocent
[63:48] because we can prove and also they have
[63:50] people overlooking them. They actually
[63:52] have people monitoring the roots. It is
[63:53] something they cannot rush out and it
[63:55] doesn't seem like they're rushing it,
[63:56] which is good. and they're not promising
[63:58] the world.
[63:58] >> I do agree. Listen, I I'm a big fan of a
[64:01] big fan of taxi drivers generally in
[64:02] part because I spend a lot of time in
[64:03] taxis and I think I'm not just getting
[64:05] in there because I want to get to from A
[64:06] to B. I'm getting in there for lots of
[64:08] other reasons.
[64:08] >> Yeah.
[64:09] >> However, when I look at the stats
[64:11] >> around what is more dangerous
[64:14] >> driving myself or having an autonomous
[64:16] vehicle drive me, there's an 68% lower
[64:19] overall crash involvement rate when
[64:21] you're in an an autonomous vehicle. Mhm.
[64:23] >> Autonomous vehicles experience roughly
[64:25] 2.1 police reported crashes per million
[64:27] miles compared to humans that are at
[64:29] roughly 4.68 per million miles. So, a
[64:32] 55% reduction when you get in an
[64:34] autonomous vehicle. And autonomous
[64:35] vehicles show an 80 to 81% reduction in
[64:38] crashes resulting in injuries versus
[64:41] human drivers.
[64:42] >> Uhhuh.
[64:42] >> So, you're 85% less likely to be
[64:45] involved in a single vehicle crash like
[64:47] hitting a wall or a tree if you're an
[64:49] autonomous vehicle
[64:51] >> versus being driven by I agree. But
[64:53] >> so it's safer
[64:55] >> in also that data is what's the sample
[64:58] size of human drivers? I mean we've got
[65:00] many many many many many many more years
[65:02] of drivers and many many many more years
[65:04] of accidents and also man does that not
[65:06] have anything to do with generative AI.
[65:08] If we were just talking about that be
[65:11] having a different conversation.
[65:12] >> I guess the question here was really
[65:13] around job disruption. Like you know we
[65:15] we look across industries and we go
[65:16] driving is a massive profession. Is
[65:18] there going to be job disruption because
[65:19] cars can now drive themselves? If we
[65:21] think about white collar, you know,
[65:22] jobs, you know, lawyers and accountants,
[65:25] people sit here and they tell me that
[65:27] lawyers and accountants would the
[65:29] profession, right? I should say some of
[65:31] the skills within the profession will be
[65:33] relegated to AIS to do.
[65:35] >> Here's the thing. Lawyers, for example,
[65:37] great example. Always hearing [ __ ]
[65:40] legal partners talking about AI. Never
[65:42] the associates. The associates are the
[65:44] ones that go out and find the president.
[65:46] They're the ones that go and do the
[65:47] grunt work. They're the ones who are
[65:48] pulling motions half the time. The
[65:50] partner is the one that might be the
[65:51] litigant. It may be the client facing,
[65:53] but the ones that are actually doing the
[65:54] day-to-day work. I'm not hearing from
[65:56] them. I'm not hearing associates being
[65:57] like, "This is [ __ ] awesome." I'm
[65:59] hearing a bunch of well- paid people
[66:02] that have sat on Chat GPT and gone,
[66:04] "Yeah, yeah, I'm the greatest lawyer
[66:06] ever." They're not the ones that I want
[66:07] to hear from the actual workers. White
[66:09] collar labor disruption is not
[66:11] happening. Open AAI had a study that
[66:13] came out I think like a week ago that
[66:15] said there was no corre connection
[66:17] between spending on AI tokens and
[66:18] revenue per employee. Like this is open
[66:21] and that's
[66:21] >> what does that mean? Could you explain
[66:22] that to me?
[66:23] >> As in the more tokens you spend has no
[66:25] no correlation at all with the amount of
[66:28] money you make. It's the second report
[66:30] they've put out. The other one was like
[66:31] hallucinations are mathematically
[66:33] guaranteed kind of almost the one thing
[66:35] I respect about that company that
[66:36] occasion they just put out a study. It's
[66:38] like, yeah, [ __ ] kind of sucks. But the
[66:41] people that are having their lives
[66:42] disrupted work-wise are art directors.
[66:46] It's people, art directors,
[66:47] transcribers, translators, who have
[66:50] bosses that don't care about the output.
[66:52] It's what they consider cheap work. And
[66:54] the problem is is those people would
[66:56] have automated your work away anyway.
[66:57] They would have sold it. They would have
[66:59] taken the cheapest for they would have
[67:00] sold it to the global self. They would
[67:02] have taken the shittiest option they
[67:03] could. That is something that AI is
[67:05] doing. And again, those people are not
[67:06] paying the actual cost of AI. They're
[67:08] using a subscription. The actual white
[67:10] collar labor force might have some
[67:14] things that are slightly changing, but
[67:16] there is no evidence of like
[67:18] productivity gains. In fact, if there
[67:20] were, they would be screaming it from
[67:22] the rooftops. There was an Oxford
[67:24] economics study last year where it's
[67:25] like, oh, young people are finding less
[67:27] jobs because of AI. We actually read the
[67:29] study, which multiple journalists did
[67:31] not. It was a single line that said,
[67:33] "Yeah, we saw some correlation." Didn't
[67:35] give a number. Didn't actually say what
[67:37] the correlation was. We are so
[67:40] conditioned to believe that the rich and
[67:42] powerful know what they're doing that we
[67:45] internalize these narratives about like,
[67:47] well, previous booms lost a lot of
[67:49] money. Well, technology takes time to do
[67:51] stuff. And they are intentionally
[67:53] playing on those mythologies. They are
[67:55] playing on these knowing that
[67:57] journalists, analysts, investors will
[68:00] believe them. And this is partly because
[68:02] our our realities are defined by stock
[68:04] prices. Because the stock prices of
[68:06] these companies went up, we're like,
[68:08] "Oh, look, it must be working, right?"
[68:10] >> Both of those things you said were true,
[68:11] though, right? Like that previous
[68:12] technologies didn't make money at the
[68:14] start and you The other one you said was
[68:16] um they'll get better.
[68:17] >> But that's the thing. Okay. Because
[68:19] another thing got better, this will get
[68:21] better.
[68:21] >> No, but there's there's got to be
[68:22] something that they're saying that is
[68:24] fundamentally not true because those are
[68:25] two true statements that okay,
[68:27] technology often starts
[68:28] >> I know. I get what you mean. What they
[68:30] are fundamentally misleading people
[68:32] about is how possible it is. How many
[68:34] actual signs they have because they
[68:35] don't have the signs. If they had the
[68:36] signs as in the signs of this getting
[68:38] cheaper as in the signs of this being
[68:40] able to autonomously do work without the
[68:42] Rub Goldberg machine and even then in a
[68:45] reliable way that was making the
[68:47] customer more money being productive in
[68:49] a way you can say with your whole chest
[68:51] without a series of asterisks and that's
[68:54] how it is across the board. The people
[68:56] that are most excited about this,
[68:59] psychopaths on Twitter in many cases are
[69:01] people that I believe there really are
[69:03] some I'm sorry, there are some people on
[69:05] Twitter because the other thing about
[69:07] this is this is really unique to the AI
[69:09] industry. I've never seen it any other
[69:11] industry outside of maybe like sports
[69:13] teams. The attachment that some people
[69:15] online have to these companies. If you
[69:17] dare dare to criticize anthropic, it's
[69:20] almost this religious attachment. Good
[69:23] example was this week Bloomberg reported
[69:25] that OpenAI was on track to hit $40
[69:27] billion in annualized revenue. Month
[69:29] times 12, four weeks times 13, we don't
[69:31] know. They don't define it. I saw
[69:33] multiple people and I going actually
[69:35] it's 60 billion. It's actually 60
[69:37] billion. I heard from someone it is like
[69:40] a cult and it's a cult of software
[69:42] driven around growth and this idea that
[69:45] by backing the right horse you will have
[69:48] some grand thing and open AI in
[69:51] particular in particular Mr. Baltman
[69:54] they have been fermenting this that Tibo
[69:56] as well the Tibbo the one of the guys at
[69:59] uh OpenAI they ferment this thing online
[70:01] they build this kind of parasocial
[70:03] relationship with both the large
[70:05] language model themselves and the
[70:07] companies and one's allegiance to the
[70:09] companies is so important it's truly
[70:12] vile if only these people gave a [ __ ]
[70:14] about I don't know Medicare for all or
[70:17] poverty or thing like actual problems in
[70:19] the world versus are we buying enough
[70:21] GPUs Do you know what's interesting is
[70:23] some of what your narrative
[70:26] one would argue actually helps them.
[70:29] How? Because you know the AI doomers
[70:31] that have come here and told you know
[70:32] some of the original founding fathers of
[70:34] AI like Jeffrey Hinton have told me that
[70:37] what they're building is highly highly
[70:38] dangerous and that it will be
[70:40] fundamentally disruptive to society. And
[70:43] it's interesting because some of the
[70:45] CEOs who you've mentioned, their
[70:46] historical narrative was also, by the
[70:48] way, this is really [ __ ] dangerous
[70:50] and there is a significant chance it
[70:51] could f we could [ __ ] up the planet.
[70:53] >> And what we've seen is this slow pivot
[70:55] away from it because now they're getting
[70:57] booed and they're being attacked.
[70:59] There've been this slow pivot away from
[71:00] it. And the pivot almost sounds a little
[71:04] bit like your narrative.
[71:05] >> It now sounds like actually no, it's not
[71:07] going to change anything and you're all
[71:08] going to be fine. And it's now there's
[71:10] just not it's nah it's not dangerous at
[71:11] all.
[71:12] >> But that's the funny thing
[71:13] >> and that's why I'm saying like you're
[71:14] you're not they I actually think there
[71:16] might be a couple PR people at these big
[71:18] AI companies thinking thank god for Ed
[71:22] some of it because you're like you're
[71:24] saying actually don't worry everything's
[71:25] going to be fine. It's not going to take
[71:26] your job. It's not going to disrupt the
[71:28] economy. It's just a fad. There's no
[71:29] technology. And I think they don't think
[71:31] that.
[71:31] >> Here's the thing. I think Alman and
[71:33] Amday are some of the most deeply
[71:34] corrupt and cynical people in the world.
[71:36] I don't think of course they were going
[71:37] to say from the it was early 2023 or man
[71:40] said we're a little bit scared about
[71:41] what we're creating. Oh, shut up. I'm
[71:44] just I hear that and I feel so
[71:46] frustrated because I've met so many of
[71:48] these rich [ __ ] liars, these people.
[71:50] And you know why he wants to say that?
[71:52] So you'll invest in his company and buy
[71:54] the software. So you'll be scared that
[71:56] if you don't use AI today, you'll be
[71:57] left behind in the future, which is
[71:59] their continual narrative that if you
[72:01] don't get on the train today,
[72:03] then you'll be left behind. By the way,
[72:05] every single scam and con starts with
[72:07] rushing you. Every single trick in
[72:10] history begins with saying you must do
[72:12] this now. And best piece of advice I
[72:14] ever got was if anyone tries to rush you
[72:16] and it's not literally a mortal thing
[72:18] like you are bleeding or on fire or the
[72:19] house is on fire, slow down. And yet all
[72:22] of these companies saying it's so scary.
[72:24] And now they're talking about slowdowns.
[72:26] But you ever noticed that Amade and
[72:28] Ortman, they say, "Oh, maybe we should
[72:29] slow down progress." And then they
[72:31] don't. Right now, Orman's saying, "Oh,
[72:33] we slow down progress because we're so
[72:34] delayed." No, they're out of compute.
[72:36] Now, they're doing it. I can guarantee
[72:37] you, by the way, their PR people do not
[72:39] like me. I know for I know I don't think
[72:40] OpenAI's PR people are super fond of me.
[72:43] >> But I bet there's elements of what
[72:44] you're saying because you're calming
[72:46] people. You You are theoretically
[72:47] calming down the general public.
[72:49] >> And you know what? I hope I am because
[72:51] >> the fear based tactics is horrible.
[72:53] These companies don't want that. These
[72:54] companies want people scared. I'm 100%
[72:56] sure.
[72:57] >> Uh I don't I just fundamentally
[72:59] disagree. I think it
[73:00] >> can I so the timelines there and I sit
[73:02] here and what I do is I log their quotes
[73:04] over time
[73:05] >> and I read them out from 2015
[73:08] >> to 2026 and the change you see is them
[73:12] going from there could be extinction
[73:14] that's the narrative the early narrative
[73:16] Elon said it himself he says it's the
[73:17] single most dangerous thing in
[73:18] >> Elon and then you track it over time and
[73:21] it evolves to this age of abundance
[73:23] we're all going to have unlimited stuff
[73:25] and then um the the new slogan at
[73:28] trackbt is intelligence for everyone.
[73:30] It's suddenly and all the and and
[73:32] whenever Daario comes out and says, "By
[73:34] the way, it's really [ __ ] dangerous."
[73:35] They attack Daario. Yeah. They hate him.
[73:38] >> That man Daario is
[73:40] >> They're like, "Dario, shut the [ __ ] up."
[73:41] >> Honestly, I I've been saying Dario, shut
[73:44] the [ __ ] up for years. But it's But the
[73:46] thing is, I get your point where it's
[73:47] like I don't think they've changed to
[73:49] calm the public down so much as they're
[73:51] desperate to not get regulated, which is
[73:53] laughable. We don't regulate tech. We
[73:55] don't regulate [ __ ] America doesn't
[73:57] regulate [ __ ] We are in the We are
[74:00] still trapped in the hands of Milton
[74:02] Freriedman, Margaret Thatcher, and
[74:04] [ __ ] Ronald Reagan. We're still stuck
[74:06] in the neoliberalistic hellscape, which
[74:09] is growth at all cost, free market
[74:11] capitalism. So, no, no one's regulating
[74:13] the regulation of these companies should
[74:15] have been, I don't know, breaking up.
[74:17] Put these bastards to the side. Break up
[74:19] these [ __ ] for sure. We shouldn't
[74:20] have companies this big. It makes things
[74:22] worse.
[74:22] >> But these technologies are dangerous.
[74:24] >> I mean, they're dangerous, but not in
[74:26] the ways they've been warning about.
[74:27] Let's if we think about cyber hacking,
[74:30] >> right? And just to be clear, those cyber
[74:32] hacking things that happened were not a
[74:33] result of they were like break out of
[74:35] the sandbox and then they set the
[74:36] sandbox up wrong. They set up the server
[74:39] they were on wrong. But I mean, you
[74:41] know, advanced AI models could very
[74:43] easily cuz they can go out onto the open
[74:45] internet as agents. They could very
[74:47] easily go and look at code bases of
[74:48] different websites, find vulnerabilities
[74:50] and exploit those vulnerabilities.
[74:52] >> Yeah. in at scale and arguably um at a
[74:56] higher intelligence and faster and wider
[74:59] than humans a human hacker could
[75:01] theoretically. So that's dangerous.
[75:02] >> Well, here's the funny thing. We don't
[75:05] know how much compute was spent to do
[75:07] the hugging face attack, the open AI
[75:09] one. We also do know that they
[75:10] improperly set up the server to keep it
[75:12] in. They thought they'd turn the
[75:13] internet off and they didn't. That's
[75:15] human error. And that's human error in a
[75:17] sense that yeah, they threw about an
[75:19] indeterminately large amount of compute.
[75:21] This is dangerous, but people keep
[75:23] saying we can't let the the Chinese get
[75:25] a hold of these models. We couldn't
[75:27] possibly because what if these models
[75:28] fall into the wrong hands? They're
[75:30] already in the wrong hands. Mark
[75:32] Zuckerberg, Sam Olman, Dario Amade. The
[75:35] wrong hands are the hands of those who
[75:37] are running these companies. We should
[75:39] not be training these models to do these
[75:41] things. I don't know why the [ __ ] we're
[75:43] doing it other than they've run out of
[75:45] other things they can train on. There's
[75:46] a ton. And the fact that they can do it,
[75:48] it's kind of interesting. But you do
[75:50] would you agree that it's an
[75:52] intelligence and I'll call it that you
[75:54] know you might disagree with that
[75:55] terminology but an intelligence that can
[75:57] go out onto the internet and click
[75:59] around and take actions is inherently
[76:03] there's risks associated with that. Well
[76:06] the second part I agree with the risks
[76:08] we've had people running automated
[76:10] scripts hacking scripts for a while
[76:11] we've had hackers doing that for years
[76:12] and years and years. This is brute
[76:14] forcing it with a bunch of compute and
[76:16] yet it is dangerous. These companies are
[76:18] doing something dangerous. That is not
[76:21] what Jeffrey Hinton at have been warning
[76:23] about. They've been saying, "Oh, these
[76:25] things could destroy society. They could
[76:26] manipulate people." When you actually
[76:28] look at the underlying things, not so
[76:29] much. Jeffrey Hinton as well talking his
[76:31] book still got his Google stock, I
[76:32] think. And weirdly enough, he left
[76:34] Google because he was worried about the
[76:35] AI there, but then immediately made a
[76:37] comment being like, "Yeah, actually
[76:39] though, Google's very responsible."
[76:40] Strange thing. But let's get back to the
[76:42] the cyber security side. I agree this is
[76:44] dangerous. These people should not have
[76:46] access to so much comput. They clearly
[76:47] don't know what to do with it. There's a
[76:49] really easy way of dealing with this.
[76:51] It's not letting them use so much
[76:52] compute. It's regulating that part out
[76:54] of existence. What if the Chinese do it?
[76:57] The Chinese were able to distill the
[76:58] models. And also,
[77:01] I don't know, regulate it and stop I I
[77:04] feel like with this particular thing as
[77:06] well, we got to this point and let the
[77:09] genie out of the bottle to use an
[77:11] annoying Samman term. We let this happen
[77:14] because we let these companies be
[77:15] unregulated and use as much computers we
[77:17] want. We had these [ __ ] enablers
[77:19] allowing them to burn as much computers
[77:20] as they want. And also we for all of
[77:24] these dire warnings about AI dangers, no
[77:26] one seems to have [ __ ] done anything.
[77:28] >> Okay, we're going to play a game, Ed.
[77:29] >> Let's play it.
[77:30] >> On these cards here,
[77:31] >> I have the things that you consider to
[77:33] be myths about the AI industry.
[77:37] >> The challenge is I want you to give me
[77:39] one sentence.
[77:40] >> [ __ ] on each myth.
[77:42] >> Oh, Christ.
[77:43] >> So, just your first reaction. You're
[77:44] going to pick it up, you're going to
[77:45] read it,
[77:45] >> and then you're going to give me one
[77:46] sentence on your opinion of that
[77:49] >> um belief.
[77:50] >> Okay, let's go.
[77:51] >> So, let's do this.
[77:56] >> What does it say in your says the the AI
[77:59] industry is creating enormous economic
[78:01] growth?
[78:02] >> No, it's not. It's nowhere in the data.
[78:05] >> Okay. Like, it's just May I do a second
[78:07] sentence?
[78:08] >> Go ahead. pretty much all of the
[78:10] economics is either Nvidia feeding money
[78:12] to it companies like Corewave or these
[78:14] three companies feeding money to these
[78:16] ones to spend it with the them.
[78:18] >> Okay. And what evidence do you have that
[78:20] there's it's not causing economic
[78:22] growth?
[78:23] >> Just to be clear, other than the spend
[78:25] on semiconductors, so the speculative
[78:27] investment in GPUs and data center
[78:29] infrastructure that's happening, but as
[78:31] far as like spend on AI goes, barely
[78:33] cracking hundred billion. And most of
[78:35] that is just these two running their
[78:37] services and paying these three
[78:39] companies, Oracle, Core, and others.
[78:41] >> But a hundred billion is a lot of money
[78:43] for a relatively new technology.
[78:45] >> Not when you've spent $300 billion in
[78:47] equity funding. And it if we're going
[78:50] with just these three, I think $600
[78:52] billion in capital expenditures.
[78:53] >> Yeah, I get that. That means it's not
[78:55] profitable. But the hundred billion is
[78:57] an expression of consumer demand
[78:58] >> when the compute is mostly driven by
[79:00] subscriptions that subsidized. No, it's
[79:02] not. When you're giving someone $20 or
[79:04] $40 for a dollar, they're going to use
[79:06] it more. If this was all on a per
[79:08] million token basis, we'd be having a
[79:09] different conversation.
[79:10] >> Okay, fair. Fine. Cool. Next one.
[79:14] >> The United States need to spend
[79:15] trillions to beat China in the AI race.
[79:19] Let's see.
[79:21] What AI race?
[79:23] That's actually That's actually my
[79:25] point. It's what AI race is there. Is it
[79:27] to make big scary LLMs? They they did
[79:30] that already without the Nvidia GPUs. By
[79:32] the way, they've got Blackwell GPUs.
[79:34] Kakashi and Jastario, two amazing
[79:35] analysts I love. They've been on this
[79:37] for years. It's like China's already had
[79:40] Nvidia GPUs that they're not meant to
[79:41] have for years. But also to do what?
[79:43] They already got the LMS. What What's
[79:45] the race to do? To make us spend more
[79:47] money than them? For us to constantly
[79:48] piss our pants worrying about China?
[79:50] Because u they won if that's the case.
[79:53] Myth number three, AI will replace all
[79:56] human jobs.
[79:58] that just isn't happening and there's no
[80:00] economic data to support it.
[80:02] >> Will it replace some jobs?
[80:04] >> I mean, it's replaced some contract
[80:06] labor that would otherwise be replaced
[80:07] with cheap labor out in the global
[80:09] south. It's a digital globalization in
[80:11] that sense, but all jobs, most jobs, a
[80:15] lot of jobs. No.
[80:16] >> What about robotics?
[80:17] >> Robotics is not what we're talking
[80:19] about. Robotics is a very different
[80:20] thing. And even then,
[80:21] >> robotics will be powered by AI.
[80:23] >> I mean, yes, but there are tons of
[80:24] different kinds of AI. We're talking
[80:26] explicitly about generative AI. And
[80:27] that's what I this mythbusters piece
[80:29] that was definitely about generative AI.
[80:31] >> Okay. But what about robotics? Like the
[80:33] thing is the Optimus robot that Elon's
[80:35] working on at Tesla.
[80:36] >> The one where even in the demo of the
[80:39] hand he like they had to have a guy
[80:41] controlling it. Wasn't doing it
[80:42] autonomously. Here's the thing. If they
[80:44] can beat all these challenges, yeah,
[80:46] robotics would be really cool. I don't
[80:48] know how long that's that's one I'd
[80:50] actually be willing to believe in a
[80:52] couple decades.
[80:53] >> Have you seen them ch them Chinese
[80:55] robots? I know you've seen them. the
[80:56] uni, what's it called? The one that can
[80:58] dance and that, but they can't really do
[81:00] human things.
[81:01] >> Well, it's just it is pretty
[81:02] mindblowing.
[81:04] >> Robotics are [ __ ] cool. I like I'm
[81:06] not going to pretend. I don't think
[81:07] robots are cool. I wish they were
[81:09] building robots and actually doing cool
[81:11] [ __ ] I wish the tech industry still
[81:12] made fun stuff and interesting stuff.
[81:14] Instead, we get these [ __ ] large
[81:16] language models. But with AI plus
[81:18] robotics is, you know, I was in San
[81:20] Francisco and I went to this massive um
[81:22] incubator there. And when I'd gone there
[81:24] three years earlier, it was all software
[81:26] startups, right? And when I went back
[81:27] three years later, it was all these
[81:29] robot startups. And I remember saying to
[81:30] the founder of the incubator, I was
[81:32] like, "Why is everything robots now?"
[81:34] There was this one robot where it was
[81:35] just the arm and it had a frying pan on
[81:37] it. Yeah.
[81:38] >> And it whole thing is it cooks for you.
[81:39] >> Yeah.
[81:40] >> So it was he was showing me it cooking
[81:41] whatever. And he goes, "Well, you know
[81:43] the arm." He goes, "The the hardware
[81:45] part, the physical parts,
[81:47] >> that's always been fairly cheap." Yeah.
[81:48] >> He goes, "The expensive part was the
[81:50] intelligence. And now that's come down
[81:52] to pennies." So what you're seeing is
[81:53] this explosion in the robotics industry
[81:55] because robotics is a function of
[81:57] intelligence plus hardware. We've always
[81:58] had the
[81:59] >> and a ton of data though as well and the
[82:00] data is very expensive.
[82:02] >> Yeah.
[82:03] >> The thing is cyber cabs rolled out real
[82:05] slow. It's going to take a long time. It
[82:08] could be a threat if they do a robot
[82:10] that could replace a human job. Sure it
[82:12] could. But that human jobs are
[82:13] multifaceted. Human jobs change with
[82:15] environments. And also a lot of human
[82:17] jobs that you might think of like I
[82:19] don't know dishwashing robot for
[82:21] example.
[82:21] >> Yeah.
[82:22] some guy at a restaurant isn't paying 10
[82:24] 20 grand for a robot to replace the job
[82:26] that they're already not paying enough
[82:28] for. The point is, yeah, it could if you
[82:31] can replace the jobs. That is not what
[82:33] we're talking about with this.
[82:34] >> Yeah. I I just I just I ask these
[82:36] questions not because I'm trying to be
[82:38] like I actually I'm trying to form my
[82:39] own opinion on these things and
[82:42] >> I I do think, you know, as it's written
[82:45] there, it says AI will replace all human
[82:48] jobs. Obviously not. Obviously, that's
[82:49] [ __ ] Yeah.
[82:50] >> But um I'm trying to figure out if the
[82:51] truth is somewhere in the middle that
[82:53] there's a certain type of job which
[82:55] actually humans probably shouldn't have
[82:57] ever been doing really.
[82:58] >> Um if you think back through history,
[83:00] there was someone's job just to sit in
[83:01] an elevator and press the buttons.
[83:02] >> That's an example of a job that humans
[83:04] probably shouldn't have been doing. And
[83:05] as technology gets more advanced, it
[83:07] takes on a lot of that
[83:09] >> sort of automated monotonous stuff.
[83:11] >> Right? The thing is with this particular
[83:14] thing that I know that this is from,
[83:15] it's a specific blog I wrote. I was
[83:17] explicitly talking about generative AI
[83:18] though. I was explicitly talking about
[83:20] people when they say this they are
[83:22] referring to that.
[83:23] >> So you're not talking about agentic AI
[83:24] which is
[83:25] >> agentic AI is LLMs. Agentic AI is just a
[83:27] fancy way of saying an LLM talking to
[83:29] another LLM with a harness on top. That
[83:31] is still LLM. Agentic AI is one of the
[83:34] big the bigger lies they to tell. It's
[83:35] like when you hear agent you're meant to
[83:37] think autonomous AI can do what you
[83:38] want. It's still LLMs. It's still LM
[83:40] talking to other LMLs
[83:42] >> taking screenshots and putting them in
[83:44] LLM and stuff.
[83:44] >> Oh god. Yeah.
[83:45] >> Okay. But but you know I could I could
[83:47] make the case that
[83:49] I'm just thinking about my personal
[83:51] usage. I definitely use agents to do
[83:54] things that I would have previously
[83:55] asked people to do. It's not to say that
[83:56] I didn't I still don't hire cuz we're
[83:57] hiring like crazy.
[83:58] >> Yeah.
[83:59] >> And I still in that particular function.
[84:00] I'm thinking about like the chief of
[84:02] staff role. So my chief of staff would
[84:04] have triaged all of my inboxes
[84:06] previously and put them somewhere and
[84:08] told me about them or maybe once upon a
[84:09] time shown me a piece of paper back in
[84:11] the day. I guess now my chief of staff
[84:13] is no longer doing that job. You still
[84:14] have a chief of staff though.
[84:16] >> This is what I'm saying. They're doing
[84:17] other things,
[84:18] >> right? But the thing is again what you
[84:20] were describing is
[84:22] fairly basic automation. I don't know
[84:23] what the tasks are triaging.
[84:25] >> Basic spend a trillion dollars on
[84:27] triaging email. Like that's the the
[84:29] promise. If they'd spent $10 billion and
[84:31] this was much smaller and you I go cool
[84:33] software. Yay. A lot of the things that
[84:35] people are impressed with like script
[84:36] stuff as well. It's just LM's doing
[84:38] Python. You should be impressed by
[84:39] Python code. Python's incredible. You
[84:41] can scrape websites. You can download
[84:43] [ __ ] It's awesome. But the point I'm
[84:45] making is none of this would be anywhere
[84:47] near as much of a problem if they didn't
[84:50] ask for all of the attention, all of the
[84:51] money, and promise the world. It's their
[84:53] promises that are the problem. And the
[84:55] journalists who went along with it, and
[84:56] the analysts and the Twitter people who
[84:58] went along with this, saying that this
[84:59] would change everything and replace
[85:00] everything and leaving the realm of
[85:02] reality. Is there any technological
[85:04] innovation through history that was
[85:06] really, really game-changing where that
[85:08] didn't happen?
[85:10] I mean
[85:12] the internet
[85:13] >> I mean people overpromised that
[85:15] >> I mean they overpromised on the
[85:16] businesses but I've read through a great
[85:19] many pieces about the early internet a
[85:21] lot of people were excited but hesitant
[85:24] they were worried that there was not
[85:26] enough demand but they were still like
[85:28] oh yeah this could have potential
[85:30] ramifications if it happened. People
[85:32] were not super negative about the
[85:34] internet. A lot of the skeptics were
[85:36] saying we're worried about an overload
[85:37] of bad information. Look at where we
[85:39] are. A lot of people were worried about
[85:41] the social consequences of everyone
[85:42] talking online, which they were correct
[85:44] about. With the economic things, they
[85:46] were specifically talking about like the
[85:47] globe, which I think made hundreds of
[85:49] thousands of dollars and had like a I
[85:51] think a billion dollar market cap, but
[85:53] they were talking.
[85:54] >> Yeah, there was massive hype in the com
[85:56] era.
[85:56] >> I read a lot of those stories. The hype
[85:57] was nowhere in it. You didn't have
[85:59] articles everywhere that were saying if
[86:01] you don't get online, you'll be left
[86:02] behind. You didn't have professional
[86:05] consequences. Nick Sesh mentioned his
[86:07] blog earlier. He described this thing
[86:08] global uh AI sisterating global
[86:11] decision-m where he said that you have
[86:13] businesses you work at where if you
[86:16] don't say that you're more productive
[86:17] with AI whether or not it's true is
[86:19] irrelevant you have professional
[86:21] consequences you can get fired there are
[86:23] people having to AI wash their jobs by
[86:25] saying AI did it otherwise their bosses
[86:28] who don't do [ __ ] will get mad at them
[86:31] this did not happen with the internet it
[86:33] was not present and part of the thing is
[86:35] social media was not like it is today
[86:37] the kind of uh was it decentralization
[86:40] of media in general has caused this as
[86:42] well and also the fact of day trading
[86:45] there's so many different things that
[86:46] are different it's crazy
[86:47] >> I I do think AI is different from the
[86:50] internet in part if you just measured it
[86:52] on the speed of adoption especially if
[86:54] we just think about generative AI AI
[86:56] >> but the this adoption of the internet
[86:58] required physical connections to your
[87:00] house the adoption of generative AI
[87:02] involves having a web browser it took a
[87:04] vast amount of effort to bring internet
[87:06] to people Even with dialup connections,
[87:08] it still required the distribution
[87:09] >> and that's why it was so slow and there
[87:11] was less, you know, there was less hype
[87:13] than AI. I do agree that there's way
[87:14] more hype and we again going back to
[87:16] this point that we're clustering AI in
[87:19] this big category of lots of different
[87:21] things.
[87:21] >> There's generative AI.
[87:22] >> There's generative AI. There's like real
[87:24] world AI.
[87:24] >> Generative AI is explicitly what I'm
[87:26] talking about here. When bosses are
[87:27] saying you need to use AI, they're not
[87:29] saying I need you to go and buy a
[87:30] Unibeam robot. They're saying use LLM so
[87:32] that I and that's the thing. They have
[87:35] this theory, the era of the business
[87:36] idiot where it's like we are ruled by
[87:38] people that don't do work because nobody
[87:39] who actually does a bunch of work who
[87:41] really is productive is harassing
[87:44] someone who works for them for not being
[87:45] productive enough.
[87:47] >> They're not they don't have the time.
[87:48] They're doing work. Someone who is
[87:50] sitting there with the ingratiation
[87:51] machine that's telling them that every
[87:52] beautiful idea out of their messy little
[87:54] skull is amazing. Yeah. They're going,
[87:57] "Damn, this thing says I'm a genius. Why
[87:58] are you not using the genius machine to
[88:00] do more work?" And yeah, if you're a
[88:02] boss that goes to lunch, leaves lunch,
[88:04] and sometimes reads your emails, LM are
[88:06] magic.
[88:06] >> I, you know, one of the most compelling
[88:08] arguments I have for the overhype of AI
[88:12] >> in a world where everybody has access to
[88:14] these tools, whatever the tools can do,
[88:17] would largely be commoditized. What the
[88:19] tools can't do, which one could say is
[88:22] the human taste, judgment, you could say
[88:25] it's people, skills, whatever you want
[88:26] to say, is now going to be the valuable
[88:29] thing because the scarce and the hard
[88:31] becomes the most valuable through
[88:33] history and the commoditized becomes the
[88:35] least valuable. So the very nature that
[88:37] we're commoditizing, the generation of
[88:39] content or whatever you want to call it,
[88:40] code means that's actually not where the
[88:42] value will acrue as for the user. And
[88:45] actually if you think about what it
[88:48] takes to now make something that is
[88:50] objectively great if an AI can do it
[88:54] then it's not the the great thing is not
[88:56] of value.
[88:57] >> So so I think a lot I've been thinking a
[88:59] lot actually about how
[89:01] >> how do you um avoid the temptation of
[89:04] sloppification of the things you make
[89:06] the value you put into the world. It's
[89:08] very simple example that people will be
[89:09] able to relate to. If you use chat GBT
[89:12] or anthropic, you know, Claude to make
[89:14] your LinkedIn posts, let's say,
[89:16] >> they will be [ __ ] LinkedIn posts because
[89:18] everybody else is using them. And
[89:19] actually, a great LinkedIn post now is
[89:21] someone who doesn't use them and makes
[89:23] something that's like irreplaceably
[89:24] human,
[89:25] >> right?
[89:25] >> And deeper and more personal N of one
[89:30] lived experience.
[89:32] >> Yeah.
[89:32] >> All these things that AI can't do. And I
[89:34] think that's a compelling argument that
[89:35] actually the commodity tools produce
[89:38] commodity outcomes. So everyone has
[89:40] access to these things and what's
[89:41] changed? Like really like what
[89:42] >> the slopification we've we've got a
[89:44] bunch of slop but these people were
[89:46] halfassing their jobs before. It's just
[89:47] a halfass arcery machine and it's just
[89:50] it's it's the thing. It's what I'm
[89:51] talking about with the slot blogs. It's
[89:53] like it's it yeah people that gave you
[89:55] dog [ __ ] before have now got the dog
[89:56] [ __ ] machine to pump out dog [ __ ] It's
[89:59] so there's a guy called Carl Brown uh
[90:01] internet bucks. Awesome guy. Great
[90:02] software engineer. He he said I might
[90:05] have said this earlier. So, it makes the
[90:06] easy things easy, the hard things
[90:07] harder. When you know you're doing a
[90:08] really distinct small script for
[90:10] something and it can plop that out. It's
[90:12] awesome. I used Claude the other day for
[90:14] something useful. My kid loves
[90:15] Minecraft. I was trying to fix a [ __ ]
[90:17] broken mod cuz he loves his wither
[90:19] storm. It's awesome.
[90:20] >> And it still took me half an hour and
[90:22] kept getting things wrong. What do you
[90:24] use AI for? Generative.
[90:25] >> I really don't. I don't use it
[90:27] >> with Bloomberg terminal. I use AskB,
[90:29] which is just when it's like requesting
[90:31] the consensus analyst estimates for
[90:32] Nvidia,
[90:33] >> but otherwise you don't use it.
[90:34] >> No. So, how do you know it's bad? I've
[90:36] used it. I've put it through its paces.
[90:38] I've used it to try and do financial
[90:39] models and found one error and
[90:41] immediately be like, "Ah, I've never
[90:42] been particularly impressed." The one
[90:44] thing I will defend it on is it's really
[90:46] good for like tech support. Like I have
[90:48] this thing called Synergy in my New York
[90:50] New York place I go to. I have this
[90:51] monitor where I have a MacBook and a PC
[90:53] laptop and this thing Synergy for using
[90:55] the same mouse and keyboard.
[90:57] >> Dropping a giant [ __ ]
[91:00] troubleshooting log into this thing and
[91:01] going, "What's wrong?" And it going,
[91:03] "This is wrong." Yeah, super useful. Is
[91:05] that trillion dollars? No. Is that a $2
[91:07] trillion company? No. Pretty use.
[91:08] >> Better than Google though, right? Better
[91:10] than Google search.
[91:10] >> I know. I mean, yeah. Remember,
[91:12] >> do you use Google search still?
[91:14] >> I try. I have to [ __ ] push the crap
[91:16] out of the way. And
[91:17] >> I can't remember the last time I did a
[91:20] Google search.
[91:20] >> Christ, I find myself using Bing
[91:22] sometimes. I know. I hate saying it,
[91:24] too. But I have to scroll past the AI
[91:26] crap cuz I want the good stuff. I want
[91:28] the I want the actual links to stuff so
[91:30] that I can read the thing and go. But
[91:33] you can ask the AI to give you the
[91:35] links.
[91:35] >> Yeah. And it doesn't do a particularly
[91:37] good job. Like my
[91:38] >> So say that the other day my iPad wasn't
[91:41] turning on and it was doing this funny
[91:42] little thing on the screen. You think
[91:43] that it's better to type that into
[91:45] Google than
[91:46] >> Oh, no. I must be clear that may be the
[91:48] only LLM use case I defend. The
[91:50] troubleshooting thing is awesome for it.
[91:52] I It's the the one weakness I have. It's
[91:54] like genuinely being able to drop a log
[91:56] into it. That's awesome. Again, that is
[91:59] not what they're selling it as. They're
[92:00] not selling it as a useful little tool.
[92:02] They're selling it as the uh software as
[92:05] the thing that will change everything
[92:07] that will replace all jobs that will do
[92:09] this and that. It's not like they sold
[92:11] it as a quirky bit of software.
[92:12] >> No, you are right. They are, you know,
[92:14] telling us that it is going to replace
[92:15] everything. But funnily enough, the
[92:17] critics are saying that as well.
[92:18] >> Which one I mean I mean
[92:19] >> they are like the Jeffrey Hintons of the
[92:21] world. you know, even people that have
[92:23] left the safety team in chat who who
[92:25] I've sat here with the these are critics
[92:27] that are that are warning of the impacts
[92:30] it's going to have on the world. It's
[92:31] weird how all these critics also have
[92:33] vested interest in AI doing well though.
[92:35] Daniel, former open AI guy, AI 2027
[92:38] written with the Star Codeex guy that
[92:40] was nothing more than badly written
[92:42] science fiction that he's already had to
[92:43] walk back.
[92:44] >> You know, he could have made more money
[92:45] by staying at chat.
[92:47] >> Could he?
[92:48] >> I mean, looks like he lost
[92:49] >> if he had options early. it sticking
[92:52] around.
[92:52] >> Did he lose the options? How much do
[92:54] they
[92:54] >> You're not saying that they're they're
[92:56] being critical. They're not critical of
[92:58] the companies themselves. They're not
[93:00] critical of the stealing. They're not
[93:01] critical of the environmental damage.
[93:03] They're not critical of the fact that
[93:04] you cannot rely on the answers. They're
[93:06] critical of this big scary boogeyman out
[93:09] in the future where it's like, "Oh, I'm
[93:12] scared of when this becomes so powerful
[93:13] and everyone should talk to me about how
[93:15] scary and powerful it is." They're not
[93:17] saying, "Hey, here are the harms today.
[93:18] Here are the things we're actually
[93:20] looking at today. Here are the social
[93:21] problems of having this automated way of
[93:25] spewing out slop, of filling our feeds
[93:27] with crap, of having information that
[93:30] will pop up that is presented even with
[93:31] the little disclaimer thing of saying,
[93:33] "Yeah, sometimes this gets [ __ ] wrong."
[93:34] So, in the tiniest words possible, they
[93:37] don't talk about the fact that these
[93:39] things are trained on stealing millions
[93:41] of people's work. But on that last point
[93:42] where you say that it's going to get
[93:44] progressively more intelligent and when
[93:45] it does, it will be a danger.
[93:46] >> Yeah. Would you agree with the statement
[93:49] that artificial intelligence has gotten
[93:51] more intelligent
[93:53] if you measure it based on any sort of
[93:55] measure of intelligence one might use?
[93:57] >> It's got better on the tests that are
[93:59] rigged for the models. It's got better
[94:00] at tests where you can train for the
[94:02] test.
[94:03] >> Okay, so it's got better at
[94:04] >> it's got better at tests that they're
[94:06] intentionally trained for.
[94:07] >> So if you logged the rate of improvement
[94:10] on a graph, it would look something like
[94:12] this,
[94:13] >> right?
[94:14] >> You agree? in terms of what it's capable
[94:16] of doing.
[94:17] There we go. Yeah,
[94:18] >> cuz it's not it's not got new features.
[94:21] You'll notice that outside of OpenAI and
[94:23] Anthropic the VA when you remove the
[94:25] coding startups, there's basically no
[94:27] successful AI startup company.
[94:29] >> So, we agree that it's got better. It's
[94:31] got more capable
[94:34] at doing things.
[94:35] >> Yeah. Okay. Over time, AI's got more
[94:37] capable. If we imagine that trajectory
[94:41] will continue, it will get more capable.
[94:43] Then at some point it does cross you
[94:46] know this is what they say to me it
[94:48] crosses human intelligence and at such
[94:50] time
[94:51] >> will it not start to do some of the jobs
[94:54] that people are doing today
[94:55] >> outside of software engineering remove
[94:57] software because I will concede software
[94:58] engineering it's got better at that
[95:00] outside of software engineering where
[95:02] >> so the chief of staff things that admin
[95:04] >> okay so it's got better admin video
[95:06] generation photo generation
[95:08] >> text generation theoretically coding
[95:11] >> right
[95:12] >> and then I'd say agentic workflows. So
[95:14] >> what is an agentic workflow?
[95:15] >> So automated workflows where you're
[95:17] doing the same I mean a good example is
[95:20] looking at the backend data of the dire
[95:21] of a CEO
[95:22] >> summarizing
[95:23] >> looking at all of the data ingesting all
[95:24] of it going out into the internet and
[95:25] searching who Ed is
[95:27] >> looking at every interview you've ever
[95:28] done ever.
[95:29] >> Uhhuh.
[95:30] >> This is summarizing and generating
[95:32] >> making a little model on you know the
[95:33] things people want to know from Ed.
[95:35] >> Producing a report sending that to my
[95:37] inbox.
[95:38] >> Me getting a 20 30 40 50page report on
[95:40] Ed before he arrives.
[95:41] >> This is all basically the same thing. I
[95:42] think it's been doing for years though.
[95:44] It's It's not really new capabilities.
[95:45] >> Research. It's It's
[95:48] >> still the same things. They've had web
[95:49] search for years. They've had report
[95:51] generation for years.
[95:52] >> Well, we couldn't generate
[95:54] highquality videos that are like
[95:56] indistinguishable from cameras. Seed
[95:58] dance and these ones that look like
[96:00] movies.
[96:01] >> I mean, they
[96:01] >> are incredible.
[96:02] >> So, I'm saying the point I'm trying to
[96:04] make is that if we imagine that over the
[96:05] last 10 years there has been a rate of
[96:06] improvement in terms of capabilities and
[96:08] output and quality. We've seen
[96:10] hallucinations drop. We've seen the
[96:12] models get more quote unquote
[96:14] intelligent, get better at, you know, if
[96:15] you did give it an IQ test, it's getting
[96:17] higher scores than it was 10 years ago.
[96:18] We agree that there's been a upward
[96:20] motion of improvement.
[96:21] >> This is pretty much how machine learning
[96:23] goes when you feed it more data.
[96:24] >> Exactly. And you put more compute behind
[96:25] it. So if this continues,
[96:29] what does the future look like? So the
[96:32] rebuttal I was expecting to hear is that
[96:33] it won't continue. And actually,
[96:35] >> I actually don't think it I think that
[96:37] there are hard limits that we're going
[96:38] to hit. So you do believe in that
[96:40] there's a hard limit somewhere.
[96:41] >> We've kind of already hit the
[96:42] diminishing returns level because for
[96:45] example video generation which is by the
[96:48] way far less an American concern
[96:50] anymore. OpenAI shut down Sora. I think
[96:52] you can still use the API but
[96:54] nevertheless look at the look around you
[96:56] with the amount of stuff in the crew you
[96:57] need to get a shot. People think the
[96:59] movies are just shot by shot by shot and
[97:00] they just magically happen. When you've
[97:02] got my my wonderful girlfriend of first
[97:04] ads, assistant directors, you've got
[97:06] gaffers, you've got lighters, and also
[97:08] simulating light is insanely difficult.
[97:10] There are so many magical things that
[97:12] happen in creating visual images that
[97:14] yeah, you could create a one minute long
[97:16] thing that might fool someone. How do
[97:18] you practically turn that into a movie?
[97:20] Because that movie, I forget what the
[97:21] name is. There was a movie that claimed
[97:22] it aired at Can. It didn't. No one. It
[97:26] aired in the city of Can during the Can
[97:28] Film Festival. It was not at the film
[97:29] festival. When it comes to the practical
[97:31] creation of actual things at the end of
[97:33] it versus magic tricks, the actual
[97:35] practical outcomes are not there. The
[97:36] reason I keep coming back to the
[97:37] capabilities thing for the example is
[97:39] yeah, they can do better at tests, do
[97:41] better number go up. When it comes to
[97:44] can this actually do distinct tasks you
[97:46] can rely on it, you can rely on it for
[97:48] summaries. You can rely on it for
[97:49] generations. The things it was doing,
[97:51] it's getting linearlyish better at. But
[97:54] again, there's a ceiling to that. Like,
[97:56] okay, so it gets really good at
[97:58] research. What does that actually mean?
[97:59] you've already kind of got the
[98:00] automation there. What is the next step
[98:02] of that? Because training it to be more
[98:04] autonomous for example, that's not
[98:06] something that comes from training data.
[98:07] That is actually a new Gary Marcus a
[98:10] neuros symbolic. You actually need to
[98:11] build a structure around the AI to make
[98:13] it work. And even then, it doesn't fix
[98:15] the
[98:16] >> So you're saying that there will become
[98:17] a point where the rate of improvement
[98:20] will plateau.
[98:21] >> We're already there and stop.
[98:22] >> We've already hit that diminishing. Gary
[98:24] Marcus said this in 2022 as well. Do you
[98:25] know there's lots of people listening
[98:26] now that like they've had their
[98:28] workflows completely transformed by
[98:30] these tools? Have they?
[98:32] >> There'll be people. Yeah, there are.
[98:33] Yeah. The thing is, first of all, every
[98:35] single one of them, did you pay for the
[98:37] tokens? That's the thing. Did you pay
[98:39] for the tokens? And also, how many
[98:41] tokens did you burn? But putting all
[98:42] that aside, what workflows? Because if
[98:43] it's, yeah, I did a bunch of web
[98:45] scraping or web searches. I'm just not
[98:46] impressed. Did you make an entire
[98:48] [ __ ] movie? No, you didn't. Is it
[98:51] speeding up your coding? Yeah, I believe
[98:52] that. I've heard that from multiple
[98:54] people. But again, how much can you
[98:56] trust this?
[98:57] >> I think I'm I was getting at is, you
[98:59] know, when in the moment of any
[99:01] technological innovation, people they
[99:04] extrapolate linearly or they view it as
[99:08] a static state, i.e. they think today is
[99:10] going to look like tomorrow or they
[99:11] think it's going to get better in this
[99:12] sort of straight line. But what we end
[99:14] up seeing a lot of the time is this
[99:15] exponential improvement. All of the
[99:17] innovations we're talking about with you
[99:18] with like with compute and all that with
[99:20] fast processes, those are hardware
[99:22] breakthroughs. The hardware breakthrough
[99:24] companies don't seem to be fixing the
[99:26] LLM problems despite the all the king's
[99:28] horses, all the king's men with what
[99:30] nine 10 generations of TPUs from Google
[99:32] now. Broadcoms building stuff with open
[99:34] AI, their halapeno chip. And yet none of
[99:37] these people can just say, "Yeah, we're
[99:38] on the path to making this profitable."
[99:40] Because they can't. If we fix the
[99:42] environmental problems and the
[99:43] profitability situation, maybe I'd be
[99:45] more generous with this stuff. But they
[99:47] don't seem to be able to. And you talk
[99:50] about these improvements and
[99:52] capabilities. There's a certain point at
[99:54] which I'm saying, "Okay, can it do even
[99:57] a tenth of the stuff they're promising?"
[99:59] Sam the other [ __ ] week was saying it
[100:00] was going to be in like 6 months will be
[100:02] like a genie that you can ask wishes for
[100:04] from like [ __ ] never watched
[100:06] Aladdin. What's he talking about? Like
[100:08] also the the genie was charming. Anyway,
[100:10] long story short, the promises do not
[100:14] line up with the capabilities or the
[100:15] capability improvements. An exponential
[100:17] improvement
[100:19] in software and software performance is
[100:22] always a result of direct hardware
[100:24] improvement. We have all the gifted
[100:26] mathematicians, all the gifted software
[100:28] engineers, all the gifted hardware
[100:30] engineers. And where are we? Trillion
[100:32] plus dollars in with the future great
[100:35] financial crisis and the world's
[100:37] greatest marketing scop.
[100:38] >> I just think in the future I do think
[100:40] that all of the devices and the
[100:41] computers we use and the physical items
[100:43] in our world will be more intelligent. I
[100:45] mean sure but is that LLMs
[100:48] >> and that will be powered by the
[100:49] underlying AI infrastructure. It will be
[100:51] the more data data centers. It will be
[100:53] energy coming down.
[100:54] >> How does a GPU full data center
[100:58] translate to a Nikon camera that can I
[101:03] don't know even what you'd think think
[101:05] like because what is the thing we're
[101:06] talking about here? Because the idea
[101:08] that devices will get smarter. Sure, I
[101:11] can see that. It's a very broad
[101:12] statement. I could see it happening.
[101:13] It's really kind of happening. What does
[101:15] that have to do with the data centers?
[101:16] Cuz these data centers again are not
[101:18] being built to make your consumer
[101:20] electronics smarter. They're not being
[101:22] built for anything other than
[101:24] speculating on the ability to capture
[101:26] demand for generative AI services.
[101:27] >> But it's not just generative AI. We went
[101:29] through that earlier.
[101:29] >> Yes. No, but those data centers, they
[101:31] are being built for generative AI. They
[101:33] are not being built for anything else.
[101:34] Would you consider generative AI to be
[101:37] the fact that on Meta's earnings call
[101:38] like a couple of weeks ago, Mark
[101:40] Zuckerberg said, "The big breakthrough
[101:41] we've had, which has resulted in 15
[101:43] basis points of increased retention, I
[101:46] believe he was referring to Instagram,
[101:48] is that we now take anything you post on
[101:50] social media and we run it through an AI
[101:53] to get full context of what it is." And
[101:55] because we can see guy sat in front of
[101:57] me called Ed with blue shirt and coffee,
[102:01] we now can train the AI to serve whoever
[102:03] wants blue shirt, Ed, and with coffee to
[102:06] the right user, which means people are
[102:08] retained longer because
[102:09] >> it'sn't 15 basis points, like 0.15%.
[102:11] >> Yeah, it's cool. But it makes a
[102:12] difference at scale. It makes a big
[102:14] difference at scale.
[102:15] >> Yeah. But 10 and something billion
[102:17] dollars in and the best you've got is
[102:18] 0.15%. If if he could be fight I mean
[102:22] how much of a difference because
[102:24] >> there's a reason he's saying basis
[102:26] points versus dollars
[102:28] >> because think about it like this if Mark
[102:30] Zuckerberg was
[102:31] >> I take your point about scale. No, I'm
[102:33] saying the point I was making was that
[102:35] that is another application of these
[102:38] data centers because it needs a data
[102:39] center that is driving revenues, but
[102:43] also that's not out that's outside of us
[102:45] thinking about just generating
[102:47] >> and that's generative
[102:49] model. Muse was it? Oh, Muse Spark is
[102:51] their LLM. Gem is their generative ad
[102:54] model. Well, Muse then then that's them
[102:56] doing the weird thing where it's like on
[102:58] Instagram and it's like Dave the cat.
[102:59] Why is Dave the cat suffering? Like it's
[103:01] the weird popup things. Meta is [ __ ]
[103:04] god damn that company sucks. Like every
[103:06] time I think about how they've ruined
[103:07] that product. But that's the thing
[103:08] though, again, why can't he just say
[103:10] with his whole chest, we've made a
[103:11] couple billion. Why can't he say that?
[103:13] Because he isn't. Because there's not
[103:14] actually a way of going, I spent all
[103:16] this money. I spent 14 billion goddamn
[103:19] dollars on scale Alexander Wong and I
[103:22] made this much. They can't. It gets back
[103:24] to a very simple point of, hey, if it
[103:27] was going well, you'd tell me how well
[103:29] it was going rather than, I don't know,
[103:31] doing this weird rain dance thing where
[103:33] you're like, well, if we move all the
[103:35] pieces around in 3 years, theoretically,
[103:37] this will happen.
[103:39] I've done almost 700 interviews with
[103:42] some of the most interesting people in
[103:43] the world. And one of the things you
[103:44] learn, which is unexpected, is that
[103:46] vulnerability is the doorway to
[103:48] connection. And after sitting here for 2
[103:50] three hours with a guest, I feel a deep
[103:53] sense of connection to them. And as they
[103:55] leave, what I get them to do is to write
[103:57] a question in the diary of a CEO. We've
[104:01] taken all of the questions from the
[104:02] diary of a CEO. We have put the question
[104:06] here on this card with the name of the
[104:09] person that wrote it. So you can sit at
[104:10] home as I do with my fiance and my
[104:13] colleagues at work and other people in
[104:14] my life. Whenever we get a minute, we
[104:16] play the diio conversation cards and it
[104:20] is incredible what happens. These are
[104:22] great if you're in a romantic
[104:23] relationship and you want to connect
[104:25] your partner more. These are also great
[104:26] if you're in a team and you want to bond
[104:28] your team together. And I have to say
[104:30] they're also great for families that
[104:31] want to learn more about each other and
[104:33] that need a good excuse to spend some
[104:35] time in a digital world in the analog
[104:38] environment connecting human to human.
[104:40] It is remarkable what the right question
[104:43] at the right time can do. Go to the
[104:46] diary.com
[104:48] and you can get these conversation cards
[104:50] right now. There should be a button just
[104:53] down below here. And if it says
[104:54] subscribed, you're already subscribed.
[104:56] If it says subscriber, that means you're
[104:58] not yet. And if you're not subscribed,
[105:00] please could you do us a favor and hit
[105:01] that button? It helps the show more than
[105:02] you know. And according to the
[105:04] algorithm, you're someone that watches
[105:06] our show, but you haven't yet hit that
[105:07] button. Thank you so much. I do think
[105:09] you're accurate and right when you talk
[105:11] about the fact that there's a lot of
[105:13] like is the word for gazy?
[105:14] >> Yeah.
[105:14] >> Where like there's a lot of people that
[105:16] have spent a lot of money and they kind
[105:17] of shouldn't have spent it and they
[105:18] [ __ ] up and now they're thinking [ __ ]
[105:20] like we've spent all this invested money
[105:21] kind of like the metaverse was a bit of
[105:23] a [ __ ]
[105:23] >> oh my god that was a bit of a joke.
[105:24] >> That's so weird.
[105:25] >> A lot of money spent. We kind of thought
[105:27] this dream was coming of this well I
[105:28] shouldn't say dream cuz it's not a dream
[105:30] I've had but
[105:31] >> dream that they had.
[105:31] >> Yeah. This sort of virtual world and
[105:33] actually it never transpired and there's
[105:35] no sign that it will in the near term.
[105:37] AI and the dotcom boom in this regard
[105:40] are the same. NFTTS were the same,
[105:43] >> you know. So crypto, one could argue
[105:44] that a lot of the crypto industry was
[105:46] the same. It's weighing that is inflated
[105:48] by the media. The difference is the
[105:49] reason the metaverse and NFTs didn't
[105:52] escape this was there weren't stocks to
[105:54] speculate on. There weren't big
[105:55] companies that you could invest in. They
[105:57] had re record earnings in 2021. There's
[106:00] a bunch of money floating in the system
[106:01] thanks to postcoid uh the PDC that
[106:04] basically government federal money
[106:06] flowed in to the banks. There was a
[106:07] bunch of easy money zero interest free
[106:09] era money was easy to find. Then after
[106:11] that there was the hangover. Growth
[106:12] started to slow down dramatically. This
[106:14] is actually my rockcom bubble theory
[106:16] which is they don't have any hyperrowth
[106:18] ideas anymore. So suddenly they started
[106:21] buying GPUs. And when they bought GPUs
[106:23] people went they're doing AI. Oh we
[106:26] better buy the stock. And the stocks
[106:27] went on an incredible run. may like
[106:28] several hundred percent grow in the last
[106:30] few years. the stock has grown by
[106:32] hundreds of percent. Despite zero proof
[106:35] and because the media was just saying,
[106:37] "Yeah, Meta's revenues growing because
[106:40] of AI, right? Microsoft's revenue is
[106:41] grown because of AI, right? The fugazi
[106:43] you're talking about was the fact that
[106:45] everyone just gave them credit in
[106:46] advance and now we're kind of getting to
[106:48] the point where it's like, hey, you
[106:50] didn't spend that trillion dollars for
[106:51] no reason, did you? Satcha Amy Amy Hood
[106:54] just going to take him out back, send
[106:56] him to the glue factory or something?"
[106:57] Like,
[106:57] >> I do think there's overspending. I I
[106:59] want to concede that but I doic
[107:01] >> yeah no I do think there is and I think
[107:03] the reason why there's overspending Ed
[107:06] is I think there is something here
[107:08] >> and what
[107:10] >> in terms of like I think there is pra p
[107:12] p p p p p p p p p p p p p p p p p p p
[107:12] practical uses for this technology and I
[107:14] think when people realize that through
[107:16] history they go crazy because they want
[107:18] to be the person that owns the
[107:19] opportunity.
[107:20] >> I'm going to be honest I just I
[107:21] fundamentally don't agree.
[107:22] >> You don't agree with which part you
[107:24] >> I don't agree that this that the
[107:25] speculation is a result of actual
[107:27] demand. I don't believe it's suspect. I
[107:29] don't think private credit is sinking
[107:30] hundreds of billions of dollars into AI
[107:32] because of actual demand. They are doing
[107:33] it because they saw the biggest
[107:34] companies in the world building data
[107:36] centers making a ton of money from two
[107:37] companies they feed money and went I
[107:39] want some of that money.
[107:40] >> I am saying that I do think there is
[107:42] value in the underlying technology. I
[107:44] think that and so I think I'm not saying
[107:47] how much value
[107:47] >> right okay I actually I get your meaning
[107:50] that's fair.
[107:50] >> I'm not saying it's proportionate to the
[107:52] investment. All I'm saying is that do
[107:54] you know what it's like? It's like if I
[107:56] take your example, the rot economy essay
[107:57] that you wrote.
[107:58] >> Yeah.
[107:58] >> Say that you're on a desert island and
[108:00] then someone says they found a banana
[108:02] tree,
[108:02] >> right?
[108:03] >> And there's there's 10,000 people on the
[108:06] island.
[108:06] >> Okay.
[108:07] >> They are going to stam [ __ ] peed
[108:10] towards where they think the banana tree
[108:11] is. They are going to [ __ ] claw each
[108:14] other to pieces. And if if your essay
[108:16] here is right that there was desperation
[108:17] cuz they hadn't found an innovation in a
[108:19] while,
[108:20] >> maybe that explains it. Maybe there is a
[108:21] bit of value here,
[108:22] >> right?
[108:23] >> And they're stam [ __ ] peeding and
[108:25] killing each other and making irrational
[108:26] decisions like hungry people would.
[108:28] >> I actually think we're then we actually
[108:30] agree. That is actually my point, which
[108:32] is these three companies in Meta, their
[108:34] main business lines are running out of
[108:36] growth. There's only so much they can
[108:37] grow. And indeed, in the next three and
[108:38] a half years, analysts think that these
[108:40] two bastards, these two, OpenAI and
[108:42] Anthropic are going to spend over $400
[108:43] billion on these people alone,
[108:46] Microsoft, Google, and Amazon. And the
[108:48] crazy thing is is that's a large part of
[108:49] their future growth. And if this money
[108:51] isn't spent, their growth slows down.
[108:53] Okay,
[108:53] >> so your point about a bananas, I
[108:55] actually agree. That is the rockcom
[108:56] bubble, it's they don't have a new thing
[108:58] and they're desperate. And indeed, they
[109:00] got rewarded for buying the GPUs. They
[109:02] got when they bought these goddamn GPUs
[109:04] from Nvidia, all the markets went
[109:07] rockard overnight. They loved it. There
[109:09] were stories about how they were sending
[109:10] armored cars with the GPUs to Microsoft
[109:13] to make sure Microsoft got the GPUs. And
[109:15] so everyone saw all that money flowing
[109:16] in. Even though they never disclosed AI
[109:18] revenues, they saw the expenditures and
[109:20] they went, "Well, I want to do what
[109:22] these people are doing. I want to get a
[109:23] little of that money, don't I?"
[109:25] >> I think the area where we have a slight
[109:27] disagreement is that I think the
[109:29] underlying technology has a lot more
[109:31] promise over the long term than you do.
[109:34] So the thing I want to push back on
[109:36] there is
[109:38] to have progress with AI just on a
[109:41] taking it in a vacuum to have progress
[109:42] for these two companies to keep going
[109:44] and to keep progressing they need to
[109:47] spend tens of billions of dollars a year
[109:49] on training.
[109:50] >> The only way that that can happen is if
[109:53] these companies and venture capitalists
[109:54] and private credit firms and Nvidia
[109:56] >> keep circulating money to them. So the
[109:58] progress
[109:59] >> that we've got so far is entirely a
[110:01] result of this circular system. So it
[110:04] means that
[110:05] >> circular you talked about VCs there
[110:06] >> venture capitalists who are by the way
[110:09] the majority of the funding that open
[110:11] AAI got in the last 6 months came from
[110:14] SoftBank Nvidia and Amazon
[110:16] >> okay yeah
[110:16] >> so just the point is is you're talking
[110:18] about progress continuing progress in
[110:21] LLM can only continue as long as the
[110:23] money keeps flowing once the money keep
[110:26] once the money stops flowing the
[110:28] progress stops which
[110:28] >> but isn't that most like early like
[110:30] Spotify didn't make money for 20 years
[110:31] >> Spotify didn't lose 20.9 9 billion in
[110:34] one year. They didn't need to raise $217
[110:36] billion in the space of 6 months.
[110:38] >> Yeah. And Uber is another example.
[110:40] >> $33 billion since inception before it
[110:42] became a messy kind of profitable.
[110:43] Amazon Web Services between 2003 and
[110:45] 2015 when it became profitable. $29.7
[110:48] billion the scale. Yeah. That's the
[110:50] total capital expenditures and that's
[110:52] not just Amazon Web Services. That's the
[110:53] entire logistics operation normalized
[110:55] for inflation.
[110:56] >> So they all lost money for a long period
[110:58] of time is the TLDDR.
[110:59] >> Yes. But the amount of money they lost
[111:01] is
[111:04] completely
[111:05] just magnitudes different on a level
[111:08] where these three
[111:09] >> Can I argue then that the that's because
[111:11] the potential of intelligence permeates
[111:14] everything whereas Amazon at the time
[111:16] was like selling books
[111:17] >> no
[111:17] >> that was that was bringing retail online
[111:19] >> when Amazon web services grew it was
[111:21] >> oh so cloud with Amazon web services the
[111:25] reason I bring that up going to repeat
[111:26] something but it's really important 2003
[111:28] it was founded
[111:29] >> and it was founded mostly because Amazon
[111:31] as a growing online store needed
[111:33] hardcore infrastructure. 2006, I think,
[111:36] is when they turned it client-f facing.
[111:38] I may be wrong on the dates there, but
[111:39] 2015 was the year it became profitable.
[111:41] >> Yeah.
[111:41] >> The total capital expenditures
[111:43] normalized for inflation with $29.7
[111:45] billion across that 12-year period.
[111:48] >> Yeah.
[111:48] >> And yeah, it lost money, but
[111:51] >> if we speak cold economics here, Amazon
[111:55] didn't have to go into the they were
[111:56] unprofitable in in a way, but their
[111:58] margins actually started improving
[111:59] because AWS was a very margin heavy
[112:01] business. It was great.
[112:02] >> Yeah,
[112:03] >> these these two Google cash flow
[112:06] negative, Amazon cash flow negative.
[112:08] These businesses, the reason you liked
[112:09] software businesses was they are meant
[112:11] to be cash heavy asset light. These
[112:16] companies along with Meta have added
[112:18] more than $700 billion of new property,
[112:21] plants and equipment. So assets, data
[112:23] centers, GPUs in the last four years.
[112:26] They have gone from being these cash
[112:28] machines to these cash furnaces.
[112:31] >> You said a second ago, this can only
[112:33] continue if if investors continue to
[112:35] invest.
[112:36] >> Yes.
[112:36] >> And I was saying I I think that
[112:38] investors are used to pumping money into
[112:40] things that are burning cash. Your
[112:42] rebuttal to me sounds like well this is
[112:44] burning more cash than ever. And then so
[112:46] I would say well is the opportunity
[112:48] bigger than those other case studies you
[112:51] referenced like AWS? And one would say
[112:54] that the opportunity of intelligence
[112:58] permeates everything. So the TAM the
[113:00] total addressable market is enormous.
[113:03] Maybe the revival back to me is about
[113:05] open source and all these kind of
[113:06] >> No, no, no. I I actually know what
[113:07] you're getting at. So what you were
[113:08] describing there is the argument that
[113:10] Sachinadella or Sam would make that the
[113:12] theoretical opportunity of large
[113:14] language models and I could have bought
[113:16] that [ __ ] into any 24 from them when
[113:18] they were like, "Oh, we see the
[113:20] opportunity. We've gone way past the
[113:22] point at which you can rationally argue
[113:24] that LLMs need this much money. And when
[113:27] I say the money needs to keep flowing, I
[113:28] am talking these two compan Open AI just
[113:32] open AI Clammy Sam has said Wall Street
[113:35] Journal and Isaagi reported a few weeks
[113:37] ago they plan to spend $750 billion on
[113:42] compute through 2030. I think they're
[113:44] going to be dead before then, but $750
[113:46] billion.
[113:48] That is an insane amount of money. That
[113:50] is crazy
[113:50] >> and a large chunk of that is training.
[113:53] So when I say progress, I mean literally
[113:55] to make the models better at stuff
[113:57] requires billions of dollars invested
[113:58] just in data
[114:00] and also tens of billions of dollars of
[114:02] taking that data. And so training
[114:04] training is actually a really
[114:05] interesting thing because when you think
[114:07] of like for Jake and Troy my trainers
[114:10] when I train with them when I lift with
[114:11] them I have a defined thing and when I
[114:13] do it and I eat right muscles get bigger
[114:15] they would. And here's the thing. When
[114:17] you train with an LLM, you're
[114:18] experimenting each and this is not
[114:20] actually a hit on the companies because
[114:22] they're still trying to work out how to
[114:24] do the thing because putting aside how I
[114:26] feel like they're trying to innovate. I
[114:28] think there are people at these
[114:29] companies that actually want to do
[114:30] something interesting. It's costing too
[114:31] much money. So once the money tap turns
[114:34] off, the money won't be there to buy the
[114:36] data or feed the data into the GPUs. Put
[114:38] aside all the thoughts I have, just the
[114:40] raw capital to get them this far has
[114:43] cost increasingly larger amounts of
[114:45] money and increasingly larger amounts of
[114:47] training money for training runs that
[114:49] sometimes can fail. GPT5 was meant to be
[114:52] this panacea for the AI industry. They
[114:55] had at least one training run that cost
[114:56] half a billion dollars and did nothing.
[114:58] And that's the thing. If we are thinking
[115:01] about progress in a in a vacuum, they
[115:03] need so much more money just to maybe
[115:05] get somewhere. There's no guarantee.
[115:07] There's never any guarantee, but there's
[115:08] a reason that Google and Amazon are cash
[115:10] flow negative now. There's a reason why
[115:11] Oracle's probably going to die as a
[115:13] result of OpenAI because Oracle's future
[115:16] depends on OpenAI spending $300 billion
[115:18] over 5 years.
[115:19] >> It's absolutely fascinating because I
[115:21] was just reading through a list of
[115:22] quotes from the big CEOs of AI companies
[115:24] to see what they would rebuttle you.
[115:26] >> Yeah.
[115:27] >> And they're all basically saying the
[115:29] same thing. They're all saying, this is
[115:31] actual an exact quote from Sundar who is
[115:33] the CEO of Google. He says the risk of
[115:36] underinvesting is dramatically greater
[115:39] than the risk of overinvesting.
[115:42] And you go down, you go through this,
[115:44] you know, Andy Jasse, CEO of Amazon,
[115:46] we're not investing approximately 200
[115:48] billion in capex in 2026 on a hunch.
[115:51] We're not going to be conservative in
[115:53] how we play this. We're investing to be
[115:55] the meaningful leader and our future
[115:58] business operating income and free cash
[115:59] flow will be much larger because of this
[116:02] investment. Then Mark Zuckerberg, CE of
[116:04] Meta, says we'll continue to invest
[116:06] aggressively in infrastructure to meet
[116:08] the demand. I'd rather risk building
[116:10] capacity before it's needed than being
[116:12] late. Makes me think of Shrek with L
[116:15] Farquad. Some of you may die, but that's
[116:17] a risk I'm willing to accept. It's like,
[116:19] you know, I'm just going to spend all
[116:20] this money. You can't fire me cuz Mark
[116:22] Zuckerberg can't be fired due to the
[116:23] unique board situation he's got going.
[116:26] So yeah, he's just going to piss the
[116:27] money away and hope he's right. And I
[116:28] know from the people who know it matter,
[116:29] he's not right. The thing is, why might
[116:32] you be wrong?
[116:33] >> I mean, this is the thing. The AI people
[116:35] who claim this is going to be the
[116:36] biggest, strongest thing in the world,
[116:37] did they ever get that? I I mean this
[116:39] like
[116:39] >> it's a good question because it's like
[116:40] they don't. And the thing is, what would
[116:42] it take for me to be wrong? A bunch of
[116:44] hardware breakthroughs to make this
[116:45] profitable. A bunch of
[116:46] >> question new mathemat because the thing
[116:48] is
[116:48] >> when it comes to being a critic or a
[116:50] skeptic,
[116:51] >> you are put on the hot seat. Not the
[116:53] people spending a trillion dollars, not
[116:55] the people promising the world. The
[116:56] person the the [ __ ] with a blog is
[116:58] the one who's like me. Trust me. If they
[117:00] came here, they'd be on the hot seat,
[117:01] too. Trust me.
[117:02] >> Oh, I Oh, they they won't talk to me.
[117:05] Don't know why, Steve. They don't know.
[117:07] It's cuz I call him Clammy Sammy. Um
[117:09] >> I think it's cuz my guests are quite
[117:10] quite critical that I don't think Solman
[117:12] wants to come here.
[117:13] >> Mr. Orman, go on Steve show. Do it. But
[117:15] this is the thing like of course they're
[117:17] going to say that. And also, if they
[117:19] thought they were right, I don't think
[117:20] they do anymore. If I was in their shoes
[117:22] and I thought that this was an
[117:23] existential thing, sure. But it gets
[117:25] back to the rocom bubble which is yeah
[117:27] this is the last thing they've got.
[117:28] >> But I really want to know that question.
[117:29] It was one of the questions I was really
[117:30] excited to ask you which is you have a
[117:32] different opinion. We said this at the
[117:34] top. You have a very different opinion
[117:36] from a lot of people. I would categorize
[117:38] the the two most popular opinions as
[117:40] >> uh AI is going to hurt everybody and
[117:42] it's going to be catastrophic and we
[117:44] need to stop.
[117:44] >> Yeah.
[117:44] >> The other opinion is age of abundance is
[117:46] going to be amazing. Let us crack on.
[117:48] yours is different from both of those
[117:50] which is as you said in your words it's
[117:53] a con and it's and there's no real
[117:55] underlying value in the technology and
[117:57] it's overhyped.
[117:58] >> Yes.
[117:58] >> And there's way too much spending. I
[118:00] mean a few people agree on the spending
[118:01] part but the other part. So with you
[118:03] it's one of probably the first person
[118:05] that I've spoken to that's had this
[118:06] opinion.
[118:08] >> So how what would it take for you to
[118:10] change your mind about what you believe
[118:14] here? There would need to be a hardware
[118:16] breakthrough that reduced the cost by
[118:18] like a thousand but it would have to be
[118:19] just a dramatic breakthrough that is not
[118:22] happening just to be clear because
[118:23] they've all been trying. So it's the
[118:25] cost for you that would have to change.
[118:26] >> It's the cost and it's also the data
[118:28] centers. I think the way they're
[118:29] building the data centers is reckless
[118:30] and damaging to communities. The fact
[118:32] that you have communities like in
[118:34] violent New Jersey where the residents
[118:35] like I don't want this but the planning
[118:37] boards vote for it because they're all I
[118:39] assume having chummy lunches with the
[118:41] people doing it. I think the use of gas
[118:43] turbines is [ __ ] disgraceful. I the
[118:45] water situation I'm not super well read
[118:47] on, so I'm not going to wait into it,
[118:48] but the use of gas turbines and behind
[118:50] the meter power is reckless and damaging
[118:52] to communities. The noise that these
[118:54] things make and also generative AI is
[118:57] this egregious pornographic
[119:00] demonstration of how unfair the world
[119:02] is. Regular people try and get a loan
[119:04] for a business, a random business. They
[119:06] want I have a good idea. They go to a
[119:08] bank, a bank of town, go [ __ ]
[119:09] themselves. They'll say, "I'm not g you
[119:11] going to make a store that sells stuff.
[119:12] Screw you. You want to build a data
[119:14] center? You Jensen Hang will back you.
[119:17] Jensen Hong will give you 25% residual
[119:19] value. You want to build a regular
[119:21] business that's even profitable? [ __ ]
[119:23] you. No, a venture capitalist won't give
[119:25] you the money. Something that's just
[119:26] growing steadily, but it's profitable.
[119:28] Screw that. No, I need 10 100x return.
[119:31] Try and get a mortgage. You have to give
[119:33] the bank a full colonic. But you want to
[119:35] get money for Jensen Hong to buy some
[119:37] GPUs? He'll give you a contract.
[119:39] Corewave is a great example. C Neocloud,
[119:41] which is just a company that builds data
[119:43] centers and puts GPUs and rent them to
[119:45] people. Nvidia, one of their first
[119:47] investors in 2023, signed a $1.3 billion
[119:51] contract to rent back their GPUs from
[119:54] Core. So that Core go to a bank and go,
[119:56] I got a customer. Yeah, it's the guy I'm
[119:59] buying the GPUs from with the debt I'm
[120:01] getting from you. If you want to buy
[120:03] GPUs, it's open season. If you want to
[120:04] live a regular life where you build a
[120:06] regular business or buy a house, highest
[120:08] interest rates ever. Screw you. Up
[120:11] yours. Yeah, you need to show us way
[120:13] more than that. I don't trust you
[120:14] regular folks. But if you're an
[120:16] unprofitable Neocloud, you get billions
[120:19] from Jensen. It doesn't matter.
[120:21] >> It's so interesting. You It's
[120:22] interesting because you are the first
[120:24] person that I've spoken to that has that
[120:25] opinion.
[120:26] >> I am prouser. Let's take another myth.
[120:29] AI will be conscious. Mhm. So
[120:34] super intelligence, artificial general
[120:36] intelligence, these are theories. Anyone
[120:39] saying this stuff will become this is
[120:42] just guessing and does not have proof.
[120:44] >> Okay.
[120:45] >> And like that's really it.
[120:46] >> Okay.
[120:47] >> Okay. Let's take another myth.
[120:50] AI systems are already blackmailing and
[120:52] escaping control. So this is a really
[120:54] specific one. Anthropic. There's
[120:56] actually two. Open AAI's GPT 3.5. I
[121:00] realize this is more than the sentence.
[121:01] I apologize.
[121:03] In their system card, and a bunch of
[121:05] media outlets covered this, saying that
[121:07] OpenAI's model blackmailed a task rabbit
[121:10] into solving a capture. What actually
[121:12] happened was a user of GPT doing the
[121:17] experiment
[121:19] got it to generate things to say to a
[121:21] task rabbit to make a task rabbit do
[121:23] stuff.
[121:24] >> A task rabbit
[121:24] >> as in a person that you rent, not even
[121:26] to do a capture. It's something you rent
[121:28] to like nail a picture up in your
[121:30] apartment. It's an insane example. This
[121:32] was covered as if these things
[121:33] blackmailed someone and and it and they
[121:36] specifically said, "Yeah, we prompted it
[121:38] to do this." And also the other note was
[121:40] that yeah, AI systems can't do
[121:42] autonomous stuff like this. Then there
[121:43] was this other one where Anthropic said,
[121:45] "Oh yeah, a model was blackmailing
[121:47] someone saying that if you don't do
[121:49] this, I'll email proof that you slept
[121:51] with someone else other than your wife."
[121:53] I think it was what actually happened
[121:54] was Anthropic explicitly trained a model
[121:57] to do this and then prompted it to
[121:59] blackmail.
[122:00] This keeps happening and the media just
[122:03] slop slot me up. I don't need no
[122:05] thoughts. Put the story in the bag. And
[122:08] it's frustrating because it scares
[122:10] people. Put aside the fact it's wrong.
[122:12] It's scary. It's scary to people. people
[122:14] living their lives who have to work
[122:16] longer hours to make less money and
[122:18] their money doesn't go far and they turn
[122:20] on the [ __ ] news and there's some
[122:21] [ __ ] being like, "Yeah, you should be
[122:23] terrified it blackmailed someone."
[122:25] >> But this is this is so counterintuitive
[122:27] of their interest to some degree and
[122:30] they've experienced it backfire.
[122:31] >> Well, they have now like it's it's
[122:33] literally backfired.
[122:34] >> It's backfired. Eric Schmidt getting
[122:35] booed at a commencement speech by 8,000
[122:38] people every time he said the word AI.
[122:40] But I mean this is this is I mean these
[122:42] serious are being attacked at home.
[122:44] >> Yeah. Which [ __ ] sucks. Which is
[122:46] >> terrible. I must be clear like you
[122:48] dislike the don't [ __ ] hurt people.
[122:49] >> Yeah. Don't don't attack people at home.
[122:51] But but the point here is that that
[122:53] narrative is backfiring in a big big way
[122:56] for them. I don't think they saw it
[122:58] coming because you have to remember you
[122:59] mentioned regulation earlier. These tech
[123:01] companies have been glazed for their
[123:03] entire existence. Travis Kick's like oh
[123:06] what? People don't like me now. And it's
[123:07] because Uber was a horribly run place
[123:09] and he was kind of a monster. Also tons
[123:12] of articles about how great Uber was at
[123:13] the time. The point I'm making is these
[123:14] companies are not used to push back.
[123:16] They thought what would happen I believe
[123:18] just guessing. They thought they do this
[123:20] scary stuff and they would just get
[123:21] floods of money and everyone would just
[123:23] be like I kneel before you. I'll do
[123:25] whatever you want. They didn't expect I
[123:28] think what has I I agree this has
[123:30] backfired on them because they were in
[123:32] articulate. They're disconnected from
[123:34] regular people. Samman drives a $5
[123:36] million car around San Francisco. So
[123:39] that that man's doing it like 9 miles an
[123:41] hour. It's hilarious. But these people
[123:43] are disconnected from everyone else. So
[123:44] they don't they don't experience real
[123:46] problems, so they can't build the
[123:47] solutions for them. And they think,
[123:48] well, if we scare people into doing what
[123:50] we want, that'll work, right? It didn't.
[123:52] They was all of this blackmail stuff was
[123:55] an attempt to make it mystic. It was a
[123:57] mysticism attempt. It was to make it
[123:58] seem like this unknowable, impossible to
[124:00] control, just this powerful thing. But
[124:02] we're the only ones. We are the o only
[124:05] us only these two angels could possibly
[124:08] control the beast we've created.
[124:10] >> This is this is quite a controversial
[124:12] statement but I think that for some
[124:14] reason I trust Dario a little bit more
[124:17] because I think he's been the most
[124:19] balanced in his writing about the risk
[124:21] profile.
[124:22] >> I
[124:22] >> whereas the others they they seem to
[124:25] kind of move with the wind.
[124:27] >> I I do you know
[124:28] >> I get what you mean. The reason I don't
[124:30] like Dario is Daario was doing the scare
[124:32] tactics thing when he worked at OpenAI
[124:34] when GPT2 came out say it's too scary to
[124:37] release. He's also gone on television
[124:39] and given AI psychosis to Axios being
[124:42] like 50% of jobs are going to go away
[124:44] because of AI.
[124:45] >> What I respect is the consistency. He's
[124:49] now being attacked by them.
[124:50] >> Good.
[124:51] >> Um but the thing is sorry I mean let me
[124:53] clarify the word attack. Darian is being
[124:56] verbally attacked by Silicon Valley and
[124:59] you know if Silicon Valley if powerful
[125:01] people in Silicon Valley are attacking
[125:03] someone.
[125:04] >> Four months ago he wasn't though. They
[125:05] were all saying he was the smartest boy
[125:07] ever.
[125:07] >> The point I want to make there as well
[125:08] is again wow you're so scared of how
[125:10] powerful this is. You're so scared of
[125:11] it. It's so scary. What are you doing
[125:13] about it? Oh nothing. Like it's just
[125:15] like what are you doing? Well we have an
[125:16] alignment team. So does every AI lab.
[125:18] Well I guess open AI cycles through
[125:20] those really quickly. Here's the thing.
[125:22] If I'm Dario Amade, I'm sitting there
[125:23] going, I'm scared of all things changing
[125:26] and I thought I had made a thing that
[125:28] would eliminate all jobs, I'd be [ __ ]
[125:30] terrified. I'd be walking around with
[125:31] like like a 10 ton weight on my back.
[125:34] The show, the responsibility, the fact
[125:36] he doesn't, the fact he wants to be this
[125:38] weird elder statesman that's too scared
[125:40] to hold Sam Orman's hand at an event
[125:42] just makes me believe that he's just
[125:44] saying it because it's convenient and
[125:45] he'll wind that back as he kind of
[125:47] already has whenever it's convenient for
[125:49] him. I think Open AAI and Anthropic are
[125:51] basically the same level of Bad Company.
[125:53] I think Anthropic is more cultlike. I
[125:56] think it's so weird like Jack Clark over
[125:58] there, one of the co-founders. That fell
[126:00] used to be at the register. He used to
[126:01] be one of the most critical journalists
[126:02] ever. Now he's it's like like something
[126:04] took over him because they talk of these
[126:06] things in these high fluent terms. But
[126:08] then again, maybe the people at
[126:09] anthropic buy their [ __ ] Maybe some of
[126:10] the people at OpenAI buy their [ __ ] I
[126:12] don't know. So going back to the central
[126:13] question we asked at the top here was
[126:15] what would have to be the case for you
[126:16] to look back and say do you know what I
[126:17] was wrong in 2026 and you said to me it
[126:20] would be mainly that the cost of
[126:23] production around AI drops dramatically
[126:26] >> and it would have to also do insane
[126:29] amounts of stuff it does it would have
[126:30] to be a truly autonomous
[126:32] >> it would have to continue its
[126:33] improvement in terms of capability.
[126:34] >> It would have to be a different product.
[126:36] It would have to be it would have to be
[126:37] indistinguishable from magic. And the
[126:38] reason they have these high standards is
[126:40] they set them.
[126:41] >> Okay. Fair. It's interesting as well
[126:42] because all these myths and all these
[126:44] conversations, it's about technology,
[126:46] but it's also it's an information war.
[126:48] It's literally
[126:50] narrative versus narrative. Everyone
[126:52] trying to escape the financials,
[126:54] everyone trying to actually escape what
[126:56] the models can do. And the big thing I
[126:58] always say about AI boosters is if I
[127:00] could regulate them, I'd regulate them.
[127:02] They can't speak in the future tense
[127:03] anymore. Just you got to talk about
[127:04] today, mate. You get two weeks in the
[127:06] future, Max. Because if they were
[127:08] constrained to what was happening today,
[127:10] it they would sound like insane people.
[127:12] >> Yeah. No, I think yeah, most I guess
[127:14] most technology companies would at the
[127:15] time. Like Uber would sound insane.
[127:18] Amazon was
[127:18] >> Uber was basically the difference.
[127:20] >> They were pissing money though, weren't
[127:21] they?
[127:21] >> They were pissing money away, but the
[127:22] unit economics were the same just
[127:24] subsidized. So you were still getting a
[127:26] service from A to B and paying a much
[127:29] lower cost. It wasn't like you paid Uber
[127:32] 200 sorry 20 bucks a month and you could
[127:34] get 500 miles of Uber and then one day
[127:36] you started paying by the mile cuz
[127:38] that's what's happening with this.
[127:39] >> Have they they've changed their business
[127:41] model for customers like me now so that
[127:43] I have to buy credits.
[127:45] >> No. So you well kind of with
[127:47] >> they asked me the other day. So with the
[127:49] anthropics fable model with some
[127:51] accounts you have to pay for usage and
[127:53] also adoption of fable has been pretty
[127:55] low because of this because of the cost
[127:57] but with enterprises so companies over
[127:59] 150 people you have to pay by the token
[128:02] now or per million token.
[128:03] >> Oh so they are moving to a token.
[128:05] >> Yeah. But when they did that everyone
[128:06] went from being like this is the most
[128:07] impressive thing ever to being like
[128:10] >> it's always we got to control these
[128:12] costs. Uber's COO said as Andrew
[128:14] McDonald I think he said that it's
[128:16] getting hard to justify cuz it's hard to
[128:18] connect spending money on tokens to
[128:20] actual useful outcomes.
[128:22] >> He said the thing like he said the
[128:24] actual thing I've been saying and it's
[128:25] so we're in an AI bubble.
[128:27] >> Yes.
[128:27] >> And when will when this AI bubble
[128:29] collapses so much of the economy is
[128:31] resting upon it.
[128:33] >> Yeah.
[128:34] >> It's going to have downstream
[128:35] consequences. So I got two questions for
[128:36] you. I guess the first question is are
[128:38] we in an AI bubble and what happens when
[128:40] the bubble pops?
[128:41] >> Yes. And it's it depends. So the big
[128:45] thing that people say is, "Oh, we'll get
[128:47] bailed out. Donald Trump scared of
[128:48] Donald Trump." Here's the problem with
[128:50] this.
[128:52] It isn't just an AI bubble. It's the
[128:54] rockcom bubble. So the AI bubble
[128:56] collapsing will probably be this company
[128:58] running out of money. Open AI.
[129:01] >> And the thing is with Open AI is they
[129:03] were meant to go public this year and
[129:04] now it's been pushed to next year a week
[129:06] and a half after I released their
[129:07] auditive financials. Wonder where that
[129:09] was. Um, but they've delayed to next
[129:11] year. Sarah Frier, the CFO, has now
[129:12] said, "Well, they'll do it earlier than
[129:15] 2027 or 2027." Great answer there.
[129:18] >> For anyone that doesn't understand what
[129:19] going public means, that means joining
[129:21] the stock market. And at such a time
[129:22] when you join the stock market, your
[129:24] investors can finally sell their equity
[129:27] that they got for investing in the
[129:29] company when it was private. So often
[129:31] times companies will flirt with the idea
[129:34] of we'll go public someday soon because
[129:36] investors will have a moment in their
[129:38] head where they'll get their money back
[129:40] at a return. So you kind of need to if
[129:43] you're in these guys shoes, you kind of
[129:44] need to be flirting with going public or
[129:45] investors won't want to invest.
[129:47] >> Open AAI up until this point has been a
[129:49] private company and their last funding
[129:51] round they were valued at $865 billion.
[129:54] Now when they tried to go public, New
[129:57] York Times Mike Isaac reported this.
[129:59] They tried to list well they wanted to
[130:02] go at a set a 1 trillion valuation.
[130:05] Apparently their advisor said no don't
[130:08] do that. That is very bad for a number
[130:10] of reasons. One open AI needs perpetual
[130:12] amounts of money. They raised $122
[130:14] billion this year. Most of it's crossed.
[130:16] There's some left but they are going to
[130:18] need to raise at least hundred billion a
[130:20] year just to survive. If they can't go
[130:22] public they will have to raise another
[130:24] funding round. The problem is it's going
[130:26] to be difficult to raise at even the
[130:28] same one they raise that. They're
[130:29] probably going to have to take a flat.
[130:30] So the same amount. Exactly. But they
[130:34] need money. They need money so bad.
[130:35] Amazon sent them $35 billion that was
[130:38] meant to be contingent on them going
[130:39] public early.
[130:41] >> They did that because they need the
[130:43] money. Now, OpenAI is the kind of
[130:46] catastrophe center here because
[130:47] Anthropic is likely going to beat it to
[130:49] go public. And once Anthropic goes
[130:50] public, it'll be borderline impossible
[130:52] for Open AI to do so because Anthropic,
[130:54] an unprofitable, unsustainable AI lab,
[130:56] but a better business that's growing
[130:58] faster than Open AI's. I believe they
[131:00] have a ceiling. They're eventually going
[131:01] to face predition, too. I think sometime
[131:04] in 2027, things are going to start
[131:05] running out of steam. Because the thing
[131:07] I said earlier, the only way these
[131:08] models get better is if you feed more
[131:10] money, tens of billions of dollars into
[131:12] them.
[131:12] >> So, you think OpenAI runs out of steam
[131:14] in 2027?
[131:15] >> I think they're already running out of
[131:16] steam. Yeah. But I think they run out of
[131:17] cash. You think they run out of cash?
[131:19] Yes. And the sequence of events here
[131:21] will be they they go out and try and
[131:22] raise
[131:23] >> and they have trouble raising another
[131:25] round. I think maybe Invidia props them
[131:27] up a little. Maybe Private Credit,
[131:29] Blackstone, Black Rockck and the like
[131:30] the ones and the reason that Private
[131:32] Credit is getting involved. So asset
[131:33] managers is because they're investing in
[131:35] the data centers and they know this
[131:36] company's most of the data center
[131:38] demand.
[131:38] >> Okay. So they run out of steam in 2027
[131:40] according to you.
[131:41] >> Yep. And maybe they try if they bum rush
[131:42] to go public they're going to have worse
[131:44] economics than anthropic. They're going
[131:45] to get savage. it. We work was a great
[131:47] example. Another SoftBank classic. Now,
[131:50] I think Open AI collapses, there are
[131:52] many different ways it could happen.
[131:54] There are many different ways it could
[131:55] end. But the crucial thing is is that
[131:57] there are multiple companies that are
[131:59] existentially tied to OpenAI. SoftBank,
[132:03] one of the largest companies in the
[132:04] Japanese stock market, a holding company
[132:06] with lots of investments. They have on
[132:08] paper about hundred billion worth of
[132:10] OpenAI stock. If they can't go public,
[132:13] they can't do diddly squat with that.
[132:15] And so Soft Bank's future, their ability
[132:17] to continue paying the people around
[132:19] them and existing as a business relies
[132:21] on their ability to continually
[132:23] liquidate funds to be to take the things
[132:25] they've invested in and have value from
[132:27] them either by selling the stock or
[132:29] taking loans out on the stock. If OpenAI
[132:31] can't go public, SoftBank can't do that.
[132:33] SoftBank probably won't run out of
[132:35] money, but we're going to see one of the
[132:36] largest holding companies in the world
[132:38] become much smaller. We will also see
[132:41] Amazon, Google, and Microsoft have to
[132:43] restate guidance. they will have to say
[132:45] actually we don't think we're going to
[132:47] grow as fast
[132:48] >> and what happens then
[132:49] >> well I think we enter a tech depression
[132:51] because the rockcom bubble the core of
[132:53] my theory is that they're out of
[132:56] hyperrowth ideas but the market doesn't
[132:57] think so the reason they're so
[133:00] maniacally spending is because buying AI
[133:03] GPUs allows them to kick the can further
[133:05] allows them to say we're still doing
[133:07] something we're working on AI don't
[133:08] think too hard and also their current
[133:10] businesses are still growing their
[133:12] current businesses will eventually slow
[133:14] there's only so many price increases.
[133:15] There's only so many tweaks to ads. Only
[133:17] so many tweaks to Google search. Only so
[133:20] only so many ways that Amazon can screw
[133:22] merchants. So in that tech depression,
[133:25] which you think it might be triggered in
[133:27] 2027, is that a cascading downstream
[133:31] economic depression? Because the stock
[133:33] market is heavily dependent on these
[133:35] companies. The stock market sees a
[133:36] pullback, investors stop investing, they
[133:38] get panicked.
[133:40] >> Yes. I think that because
[133:41] >> what's the sort of downstream
[133:42] consequence the sort of domino effect
[133:44] >> there's so much to imagine that it's
[133:46] difficult to capture everything but
[133:48] there are a few things that worry me
[133:49] first of all a ton of American money
[133:51] just regular people's money retail
[133:53] investors are in these companies and
[133:55] they bought into the magnificent 7
[133:56] thinking the number go up forever is the
[133:58] largest company on the Fortune 500 and
[134:01] NASDAQ as well and like 7 to 8% of the
[134:04] S&P 500 that company when in when the
[134:07] bottom falls out from Nvidia and we
[134:08] haven't really got into it but Nvidia is
[134:10] doing the most circular of financing,
[134:11] feeding companies money so that they can
[134:13] raise debt to buy more GPUs. I think
[134:16] Nvidia's revenue could go 50 to 70%
[134:18] down. I think that Nvidia could put
[134:20] Nvidia back in 2022 was making
[134:22] singledigit billion dollars.
[134:23] >> And what happens though, I'm thinking
[134:24] about like Jenny and Dave that are
[134:26] watching this right now and they are
[134:27] just normal people
[134:29] >> with normal jobs.
[134:30] >> People's retirements are going to
[134:32] contract severely and I don't believe
[134:34] they're going to return to those values.
[134:36] And I think that because so much of the
[134:38] value of the S&P 500 and Russell 1000
[134:40] index comes from these four companies
[134:42] and the rest of the magnificent 7. So
[134:44] Apple, Tesla, Meta as well. And the
[134:47] thing is I don't know what happens after
[134:50] that because venture capital has also
[134:53] more than half of venture capital last
[134:54] year went into AI. I think most venture
[134:56] capital investments in AI are going to
[134:58] zero because when it comes to building a
[135:00] company on top of an LLM, all of those
[135:01] are unprofitable too. And the thing is
[135:04] LLM companies have not really been
[135:06] acquired. The exception being Cursible
[135:08] by Elon Musk for the coding side, but
[135:11] you have Cognition, which is just
[135:12] another LLM company raising a $26
[135:15] billion valuation. That means that
[135:17] company has to go public cuz who's
[135:18] buying a company at $26 billion other
[135:20] than Elon Musk. And there were rumors
[135:22] that Elon Musk was trying to buy them as
[135:23] well. Is Elon Musk just going to pick
[135:25] off every like LLM company like going to
[135:27] [ __ ] TJ Maxx for AI? Like Jesus
[135:29] Christ.
[135:29] >> So is that a recession you're
[135:31] describing? It is a recession, but it's
[135:33] also a depression within people's
[135:35] retirements. Like I'm talking about 20,
[135:37] 30, 40% off the top of these companies
[135:39] stock value.
[135:40] >> Economic contractions, recessions
[135:41] consistently lead to job losses and
[135:43] rising unemployment. When an economy
[135:44] contracts, the mechanism driving job
[135:46] losses typically follows a predictable
[135:47] sequence. Falling demand, consumers and
[135:50] businesses spend less money, causing
[135:51] revenues across most industries to drop.
[135:53] margin compression. With lower revenue
[135:56] and often fixed overhead costs like rent
[135:58] or debt, corporate profit shrink, and
[136:00] lastly, cost cutting measures to survive
[136:01] or protect profit margins, businesses
[136:03] freeze hiring, reduce hours, and resort
[136:05] to layoffs. Yes, that's that would all
[136:08] happen. But the thing is, we're talking
[136:09] about equity values dropping and we're
[136:11] talking about there not really being a
[136:13] home for that value or that money. So
[136:17] much is riding on these companies, but
[136:19] you can't bail it out. You can
[136:21] theoretically bail out OpenAI. I don't
[136:23] think it happens. You could pump these
[136:24] dogs full of money and keep them alive
[136:26] for a bit, but at some point they're
[136:28] going to have to start. They have
[136:29] between these two companies, Anthropic
[136:30] and Open AI, you have $1.1 trillion of
[136:34] commitments.
[136:35] >> Just OpenAI.
[136:36] >> Oracle is building 7.1 gawatt of data
[136:39] centers. So over $400 billion worth just
[136:42] for OpenAI. There is not a customer on
[136:44] Earth. And Oracle's revenue has been
[136:45] flat the last 15 years when you adjust
[136:47] for inflation. Without Open AI, Oracle
[136:49] dies. So you think open AAI is going to
[136:51] crash and run out of money and that's
[136:52] going to cause this domino effect across
[136:54] these other big tech companies which is
[136:56] going to impact the stock market and
[136:57] impact the broader economy.
[136:59] >> Yes. And also the tens of thousands of
[137:01] people that will be laid off from the
[137:02] tech sector. But also the venture
[137:04] capital thing is significant because
[137:05] venture capital has been having one of
[137:08] the most historic
[137:10] bad runs in history since 2018. The
[137:14] average return from venture capital
[137:16] total value put in. So the amount of
[137:17] money you get back for your dollar is
[137:19] between8 and 1.21 meaning for every
[137:21] dollar you invest you get 80 cents to
[137:23] $120
[137:24] >> paper gains.
[137:25] >> Well no that's just actual g like actual
[137:27] returns. Paper gains they'll give you
[137:28] but even then internal rate return which
[137:30] is a whole separate thing even that's
[137:32] not very happy. But long story short
[137:34] very simple venture capital is not
[137:36] making money come out. Venture capital
[137:38] is not actually providing returns.
[137:40] >> They're celebrating paper gains.
[137:42] >> They're celebrating paper gains
[137:43] >> and they're raising off paper gains.
[137:44] >> Mhm. And actually paper gains I mean
[137:46] just being able to say oh look the
[137:47] valuation of anthropic went up. So
[137:49] that's
[137:49] >> but that's that's what Google and Amazon
[137:51] were doing. Google's last quarter they
[137:53] boosted their net profits profits on
[137:55] paper by $99 billion because of the
[137:58] increased value of their SpaceX holding
[138:00] and their anthropic holding. And again
[138:03] the fact that this is happening is
[138:05] insane and the fact it's not a scandal
[138:07] is insane but we live in this culture I
[138:09] guess. But everyone is really benefiting
[138:12] right now. Oh, it's really that it's
[138:14] that great tweet. It's like when you're
[138:15] reaping, it's like, "Yeah, [ __ ] yeah,
[138:17] this rocks." Sewing. Ah, [ __ ] This
[138:19] sucks. Because right now, they're all
[138:20] like, "Yeah, all the speculative gains
[138:22] are awesome. The paper gains are
[138:23] awesome. The theoreticals of anthropic
[138:25] being worth $2 trillion. Wow. The
[138:27] articles we can write, the promises we
[138:29] can make. Then when the rubber meets the
[138:31] road, it's going to be pretty rough on
[138:33] them because the valuation of Amazon,
[138:36] Google, Microsoft, and Meta is based on
[138:38] this idea that they will grow eternally,
[138:39] that they will grow forever. If that
[138:41] changes, to quote Ed Elson from ProfitG
[138:43] Markets again, it's this. They're all
[138:45] doing Botox right now. They're sinking
[138:46] money into it to make themselves feel
[138:48] young again and the market believes
[138:49] them. When the market doesn't, we're not
[138:51] just talking about a depression. I'm
[138:53] talking about the market valuing them
[138:54] like airlines and saying, "Yeah, you're
[138:56] real big and you make money off your
[138:58] existing products, but guess what? You
[139:00] don't have new [ __ ] You're just going
[139:02] to be doing this forever and we're going
[139:04] to value you as such."
[139:05] >> So, if it's Jenny and Dave, should they
[139:08] do anything differently? Should they be
[139:10] conserving money? If there's a recession
[139:11] or depression coming, should they be a
[139:12] little bit more conservative? Should
[139:13] they
[139:14] >> I Yes. I actually I actually think it's
[139:16] I don't know. I don't have money in the
[139:18] market. I think it's a casino. Casino
[139:20] pumped up by the media.
[139:21] >> Should they invest in the S&P 500?
[139:23] Should they invest in Open AI?
[139:24] Unfortunately,
[139:24] >> oh god, no. I honestly I live in cash
[139:27] right now. I live in cash. Yeah. I don't
[139:29] [ __ ] trust the market, man. Try and
[139:31] get some gains here. I'm like I'm not
[139:33] comfortable giving financial
[139:34] >> advice, but it's like if you like it's
[139:37] like you're gambling.
[139:38] >> Okay. be conservative. Things might get
[139:39] volatile.
[139:40] >> Yeah, it really is. It's going to be act
[139:41] as you would with volatility. Take the
[139:43] gains when you've got them.
[139:45] >> Don't sell everything, but be suspicious
[139:48] of tech. Like, that's actually the
[139:49] biggest thing. It's like be suspicious
[139:50] of what they're promising. If you're
[139:51] acting based on their promises, don't
[139:54] trust the promises. Trust that they are
[139:57] going to say what will make the stock
[139:59] run rather than what's actually
[140:01] happening. and that they will find every
[140:04] dodgy way to make you think something is
[140:07] happening rather than it's actually
[140:09] happening. Annualized run rate, great
[140:10] example. Microsoft said that they had 38
[140:13] $37 billion of annualized run rate in
[140:15] AI. You hear that, you go, they made 38
[140:18] $37 billion, right? Wow, that's so much
[140:21] run rate maybe month times 12. They
[140:24] don't even define it, but it's built to
[140:26] manipulate. And they do that because we
[140:28] don't have a functional SEC and we don't
[140:30] have a media environment that actually
[140:32] where skepticism is the priority and
[140:34] where protecting the readers is
[140:36] necessary.
[140:36] >> What would they say? They would say Ed
[140:38] this technology is going to be so great
[140:41] and so transformative that we are
[140:43] investing a ton of money
[140:45] >> um in advance of the value and utility
[140:49] showing up. That's what they would say,
[140:51] >> right?
[140:52] >> And I've heard your rebuttal, but I just
[140:53] wanted to express I think that's their
[140:55] sentiment. I'm not defending them or
[140:57] anything. I'm just I'm trying to provide
[140:58] enough like balance to we see if we can
[141:01] dance between these these two
[141:03] perspectives.
[141:05] >> And a lot of people would say that
[141:08] there's going to be a blood bath because
[141:09] they can't all win big in the way that
[141:12] they're kind of describing. So,
[141:13] someone's going to have to lose. And
[141:14] >> when one of these players starts to lose
[141:16] big, I think it could, as you say, there
[141:18] could be some kind of domino effect or
[141:19] contraction.
[141:20] >> Yeah. And I think the thing that people
[141:22] want to believe is they the com bubble
[141:24] thing. It's like it worked out
[141:25] afterwards because Amazon, Oracle, they
[141:29] didn't die after the com bubble. They're
[141:30] actually fine. This isn't like that.
[141:32] They're bigger companies. They're have
[141:34] bigger promises. And even I'm not like
[141:36] Oracle I actually think could die. I RIP
[141:39] Larry. What couldn't happen to a nastier
[141:41] man? They'll probably
[141:42] >> You don't like these people, do you?
[141:43] >> No, I No. Again, I asked this question
[141:46] purely because I want an answer, not
[141:47] because I agree or disagree. But um why
[141:50] don't you like these these people? I
[141:53] don't like being misled and I don't
[141:55] think regular people like being misled
[141:57] either. And I really don't think that
[141:58] the average person can get away with
[142:01] bullshitting as much these companies do.
[142:03] And I don't think the average person
[142:04] gets anywhere near the level of
[142:06] affordance for failure and lying as
[142:08] these companies do. And I think there is
[142:10] a real economic and human cost to
[142:12] allowing these companies to run rampant
[142:14] and promise the world and never really
[142:16] get called up on it. The tepid nature of
[142:19] criticism these days is so frustrating.
[142:21] There are some really great critics out
[142:23] there that really great people, but it's
[142:25] like
[142:27] seeing these ultra rich, ultra wealthy,
[142:29] ultra powerful people lie through their
[142:31] [ __ ] teeth or misstate or whatever
[142:33] people want to call it, it turns my
[142:35] stomach. And I hate seeing people being
[142:38] misled. And I feel like I write at such
[142:40] length because I really want people to
[142:42] see why I've come to a conclusion. Am I
[142:43] right? Am I wrong? I think I am. Of
[142:45] course I do. But I also
[142:48] I just find it loathome. I find these
[142:51] companies don't make good products
[142:52] anymore. They don't care about their
[142:54] customers and and they treat their
[142:56] customers with contempt.
[142:59] >> If people want to go read more about
[143:01] your work, um you have a great Substack
[143:03] >> Ghost actually. It looks exactly like I
[143:05] moved off of Substack in 2024.
[143:06] >> Oh, okay. And you also have a podcast
[143:09] you do.
[143:09] >> Yeah, Better of Flame.
[143:10] >> Um I'm going to link both of them below.
[143:12] So, if anyone wants to read more, get
[143:13] more detail and and follow Ed. I think
[143:15] it's
[143:15] >> I would highly recommend. It's it is
[143:17] fascinating. And you know what? One of
[143:19] the things people um sometimes struggle
[143:20] with when they listen to podcasts is you
[143:22] get lots of different opinions. And
[143:24] weirdly, I think they think of some
[143:25] people assume podcasts are going to be
[143:26] like one person saying the same thing as
[143:29] the next person and then the next
[143:30] person. That is just not the nature of
[143:32] information in the world and opinions
[143:33] and progress and discussion. What what
[143:35] happens is people have different
[143:36] opinions. And I think my job, but also
[143:38] the listener's job is to try and pass
[143:40] through it and over time collect more of
[143:42] these reference points from different
[143:44] people and and do your own research.
[143:47] >> Yeah. whether it's on your health or
[143:48] whether it's on something like this is
[143:49] to watch endear and research and to
[143:51] learn and I would say also never believe
[143:54] one person never believe one particular
[143:56] perspective religiously you know collect
[143:59] a body of evidence and follow follow the
[144:01] evidence yourself but I love watching
[144:03] your YouTube um because it provides a
[144:06] different opinion and that challenges me
[144:09] to think beyond my current opinion about
[144:13] what might be possible so when I've
[144:14] heard you talking about how this is an
[144:16] economic bubble and I've heard you talk
[144:18] about the capex spend on with these big
[144:20] sort of frontier AI labs. It really did
[144:23] make me pause for a second and it really
[144:25] did make me consider
[144:27] that there could be a bit of fazy going
[144:30] on here.
[144:30] >> Yeah.
[144:31] >> And then it made me reflect on history
[144:32] and go, you know, through history
[144:33] there's always a bit of fazy in these
[144:34] moments and oh that's an interesting
[144:36] take on what's going to happen in 2027
[144:38] 2028 when there's a bit of a market
[144:39] pullback and so I highly recommend
[144:41] people go watch because you do you
[144:42] challenge me to think differently. Um,
[144:44] >> yeah.
[144:44] >> And we need some of those contrarian
[144:46] voices to to have honest discussions.
[144:49] So, thank you for doing what you do.
[144:50] Really appreciate it. And I find you to
[144:51] be a very compelling, captivating
[144:53] communicator. And I've I feel like I've
[144:54] learned a lot today. So, I appreciate
[144:56] that. We have a closing tradition.
[144:58] >> Yeah.
[144:58] >> Where the last guest leaves a question
[144:59] for the next guest not knowing who
[145:00] they're leaving it for. And the question
[145:02] left for you is given that high quality
[145:04] relationships are important for health
[145:06] and longevity, what should we be doing
[145:08] to improve our relationships and social
[145:11] connection? So this is actually
[145:14] connected to the AI bubble. So I am a
[145:17] critic. I'm a skeptic. What quote I have
[145:20] found that showing and appreciating and
[145:24] loving the people around you and
[145:25] uplifting them and me and and raising
[145:27] them up as you succeed is the way we do
[145:29] that. Your success should be everyone
[145:30] around you. It's not economic. It's
[145:32] talking about Matt Hughes for a while
[145:34] made me really happy. This whole thing
[145:37] has been at times quite grueling and
[145:39] quite negative and quite brutal. But the
[145:41] love I found and the joy I found from
[145:44] community and the people around because
[145:46] even in the in the small groups of
[145:48] haters even like Gary Marcus and sort of
[145:50] the people I talked to Edward on Grao
[145:52] Jr. Molly White, Brian Merchant, there
[145:54] are so many people who have been loving
[145:56] and caring. And I think within
[145:58] especially these very critical moments
[146:00] when you're like very much dialing in on
[146:02] how negative things are, how bad things
[146:04] are, finding the people who maybe find
[146:08] it repulsive, too. Finding the people,
[146:10] >> finding your people who can be and the
[146:12] people who will talk to you about it.
[146:13] Even like Troy and Jake, my my trainers
[146:16] who's so excited about this. um even
[146:18] talking to them about the [ __ ] as normal
[146:19] people knowing that there are people
[146:21] there going through their own struggles
[146:22] but also to just give you the
[146:25] perspective and also remind you that you
[146:27] are human to and focus I know this is
[146:29] kind of a all over the place point but
[146:30] it's just it's really easy to get hard
[146:32] locked on everything in life and to
[146:35] >> kind of get away from why you do things
[146:37] and focus too much on the work when the
[146:39] most important thing at times is just to
[146:41] know there are other people feeling the
[146:42] way you do and when I hear from my
[146:44] listeners and my readers a lot the most
[146:45] common thing they feel is they feel like
[146:47] they have a voice and they feel like
[146:48] someone is there for you.
[146:50] >> And I don't think it can be understated
[146:52] how much it means when you just reach
[146:54] out to someone you love and tell them
[146:55] you love them. Tell them their [ __ ]
[146:56] rocks. Say that their [ __ ] bangs. Tell
[146:58] everyone you when you like an artist or
[147:01] a writer they were a podcast like this.
[147:02] Tell them you [ __ ] love it. We don't
[147:04] do this enough and we need to do it
[147:06] more. Well, that's a good closing
[147:08] message. So, if you do have you have
[147:10] enjoyed the conversation today with Ed,
[147:11] please do let Ed know that you love it
[147:13] down below. Um, but please do leave your
[147:15] opinions down below and I shall read all
[147:16] of them. Ed, thank you so much. I'll
[147:18] link to your website, but also to your
[147:20] YouTube channel where people can learn
[147:22] more and I would highly recommend you do
[147:23] because it is truly fascinating and I
[147:25] think we need more voices that are
[147:26] demystifying a lot of the fugazi and the
[147:28] narrative in this moment in time and you
[147:30] are certainly one of them. I really
[147:30] enjoyed the conversation. Thank you so
[147:32] much.
[147:32] >> YouTube have this new crazy algorithm
[147:34] where they know exactly what video you
[147:36] would like to watch next based on AI and
[147:38] all of your viewing behavior. And the
[147:40] algorithm says that this video is the
[147:43] perfect video for you. It's different
[147:45] for everybody looking right now. Check
[147:47] this video out and I bet you you might
[147:49] love it.
