Sacks Says White House AI Open-Source Ban Would Cripple US Tech
All-In Podcast
Restricting American developers from using Chinese open-source AI models will destroy US competitiveness and hand global manufacturing leadership to Beijing.
## Background
Published on YouTube by the All-In Podcast on July 24, 2026, this episode features the Besties discussing major technology and political news. According to the page description, the episode covers the panic around China's Kimi K3 AI model, Anthropic and OpenAI's growth rates, Anthropic's $1.5B piracy settlement, tech stock reactions to rising AI capex, and a segment on NYC rental property rights.
## TL;DR
The hosts debate the high-stakes policy fallout over Chinese open-source AI model Kimi K3, warning that White House proposals to ban open weights will harm American innovation and lead to regulatory capture by companies like Anthropic and OpenAI. The discussion analyzes Anthropic's record-breaking $1.5 billion copyright settlement, evaluating its legal precedents and corporate hypocrisy around model outputs versus inputs. Shifting to public markets and housing policy, the panel breaks down massive infrastructure capex surges from Google and Tesla before critiquing socialist NYC tenant proposals that treat evictions as violence and create thousands of vacant ghost apartments.
## Atmosphere & style
The episode captures a fast-paced, intellectually combative, and highly opinionated atmosphere characteristic of veteran tech insiders. Sacks and Chamath drive intense, analytical policy and financial debates with sharp candor, while Friedberg grounds arguments in historical macroeconomics and philosophy. The serious discussions around AI capture and economic policy are continually balanced by playful camaraderie, self-deprecating banter, and witty comedic interruptions among the Besties.
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=0s) Open-source AI model Kimi K3 sparked a debate in the White House over whether to ban Chinese open-source AI models.
- [03:04](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=184s) David Sacks argues that banning open-source models would be a tragic mistake for U.S. competitiveness and that Anthropic's warnings about distillation represent regulatory capture.
- [08:13](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=493s) Chamath Palihapitiya warns that forcing U.S. enterprises onto overly expensive closed models creates a 'token tax' that could destabilize capital markets.
- [14:53](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=893s) David Friedberg compares model distillation to traditional engineering benchmarking, arguing that learning from outputs is not software theft.
- [16:39](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=999s) The hosts emphasize the legal distinction between proprietary model weights and model outputs, noting the hypocrisy of frontier labs complaining about output training after scraping global data.
- [20:15](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1215s) Open-source AI operates similarly to open-source web infrastructure like Apache, democratizing access and diffusing economic value across the broader economy.
- [24:03](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1443s) Monopolistic pricing power for foundational AI models is evaporating rapidly as base intelligence becomes commoditized, shifting durable profits to hyperscale cloud infrastructure.
- [28:53](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1733s) David Sacks maintains that U.S. frontier labs retain a significant capability lead and immense ARR growth despite open-source competition.
- [41:33](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2493s) China's macroeconomic strategy aims to commoditize the Western knowledge economy through open-source AI while retaining dominance in physical manufacturing and energy infrastructure.
- [48:25](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2905s) Anthropic agreed to a historic $1.5 billion copyright settlement over its use of pirated books from shadow libraries like LibGen.
- [50:47](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=3047s) Anthropic's public framing of model distillation as 'distillation attacks' was a strategic attempt to gain regulatory protections without undermining its own copyright fair-use defenses.
- [01:07:07](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4027s) Google and Tesla reported massive surges in AI capex, driving negative cash flows and stock drawdowns despite strong top-line business growth.
- [01:09:32](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4172s) Google's massive investments in cloud and custom silicon are backed by a historical 32% return on invested capital, positioning it to win regardless of model fragmentation.
- [01:15:08](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4508s) Building enterprise-grade hyper-scale cloud infrastructure with five-nines reliability requires hundreds of billions in capital investment, creating an enormous moat for existing cloud providers.
- [01:17:36](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4656s) New York City tenant protection proposals and anti-eviction activism threaten basic private property rights and landlord financial stability.
- [01:18:57](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4737s) David Friedberg and David Sacks defend private property rights as fundamental to Western democracy, warning that undermining them degrades housing quality.
- [01:25:08](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=5108s) Over-regulation and rent controls in markets like New York cause market distortions, resulting in an estimated 50,000 vacant 'ghost apartments'.
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=0s) The Kimi K3 Panic and White House Open Source Debate
The episode kicks off with the Besties diving straight into the defining technology story of the moment: the fallout from Moonshot AI’s release of Kimi K3. The Chinese open-source model has sent shockwaves through Silicon Valley and Washington, delivering performance on par with closed frontier models like Opus 4.8 and GPT-5.6 while running roughly 50% cheaper. Reminiscent of the early 2025 DeepSeek panic, Kimi K3 has ignited an intense debate reaching all the way to the White House over whether the United States should ban Chinese open-source AI models.
The political and policy fallout has escalated rapidly. Axios reported that the White House was actively weighing a ban on Chinese open-source models after former Chief Technology Officer Michael Kratsios pointed out evidence that Moonshot AI distilled Anthropic’s models to build Kimi K3. However, Wired later reported internal division, noting that Secretary of Commerce Howard Lutnick opposes a ban, favoring instead incentives for American frontier labs to build superior open-source alternatives. Meanwhile, prediction markets reflected the mounting uncertainty, with Polymarket odds of the U.S. government banning an open-source model in 2026 jumping from 22% to 45% in just a few days. David Sacks points out that this regulatory panic aligns precisely with what he predicted months prior—a coordinated push using "safety guardrails" rhetoric to justify shutting down open source.
* **Performance Parity:** Kimi K3 matches top-tier closed models (Opus 4.8, GPT-5.6) at approximately 50% lower cost.
* **White House Divide:** Axios reports potential bans under discussion; Wired reports Howard Lutnick prefers incentivizing U.S. open-source development over bans.
* **Prediction Markets:** Polymarket odds on a 2026 U.S. open-source AI ban surged to 45%.
---
[03:04](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=184s) Sacks on White House Deliberations and Anthropic's "Regulatory Capture"
David Sacks provides inner-circle context on the White House's stance, clarifying that no official decision has been made to ban open-source models and that President Trump is actively soliciting diverse perspectives before determining policy. Sacks strongly cautions against any action targeting open source, warning that prohibiting American developers from utilizing public domain contributions from abroad would be a "tragic mistake" that would severely backfire on U.S. competitiveness. He argues forcefully that "you cannot punish American developers" or cut off access to global innovations while foreign competitors freely leverage them.
Sacks directs sharp criticism at Anthropic, accusing the company of aggressive regulatory capture under the guise of national security concerns. He highlights Anthropic’s astronomical financial trajectory—scaling from $10 billion to over $70 billion in annual recurring revenue (ARR) in a single year, making it the fastest-growing tech company at scale in history. Sacks argues that Anthropic does not need government protection and that their claims about distillation are disingenuous. If industrial-scale distillation were truly a critical threat, Anthropic could implement basic "Know Your Customer" (KYC) identity verification or IP blocking to stop Chinese actors from scraping their API, rather than lobbying the government to ban competition.
> "If stopping distillation was their primary objective, Anthropic would push to ban Chinese access to American models, not American access to Chinese models."
* **Anthropic's Growth:** Exploded from $10B to over $70B ARR within a single year.
* **Proposed Distillation Mitigation:** Anthropic could curb industrial scraping through KYC requirements and account verification, but avoids doing so to prevent slowing their user growth.
---
[08:13](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=493s) Chamath on Distillation Realities, Model Commoditization, and the "Token Tax"
Chamath Palihapitiya breaks down the mechanics of model distillation—the process of querying a premier model millions of times to capture its reasoning outputs and train a smaller model. He dismisses the outrage over distillation as a "red herring," pointing out that virtually every major AI player has engaged in similar data extraction, comparing the finger-pointing to the classic "nine Spider-Men pointing at each other" meme. Anthropic recently settled a $1.5 billion copyright lawsuit with publishers, OpenAI remains embroiled in litigation with *The New York Times*, and Chinese labs distill American outputs.
The underlying driver of this panic, Chamath explains, is the rapid commoditization of foundational models. Because performance criteria become replicable within weeks, foundational models offer no sustained long-term economic moat. Closed frontier labs are desperately engaging in a "valuation preservation game" to protect inflated 25x–50x revenue multiples. Chamath warns that if the U.S. government steps in to ban open-source alternatives, it would force American enterprises to pay an exorbitant "token tax"—paying 50 to 100 times more for closed U.S. AI models than international competitors pay for open alternatives, ultimately destabilizing capital markets and harming U.S. economic posture.
> "If the United States government intervenes, it will tank the stock market. Period. Not debatable."
* **Distillation Mechanics:** Involves extracting knowledge by collecting outputs from tens of millions of prompt queries to train downstream models.
* **Valuation Mismatch:** Closed models command 25x to 50x valuation premiums over open alternatives despite rapid performance convergence.
* **Enterprise Risk:** Banning open source forces American companies like Coca-Cola onto an "island of overly expensive AI," putting them at a severe cost disadvantage against global rivals.
---
[14:53](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=893s) Friedberg on Distillation as Traditional Industry Benchmarking
David Friedberg broadens the debate by contextualizing distillation as a standard, time-tested engineering practice rather than inherently malicious theft. Drawing parallels across traditional industries, Friedberg explains that analyzing an end product to refine one's own design occurs in every sector, from automotive engineering to software development. He recalls his early tenure at Google, where engineers systematically submitted millions of queries to rival search engines like Yahoo and Microsoft to benchmark result quality and refine Google's ranking algorithms.
Friedberg emphasizes that examining model outputs is no different from traditional competitive benchmarking; it does not constitute stealing proprietary source code or infiltrating private servers. While terms-of-service violations or copyright boundaries must be managed by the service providers hosting the APIs, treating output-based learning as a national security crisis misinterprets how technology advances. He agrees with Sacks and Chamath that kneecapping open source in the name of curbing distillation would severely damage the broader innovation ecosystem.
[16:39](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=999s) Model Output vs. Model Weights: The Copyright and Distillation Debate
The conversation delves into the legal and technical nuances surrounding model distillation, intellectual property, and copyright. The hosts emphasize the crucial distinction between proprietary **model weights**—the actual software code and numerical parameters—and **model outputs**. Stealing weights from frontier labs like Anthropic or OpenAI would constitute unambiguous software theft. However, using model outputs to train and refine new models (distillation) is a fundamentally different process. The hosts point out the glaring hypocrisy of American AI giants complaining about Chinese distillation: OpenAI and Anthropic have aggressively maintained in court (such as in the ongoing *New York Times* copyright lawsuit) that they are legally entitled to scrape and train on the world's published creative outputs regardless of terms of service or creator consent.
Former government official and tech executive Michael Kratsios's perspective is referenced, with the speaker emphasizing their own national security credentials as an early voice warning about distillation after DeepSeek's release in January 2025 and co-authoring the *"Winning the AI Race"* report. While the White House and policymakers view covert, industrial-scale terms-of-service violations by foreign actors as an ethical and regulatory challenge, the besties argue that banning open-source models in response would be a massive self-inflicted wound. Banning open source would handicap American enterprise while the rest of the world continues to leverage open tech. The burden should remain on frontier labs like OpenAI and Anthropic to technically secure their APIs and enforce their own terms of service rather than relying on heavy-handed government bans.
* **Model Weights vs. Outputs:** Weights are the underlying numerical parameters/software code; outputs are the generated text/data.
* **The *New York Times* Parallel:** OpenAI argues training on NYT text to derive weights is legal fair use; Chinese labs use the exact same logic when training on American model outputs.
* **DeepSeek & Policy Context:** Distillation concerns gained mainstream national policy attention following DeepSeek's breakthrough in January 2025.
---
[20:15](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1215s) Open Source, Free Speech, and Economic Value Creation
David Friedberg lays out two primary arguments for protecting open-source AI: the logistics of free speech/enforcement and macroeconomic value distribution. From a technical and legal standpoint, an open-source model is simply a downloadable package of code and parameters—much like downloading a digital book or text file. Once stored locally, it can run entirely offline without an internet connection. Attempting to restrict or outlaw open-source software opens a legal and administrative "can of worms," as it effectively criminalizes the execution of code already sitting on a user's computer, setting a dangerous precedent with little legal backing.
Drawing on historical internet paradigms, Friedberg compares the current AI landscape to the early web. Netscape originally commercialized proprietary web browsers and web servers, building a high-margin closed system. However, open-source innovations—such as the Mozilla Foundation's Firefox (later superseded by Google Chrome) and the Apache Software Foundation's HTTP server—democratized infrastructure. Because anyone could deploy a free Apache web server, value accrued to the broader web ecosystem (giving rise to Google, Amazon, eBay, Etsy, and millions of small businesses) rather than a few proprietary gatekeepers like Netscape, Microsoft, or walled gardens like AOL and CompuServe. In the same way, open-source AI ensures that economic benefits diffuse globally across thousands of AI-integrated enterprises rather than remaining concentrated in the hands of three or four tech oligarchs.
> "If open-source AI takes off, then all the worries that Bernie Sanders and Elizabeth Warren and all the socialists are harping and larking about are not going to be the case anymore... AI becomes a power for good for creating an open economy."
* **100x Cost Reduction:** Open-source AI models deliver inference up to 100 times cheaper than closed proprietary APIs, directly benefiting enterprise productivity and consumers.
* **Historical Precedents:** Netscape's proprietary server dominance was undone by the open-source Apache HTTP server, unlocking the modern internet economy.
---
[24:03](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1443s) Value Evaporation in AI Frontier Labs vs. Cloud Infrastructure
Chamath Palihapitiya highlights an unprecedented phenomenon in Silicon Valley history: the speed at which pricing power and value capture are evaporating for frontier AI model developers. In previous tech cycles, companies that absorbed hundreds of billions of dollars in capital enjoyed multi-year or decade-long windows of monopolistic pricing power before facing commoditization. In AI, that transition from exclusive provider to commodity vendor is compressing into a matter of months. Capital allocators evaluating these businesses 10-year discounted cash flows are increasingly concluding that base intelligence is becoming a ubiquitous commodity with a clearing price approaching zero.
Jason Calacanis displays third-party revenue tracking charts showing a recent revenue growth deceleration for Anthropic, alongside data from OpenRouter demonstrating that open-source models (including leading Chinese models) now capture well over 50% of token volume. Furthermore, a massive volume of "dark tokens"—inference run on private, in-house enterprise servers using open-source models—goes entirely unrecorded on revenue charts. Palihapitiya notes that while closed frontier labs may struggle to retain astronomical long-term valuation premiums, the real durable profits are flowing directly to hyperscale cloud providers (AWS, Google Cloud, Tesla/xAI) supplying the raw compute, energy, and hosting infrastructure.
> "I've never seen it in my 25 years in Silicon Valley where a sector of the economy can absorb hundreds and hundreds of billions of dollars... and economic pricing power effectively evaporates in months."
* **Anthropic Target vs. Headwinds:** Anthropic originally targeted a 10x ARR growth path (scaling toward $100 billion ARR), but third-party data indicates recent growth headwinds due to open-source competition.
* **Dark Tokens:** Inference conducted on self-hosted, on-premise hardware represents a huge unmeasured volume of AI usage that generates zero revenue for proprietary API providers.
* **Value Capture Shift:** Similar to web browsers acting as mere gateways to web applications, base AI models act as conduits while long-term value accrues to applications and cloud infrastructure.
---
[28:53](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1733s) Sacks' Counterargument: Frontier Lab Growth, Kimi K3, and GPT-6
David Sacks pushes back against the notion that closed-source frontier labs are dying or that Chinese open-source models have permanently caught up. Addressing the recent panic around Moonshot AI's **Kimi K3** model, Sacks argues that initial fears were overblown. Detailed technical breakdowns (such as analysis by Stratechery's Ben Thompson) reveal that Kimi K3 does not possess a dramatic cost advantage over US models and only excels in specific niche benchmarks like front-end web development, rather than outperforming American models overall. Furthermore, Sacks emphasizes that US frontier labs have major unreleased models in reserve, noting reports of OpenAI's Sam Altman briefing Washington officials on **GPT-6**, which is expected to represent another massive step-function capability leap.
Far from stagnating, Sacks asserts that OpenAI and Anthropic remain the fastest-growing technology companies at scale in history. While Anthropic experienced a minor temporary blip, it remains well on track past its massive revenue milestones. Meanwhile, OpenAI's growth has re-accelerated significantly—driven by products like Codex—putting them on pace to smash through their annual ARR projections. Sacks contends that both open-source and proprietary closed models will coexist as massive commercial successes, and that the United States retains a clear 6-month capability lead over Chinese competitors.
* **Kimi K3 Reality Check:** High scores on front-end coding arena benchmarks do not translate to overall frontier dominance, and operating cost parity remains close to US equivalents.
* **OpenAI ARR Acceleration:** OpenAI's internal revenue projections were revised upward toward $60 billion ARR by year-end as coding tools like Codex drive enterprise adoption.
* **Pipeline Capabilities:** US frontier labs maintain an estimated 6-month lead with next-generation architectures like GPT-6 preparing for release.
[31:53](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1913s) The Frontier Model Duopoly and the IPO Headwind Debate
David Sacks and Chamath Palihapitiya analyze the financial momentum of the U.S. frontier model duopoly—OpenAI and Anthropic. Sacks highlights that despite narrative panics over Chinese open-source distillation like Kimi K3, Anthropic is forecasting roughly $75 billion in exit Annual Recurring Revenue (ARR), while OpenAI’s ARR run rate surged from $33 billion in May to $41.3 billion in July. He argues that distillation has been happening continuously since DeepSeek’s emergence in January 2025 without stopping their exponential growth. Instead of suffering real commercial harm, Sacks views these companies' regulatory pleas as a "LeBron-style flop"—exaggerating competitive threats during their IPO roadshows to bait the government into granting them a protected duopoly against open-source alternatives.
Jason Calacanis vigorously pushes back, arguing that open-source models have reached a critical tipping point that will severely disrupt Anthropic’s and OpenAI’s upcoming IPOs. He shares that startups across his portfolio are actively abandoning closed APIs in favor of self-hosted, local open-source models to avoid massive cost overheads. Calacanis warns that overspending on compute combined with customer churn will lead to massive margin compression and trap these AI giants in a bad financial position. While Chamath agrees that open-source models are taking over 95% of routine enterprise tasks, he argues the frontier companies will retain high terminal value by moving up the application stack into specialized verticals like life sciences or cybersecurity.
- **OpenAI Run Rate Growth:** Rose from $33 billion in May to $41.3 billion in July.
- **Anthropic ARR Projections:** Estimated around $10 billion mid-year, targeting roughly $75 billion in exit ARR.
- **Developer Workflow Tools:** Calacanis cites using Perplexity Computer as his daily harness to evaluate models across Grok, GLM-4.5, and Claude.
---
[39:28](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2368s) Open Source Innovation vs. The Threat to American Developers
Sacks argues that pushing for government bans on open-source models under the banner of "national security" or "IP theft" actively cripples American AI developers. He presents two concrete examples of U.S. teams building cutting-edge products using open weights: Mira Murati's new venture, Thinking Machines, which bootstrapped its flagship American open-source model off the Chinese base model Kimi K2.5, and Cursor's *Composer 2*, which was post-trained on top of Kimi K2.5 using Cursor’s proprietary coding data. Sacks stresses that once weights are placed in the public domain, they are simply local files hosted in American data centers with zero data packets flowing back to China. Treating derivative works as "tainted" threatens to drive a dagger through the entire domestic open-source ecosystem.
> "You take things that are in the public domain, you fork them, you make them your own... An American company has taken open-source contributions in the public domain, made it their own, and then developed their own model. And if you say that American companies can't do that or that somehow it's tainted with IP theft, you are basically going to put a dagger through the heart of the entire American open source ecosystem."
---
[41:33](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2493s) China's Long Game: Commoditizing Knowledge to Own the Physical Economy
David Friedberg zooms out to explain China's strategic macroeconomic playbook over a multi-decade horizon. For the past 50 years, the U.S. captured immense GDP by dominating the high-margin "knowledge and services economy"—converting bits to bits—while outsourcing low-margin physical manufacturing to China. Friedberg posits that all human technology ultimately reduces to "molecule conversion": using energy to rearrange matter into useful products like microchips, materials, or furniture at the lowest possible cost. By flooding the world with open-source AI models, China aims to collapse the monetary value of the Western knowledge economy down to zero, leaving global power anchored strictly in physical manufacturing and energy infrastructure.
```
Knowledge & Services Economy (Bits) ---> Commoditized by Open-Source AI ---> Value Drops to $0
Physical Manufacturing & Energy (Molecules) ---> Retained Advantage by China ---> Captures Global Value
```
- **U.S. vs. China Electricity Production:** The U.S. has ~1 terawatt of electricity capacity versus China's trajectory toward 8 terawatts.
- **Manufacturing Footprint:** The U.S. possesses ~10 billion square feet of manufacturing space compared to China's 200 billion square feet (a 20x advantage).
- **Incentive Bias:** The panel jokes about venture capitalists turning into overnight "China hawks," with Sacks noting that investors claiming open-source is a national security risk should first disclose whether they are sitting on Anthropic or OpenAI's cap table.
[47:10](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2830s) Enterprise AI Strategy and the "Promo City" Banter
The conversation opens with a discussion on open-source AI adoption in the enterprise. Most open-source models can easily handle 95% of standard tasks, meaning companies don't necessarily need to scramble for the absolute top-tier proprietary model just to get basic work done. Instead, businesses can rely on open source for the vast majority of their needs and only pay a premium or seek specialized models for the remaining 5%. However, large enterprises requiring dedicated support, wrappers, and employee tools should still partner with major providers like Anthropic, OpenAI, Grok, or Gemini.
The conversation quickly devolves into lighthearted banter and playful self-promotion among the hosts. They joke about offering absurd perks for large contracts or investments:
* **Jason Calacanis:** Jokes that signing a $10 million enterprise contract with "8090" gets customers a 5-minute Zoom call with three out of the four Besties (or a Zoom call with your mom).
* **David Friedberg:** Teases that a $1 million pre-purchase of potato seed unlocks the promo code "Sacks Poop."
* **Jason Calacanis:** Adds that anyone writing a $25 million check into Launch Fund 5 gets to guest-moderate an episode alongside "your boy J-Cal," officially declaring the segment "Promo City."
Jason pivots to the news that Anthropic settled a massive AI copyright lawsuit for **$1.5 billion**—the largest copyright settlement in U.S. history and the first major AI training lawsuit to settle. Anthropic was sued after downloading over 7 million books from pirated shadow-library sites like Library Genesis (LibGen) to train its Claude models.
Under the terms of the settlement, $101 million goes to legal fees, while affected authors receive roughly $3,000 per book, covering approximately 500,000 books. So far, 91% of eligible authors have claimed their payout. Jason sees this as a major validation of content providers standing together to demand compensation, arguing that publishers and news outlets must demand terms rather than letting AI companies freely index their work.
* **Settlement Amount:** $1.5 billion (largest in U.S. copyright history).
* **Scope:** ~500,000 books covered; 7 million total pirated books ingested from LibGen.
* **Payout Breakdown:** $101 million for lawyers; ~$3,000 per book to authors (91% claim rate so far).
---
[50:47](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=3047s) David Sacks on Regulatory Capture and the "Distillation" Hypocrisy
David Sacks pushes back on Jason's interpretation, arguing that the settlement does *not* prove that training AI on copyrighted content is inherently illegal. Sacks explains the legal nuance: Anthropic was caught red-handed because they pirated books off LibGen without paying for a single legitimate copy. Had Anthropic purchased even just one legal copy of each book, they would have had a strong fair-use defense in court.
Sacks goes on to slam Anthropic for extreme corporate hypocrisy. While Anthropic maintains that it is fair use for them to train on every creator's work on Earth (as long as they buy one copy), they simultaneously argue that no one should ever be allowed to train on *their* model outputs.
> "They believe it is fair use to train their models and derive their own weights based on fair use. However, they say that the one type of content that you should never be able to train on is *their* output... IP for we, but not for thee."
Sacks highlights how Anthropic selectively framed the issue:
* **The "Distillation Attack" Narrative:** In February 2026, Anthropic published a blog post coining the term "industrial-scale distillation attacks" to brand model distillation—a standard industry practice—as a rogue security threat.
* **Avoiding "IP Theft":** Notably, Anthropic avoided using the phrase "IP theft" in that post. Sacks points out they lacked the "chutzpah" to call it IP theft publicly because doing so would immediately sabotage their own fair-use defenses in court against copyright owners like *The New York Times*.
* **The Real Motivation:** Instead, Anthropic tried to convince Washington policymakers that Chinese distillation was a "national security threat" due to a lack of safety guardrails, hoping to secure regulatory capture and shut down open-source competition.
---
[55:22](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=3322s) Information Diffusion, Reviews, and the Limits of Copyright Law
David Friedberg introduces a thought experiment to examine where fair use actually begins and ends. He asks the group to imagine an author who publishes a book, keeps it closed to AI, and charges $30 a copy. If a reader buys the book, writes a detailed online review, and an AI ingests those public reviews to learn about the book, has the AI violated the author's copyright?
Friedberg argues that knowledge is naturally diffuse and cannot be completely locked away. While direct lifting and republishing of text is a clear copyright violation, an AI processing metadata, public reviews, and third-party analysis to form new abstract knowledge falls squarely into standard learning and transformation.
The discussion briefly derails into comedic teasing of Jason's book (*Angel*):
* **Jason Calacanis:** Reads a parody five-star review for a fake "great American novel" attributed to him, mocking its "eloquently brash blueprint."
* **Chamath Palihapitiya & David Sacks:** Joke about Jason's "slop cannon" of open-source AI and accuse him of "master virtue signaling."
---
[58:40](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=3520s) Legal Precedents, Market Competition, and the "Self-Own"
Returning to the legal reality, Jason points out that while knowledge transformation is complex, existing lawsuits are testing strict boundaries—citing *Thomson Reuters v. Ross Intelligence* (involving WestLaw data) as a key battleground where AI outputs directly compete with the original product.
Jason proposes that AI companies form a self-regulatory body and set aside 10% of their revenues into a settlement and licensing pool to compensate creators continuously. Sacks counters that content owners (especially the music industry "rabid dogs") will never settle for 10% and will demand 100%.
Sacks concludes by pointing out a potential fatal error in Anthropic and OpenAI's public relations maneuver. By weaponizing the term "distillation attack" against competitors while claiming "IP theft" behind closed doors, they have created a massive self-own (a "self-cell"). Had they simply argued that competitor distillation violated their terms of service or constituted a deceptive business practice, they could have protected their products without accidentally blowing up their own legal defenses in the fair-use lawsuits that threaten their core business models.
[01:02:50](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=3770s) The Publisher Backlash and Anthropic’s Regulatory Play
The discussion shifts to how content owners—from newspapers and magazine publishers to book authors—are attempting to protect their IP against AI indexing. Jason notes a emerging trend where content providers are demanding that Google separate its standard search crawler from its AI crawler. Currently, Google uses a unified bot for both, forcing publishers into a tough dilemma: allow AI training on their work or lose search engine traffic entirely. Historically, publishers lacked a united front, failing to extract royalties when web indexing first took off.
```
┌────────────────────────────────────────┐
│ Google Unified Bot │
└──────────────────┬─────────────────────┘
│
┌────────────────┴────────────────┐
▼ ▼
┌────────────────────┐ ┌────────────────────┐
│ Search Engine Index│ │ AI Model Training │
└────────────────────┘ └────────────────────┘
│ │
▼ ▼
(Drives Web Traffic) (Uncompensated IP Usage)
```
David Sacks points out the severe strategic flaw in Anthropic’s recent lobbyist push in Washington. By telling regulators that training on creator output without explicit consent constitutes "IP theft," Anthropic essentially confessed that its own product was built on stolen data. This declaration backfired spectacularly, activating the startup ecosystem—demonstrated by Y Combinator's Garry Tan and over 200 startup founders signing a protest letter—because framing model derivative output as IP theft risks declaring all downstream AI applications legally tainted.
* Anthropic’s declaration of unconsented model training as IP theft effectively creates legal exposure for its own foundational models.
* Garry Tan and over 200 startup founders submitted a joint letter warning that treating Chinese open-source models or trained derivative works as "stolen IP" threatens the broader technology startup ecosystem.
* The co-hosts joke about their upcoming 2027 book lineup under the fictional "All-In Books" publishing label, featuring gag titles like David Sacks's *Ferrari on My Wrist* and *The Fight to Save America*, alongside Chamath Palihapitiya's *A Sexual Italian Summer*.
> "If you go to Washington and you lay down with the dogs, don't be surprised if you wake up with the fleas... they may have poked the tiger."
---
[01:07:07](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4027s) Big Tech Earnings, AI Capex Surges, and SpaceX Valuation
Turning to public markets and fresh quarterly earnings, Google and Tesla both reported massive spikes in capital expenditures to build out AI infrastructure, resulting in negative free cash flows. Google's capex forecast rose to a massive $195B–$205B range for the year, sending its stock down 7–10% despite Google Cloud blowing past expectations to reach a $100 billion annual revenue run rate. Tesla's capex surged 140% year-over-year toward $25 billion, pulling its stock down 14%. Meanwhile, SpaceX is trading down 30% from its post-IPO high, settling at a $1.5 trillion market cap following its historic $2 trillion public debut.
| Company | Capex / Valuation Metric | Market Reaction / Note |
|:--- |:--- |:--- |
| **Google (Alphabet)** | $195B–$205B projected 2026 Capex | Stock fell 7–10%; negative free cash flow for the first time |
| **Google Cloud (GCP)** | $100B annual revenue run rate | Accelerated growth driven by enterprise AI infrastructure |
| **Tesla** | $25B projected Capex (+140% YoY) | Stock fell ~14% following cash flow compression |
| **SpaceX** | $1.5T current valuation (down from $2T) | Down 30% from peak post-IPO pricing amidst lockup expirations |
* Google reported negative free cash flow for the first time in its history as a public company due to unprecedented infrastructure spending.
* Tesla’s capital expenditure grew 140% year-over-year as it expands compute capabilities like its Colossus clusters.
* Google experienced a $100 billion mark-to-market write-up on its Anthropic holdings within a single quarter.
---
[01:09:32](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4172s) Why Google’s $200B Bet is a Long-Term Win
Chamath and Friedberg defend Google’s aggressive spending, arguing that Wall Street's negative reaction reflects short-term shock rather than underlying weakness. Chamath highlights that Google has maintained an astounding 32% average return on invested capital (ROIC) over its 25 years as a public company. Because of this historic track record, investors should trust management's decision to reinvest profits into compute rather than handing it back as dividends. Furthermore, Google is uniquely positioned: whether model development fragments into hundreds of open-source variants or concentrates among a few giants, Google profits at the silicon (TPUs), cloud infrastructure (GCP), and application layers.
Friedberg adds that Google Cloud Platform (GCP) is the best-positioned enterprise cloud layer to capture AI value because companies already host their core enterprise data—emails, documents, and drives—within Google Workspace. Even in a absolute worst-case scenario where Google's consumer AI apps fail, its low-cost compute infrastructure alone will generate massive cash flows by hosting third-party workloads. The co-hosts contrast this strategy with Apple, which spent over $900 billion on share buybacks and dividends over the last decade rather than building out massive web services or hyper-scale infrastructure.
* Google's 25-year historical Return on Invested Capital (ROIC) stands at an average of 32%.
* Google's capex spend represents approximately 20% of the entire U.S. annual military budget.
* Apple allocated over $900 billion over the past decade toward capital returns ($755 billion in buybacks and $140 billion in dividends) instead of building out a proprietary cloud infrastructure layer.
> "When you are a machine and a business model that compounds money at 32% over a 20-year average, you give these guys the benefit of the doubt. These are not people that are flying fast and loose."
---
[01:15:08](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4508s) The Engineering Reality of Cloud Infrastructure
The conversation addresses why building a competitive hyper-scale cloud infrastructure is extraordinarily difficult for latecomers like Apple. Chamath and Friedberg explain that becoming a true Cloud Service Provider (CSP) requires offering "five nines" (99.999%) of system reliability and uptime. Achieving basic uptime is relatively inexpensive, but each incremental order of magnitude in reliability causes capital costs to skyrocket exponentially:
Because Amazon Web Services spent 17 years and Google spent 13 years perfecting this underlying infrastructure, there remain only three dominant, hyper-scale cloud platforms globally. While Apple pursued a "do no harm" capital allocation strategy that favored financial engineering over technical infrastructure, the co-hosts ponder whether incoming leadership might pivot toward engineering-led infrastructure investments.
* Achieving 99.999% reliability ("five nines") requires hundreds of billions of dollars in continuous data center and hardware investment.
* Amazon spent roughly 17 years and Google spent over 12 years engineering their hyper-scale networks to achieve enterprise-grade redundancy.
<FollowUp label="Want me to walk through Part 6 of the transcript next?" query="Please provide the walkthrough for Part 6 of the transcript as soon as it's ready."/>
[01:17:36](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4656s) The "Violence of Evictions" and New York Housing Proposals
The conversation shifts to "Socialism Corner," focusing on a recent housing hearing held at New York City's Tenement Museum. During the event, a tenant advocacy group introduced a "rental ripoff report" containing proposals that would significantly restrict landlord authority. The draft policies propose barring landlords from charging fees for credit checks and preventing them from enforcing both a credit check *and* the traditional 40x monthly rent income standard simultaneously—forcing them to choose only one. Additionally, the legislation seeks to legally recognize tenant unions and institute a one-year rent freeze.
The hosts react with disbelief to a 22-second video clip from the hearing. In the clip, an activist wearing a heavy, industrial-style COVID mask explicitly frames the current system as violent, stating that the local administration is "emboldening us so that we no longer tolerate the violence of evictions as a matter of business as usual." The Besties poke fun at the clip's aesthetics, comparing the activist’s oversized hat and mask setup to characters from *Fat Albert* and expressing surprise that such intensive masking is still happening.
---
[01:18:57](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4737s) The Constitutional Foundations of Private Property Rights
David Friedberg launches into a historically grounded philosophical defense of property ownership, throwing back to late 18th-century American political thought. He quotes John Quincy Adams’s 1787 work, *A Defence of the Constitutions of Government of the United States of America*, noting Adams's warning that "the moment the idea is admitted into society, that property is not as sacred as the laws of God... anarchy and tyranny commence." Friedberg explains that the foundational pillar of the American experiment was an explicit rejection of European monarchies and feudal systems, where a tyrannical lord or cabal could seize a citizen's farm or belongings at any moment.
Friedberg argues that the socialist framing used by Democratic Socialists of America (DSA) types—labeling property owners as inherently evil and accusing them of "violence"—is a slippery slope. Because anarchy is merely a temporary state of lawless chaos where people steal from one another, it inevitably coalesces into tyranny when the strongest group takes permanent control. By stripping away private property rights under a guise of moralistic intent, these policies erode the very basis of American liberty. He insists that slipping even the tiniest bit on these rights creates a catastrophic cascading effect.
---
[01:23:00](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4980s) "Luxury Beliefs" and the Degradation of Multi-Family Buildings
David Sacks expands on Friedberg’s argument by looking at the practical consequences for the rest of the tenants living in rent-controlled buildings. When landlords are burdened with delinquent tenants who cannot be evicted, building income dries up, leading directly to deferred maintenance and dilapidated housing stock. Furthermore, Sacks shares an anecdote from a friend who manages apartment complexes, noting that the unruly squatters who escape eviction often make life miserable for their neighbors through midnight noise, loud music, punching walls, foul odors, and disorderly conduct in common areas. Elderly, long-standing residents of modest income bear the brunt of this downward spiral.
The hosts critique these policies as "luxury beliefs" held by highly educated, affluent progressives who live in gated communities or wealthy enclaves like the Upper East Side or Marin County. Because these wealthy advocates do not rely on public amenities like subways, buses, or city parks, they can easily defend homeless drug addicts or problematic tenants from an abstracted distance. The working class, conversely, is forced to navigate public spaces plagued by open drug use and public defecation, or live in buildings that effectively turn into poorly maintained housing projects.
---
[01:26:03](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=5163s) Supply and Demand Economics vs. The Socialist Framework
The Besties contrast New York's restrictive approach with places that have successfully managed housing costs through basic supply-and-demand market forces. Jason notes that cities like Tokyo, as well as high-growth states like Texas, Florida, and Nevada, figured out the solution long ago: relax permitting constraints and build more units, whether they are luxury apartments, single-family homes, or multi-family complexes. Chamath cites recent real estate data from Austin, showing a scientific correlation where aggressive permitting directly expands housing supply and lowers average rents.
Instead of pursuing permitting reform, politicians push counterproductive regulations that completely break the rental market's economics. Chamath points out that if a landlord is legally blocked from running credit checks, vetting a tenant's income, or executing evictions, the risk profile skyrockets. To absorb the inevitable losses from non-paying occupants, landlords will simply adjust by setting initial asking rents three to four times higher or demanding massive, multi-month upfront wire transfers.
---
[01:28:28](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=5308s) The Rise of "Ghost Apartments" in New York City
The discussion concludes with a look at the ultimate market failures resulting from over-regulation. Jason explains that due to New York City's strict renovation codes, bringing an older apartment up to legal standards can cost hundreds of thousands of dollars. When rent control prevents landlords from recouping those expenses, it becomes financially logical to leave the units entirely empty. This dynamic has created an estimated 50,000 "ghost apartments" across the city that are completely held off the market.
When Sacks asks if these empty units are just being diverted to Airbnb, Jason notes that New York effectively banned short-term rentals as well, leaving property owners with zero viable paths to profitability. Sacks concludes that the hostile regulatory environment leaves landlords with few choices but to let their properties sit empty, wait out the political cycle, or sell their assets and move their capital to more business-friendly jurisdictions.
* **Rental RiPoff Report Details:** A proposed New York policy framework that blocks landlords from executing credit checks or verifying the 40x rent income standard concurrently, while formalizing tenant unions and freezing rents.
* **Austin Housing Data:** Serves as a direct counter-example showing that relaxing permitting restrictions boosts housing supply and systematically drives down market-rate rents.
* **The Ghost Apartment Phenomenon:** An estimated 50,000 apartments in New York City sit vacant because heavy renovation codes combined with strict rent ceilings make it unprofitable for landlords to lease them.
[01:32:00](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=5520s) Ghost Apartments and the All-In Summit 2026 Pitch
Wrapping up their discussion on real estate regulations and private property rights, the hosts touch on the natural consequence of overly restrictive housing policies—landlords simply choosing to leave units vacant rather than risk bad tenants or unworkable rent caps, resulting in "ghost apartments," provided the property owners aren't weighed down by too much debt. From there, the hosts seamlessly pivot into a enthusiastic promotional pitch for the upcoming All-In Summit 2026. They encourage listeners who have been on the fence to make this the year they finally attend, highlighting a world-class lineup of speakers, top-tier networking, and an overall sharper focus on connecting attendees with industry peers.
The pitch underscores the high caliber of the summit's audience, noting that roughly 80% of attendees are founders, venture investors, or high-level tech operators. Beyond the professional networking and content, the hosts reminisce about the legendary entertainment, food, and drink from previous years, casually dropping past musical performers like Grimes and Diplo to showcase the event's iconic party vibe. Viewers are directed to `[allin.com/events](https://allin.com/events)` to secure their applications before the episode officially wraps up.
* Approximately **80% of summit attendees** are founders, investors, or high-level operators.
* Past summits featured high-profile musical headliners, including **Grimes** and **Diplo**.
[01:32:33](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=5553s) Wrap-Up, Outro Clips, and Credits
With the main topics covered, the group brings the episode to a close in their signature style. Jason Calacanis signs off on "another amazing, all-time legendary episode of the All-In podcast," trading heartfelt "love you boys" farewells across the panel before launching the show's classic outro music and catchphrases.
The final minute of the episode features a montage of comedic, unfiltered bloopers and banter. Among the clips, the hosts laugh about open-sourcing content to fans who have "just gone crazy with it," joke about absurd personal moments caught on camera, and poke fun at the intense "sexual tension" between the Besties. The show ends with its iconic signature tag: "I'm going all in."
## Notable quotes
> [03:04](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=184s) If stopping distillation was their primary objective, Anthropic would push to ban Chinese access to American models, not American access to Chinese models. — David Sacks
> [08:13](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=493s) If the United States government intervenes, it will tank the stock market. Period. Not debatable. — Chamath Palihapitiya
> [20:15](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1215s) If open-source AI takes off, then all the worries that Bernie Sanders and Elizabeth Warren and all the socialists are harping and larking about are not going to be the case anymore... AI becomes a power for good for creating an open economy. — David Friedberg
> [24:03](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1443s) I've never seen it in my 25 years in Silicon Valley where a sector of the economy can absorb hundreds and hundreds of billions of dollars... and economic pricing power effectively evaporates in months. — Chamath Palihapitiya
> [39:28](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2368s) You take things that are in the public domain, you fork them, you make them your own... An American company has taken open-source contributions in the public domain, made it their own, and then developed their own model. And if you say that American companies can't do that or that somehow it's tainted with IP theft, you are basically going to put a dagger through the heart of the entire American open source ecosystem. — David Sacks
> [50:47](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=3047s) They believe it is fair use to train their models and derive their own weights based on fair use. However, they say that the one type of content that you should never be able to train on is *their* output... IP for we, but not for thee. — David Sacks
> [01:02:50](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=3770s) If you go to Washington and you lay down with the dogs, don't be surprised if you wake up with the fleas... they may have poked the tiger. — David Sacks
> [01:09:32](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4172s) When you are a machine and a business model that compounds money at 32% over a 20-year average, you give these guys the benefit of the doubt. These are not people that are flying fast and loose. — Chamath Palihapitiya
> [01:17:36](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4656s) emboldening us so that we no longer tolerate the violence of evictions as a matter of business as usual. — Activist
> [01:18:57](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4737s) the moment the idea is admitted into society, that property is not as sacred as the laws of God... anarchy and tyranny commence. — John Quincy Adams
## Mentioned
Moonshot AI, Kimi K3, Opus 4.8, GPT-5.6, DeepSeek, Axios, Michael Kratsios, Anthropic, Wired, Howard Lutnick, Secretary of Commerce, Polymarket, David Sacks, Donald Trump, Chamath Palihapitiya, OpenAI, The New York Times, Coca-Cola, David Friedberg, Google, Yahoo, Microsoft, Mozilla Foundation, Firefox, Apache Software Foundation, Netscape, AOL, CompuServe, Bernie Sanders, Elizabeth Warren, Jason Calacanis, OpenRouter, AWS, Tesla, xAI, Ben Thompson, Stratechery, Sam Altman, GPT-6, Codex, Grok, GLM-4.5, Claude, Perplexity Computer, Mira Murati, Thinking Machines, Kimi K2.5, Cursor, Composer 2, Library Genesis, LibGen, Thomson Reuters, Ross Intelligence, WestLaw, Y Combinator, Garry Tan, Ferrari on My Wrist, The Fight to Save America, A Sexual Italian Summer, Google Cloud, SpaceX, Apple, Tenement Museum, Democratic Socialists of America, Upper East Side, Marin County, John Quincy Adams, Austin, Airbnb, All-In Summit 2026, Grimes, Diplo
The All-In hosts argue that aggressive lobbying against open-source AI models represents regulatory capture rather than national security prudence, examining how model commoditization, copyright liability, and infrastructure economics are reshaping the technology sector alongside broader debates on property rights.
Open-Source Disruption and the AI Distillation Debate (0:00)
The emergence of Moonshot AI's Kimi K3—delivering performance comparable to closed models like Opus 4.8 and GPT-5.6 at half the cost—has sparked intense debate in Washington. While Axios reported the White House is weighing bans on Chinese open-source models over distillation concerns, Wired indicated Commerce Secretary Howard Lutnick favors incentivizing domestic open-source alternatives. David Sacks warns that prohibiting American developers from using publicly available weights would severely undermine domestic competitiveness, pointing out that domestic projects like Cursor's Composer 2 and Mira Murati's Thinking Machines have successfully built upon open weights.
Frontier Lab Growth Versus Hyperscaler Infrastructure Moats (24:03)
Chamath Palihapitiya contends that base foundation models face rapid commoditization, noting that open-source models now capture more than 50% of recorded OpenRouter token volume alongside extensive unmeasured on-premise 'dark tokens.' While Sacks counters that proprietary labs maintain momentum—with Anthropic projecting $75 billion in exit annual recurring revenue (ARR) and OpenAI revising projections toward $60 billion ARR ahead of GPT-6—the panel agrees that durable value is shifting to hyperscale cloud providers. Chamath and David Friedberg defend Alphabet's projected $195 billion to $205 billion capital expenditure, arguing that enterprise-grade 'five nines' reliability creates an enduring moat backed by Google's historic 32% return on invested capital.
Anthropic's $1.5B Settlement and Intellectual Property Tensions (48:25)
The panel examines Anthropic's historic $1.5 billion copyright settlement, which awards approximately $3,000 per book across roughly 500,000 titles downloaded from shadow-library sites like Library Genesis, alongside $101 million in legal fees. Sacks characterizes Anthropic's public complaints regarding Chinese 'distillation attacks' as corporate hypocrisy, noting that frontier labs rely on expansive fair-use doctrines to train on copyrighted public web data while seeking to prevent others from learning from their own model outputs. Furthermore, Sacks highlights that framing distillation as illicit intellectual property theft backfired across Silicon Valley, prompting an opposition letter signed by Y Combinator's Garry Tan and over 200 startup founders.
Private Property Rights and Market Distortion in Housing (77:36)
Turning to local governance, the hosts critique recent New York City tenant advocacy proposals that label evictions as violence and seek rent freezes alongside prohibitions on concurrent income and credit vetting. Friedberg mounts a philosophical defense of private property, arguing that eroding ownership protections degrades civil society and disincentivizes maintenance. Palihapitiya and Jason Calacanis emphasize that restricting landlord screening raises operational risk, leading to higher upfront rent demands and exacerbating the city's estimated 50,000 vacant 'ghost apartments,' whereas permissive permitting in markets like Austin systematically lowers rents by expanding supply.
The discussion highlights an unresolved tension across both technology and municipal governance: whether attempts to insulate incumbents through regulatory interventions—whether via AI import bans or restrictive housing controls—inevitably distort market incentives and impose greater long-term economic costs.
## Background
本期节目由《All-In Podcast》官方频道于 2026 年 7 月 27 日发布在 YouTube 平台。节目简介显示，四位主持人深入探讨了月之暗面 Kimi K3 震撼发布引发的开源 AI 存亡博弈、Anthropic 与 OpenAI 的监管俘获策略、Anthropic 达成的 15 亿美元盗版侵权和解案与知识产权双标、谷歌与特斯拉因 AI 资本支出飙升引发的股价暴跌，以及纽约极左翼“驱逐房客等于暴力”提案对私有财产权的冲击。
## TL;DR
围绕中国开源大模型 Kimi K3 追平美顶尖闭源架构所引发的白宫博弈，主持人们痛批 Anthropic 等巨头借“蒸馏威胁”之名实施监管俘获以保卫其虚高估值，并揭露其一边刚支付 15 亿美元图书版权和解金、一边反咬开源盗窃的极端伪善。讨论指出基础大模型正在极速大宗商品化，强行封禁开源无异于向全美企业强征“Token 税”并自断经脉，产业超额利润正加速流向云端基础设施与上层垂直应用。针对谷歌和特斯拉因巨额 AI 算力开支导致自由现金流转负与股价下挫，嘉宾们从投入资本回报率（ROIC）与算力底盘角度深度论证了长期看多谷歌的逻辑。最后在“社会主义专栏”中，嘉宾们从立国基石与第一性原理出发，系统批判了纽约极左翼禁止驱逐租客与冻结租金的激进提案，痛斥其违背基本经济学常理并直接制造了 5 万套“幽灵公寓”。
## Atmosphere & style
整期节目散发着兼具精英洞察与肆意张狂的典型硅谷内幕会谈质感，四位 Besties 在嬉笑怒骂、荒诞“自卖自夸”与火药味十足的激烈争吵（如 Jason 的拍桌咆哮）之间切换自如。对话从商业 ARR 曲线与暗代币流量的技术性拆解，一路升华至比特经济对决分子经济、以及弗里德伯格援引美国建国先贤捍卫私有产权的历史檄文，信息密度极高且情绪饱满。听众能真切感受到顶级风险投资人与政商操盘手在面对行业巨变时那种既亢奋又警惕、在冷酷算计与原则坚守之间来回拉扯的真实张力。
## Bulletin
- [00:00](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=0s) 月之暗面发布开源模型 Kimi K3，性能直逼 Claude Opus 4.8 与 GPT-5.6 且推理成本降低近半，引发白宫关于是否封禁中国开源大模型的激烈政策博弈。
- [48:25](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2905s) Anthropic 与创作者达成 15 亿美元的史上最大版权和解案，但其指责外部蒸馏属于‘IP 盗窃’的做法被批为自相矛盾，直接削弱自身‘合理使用’辩护立场。
- [24:03](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1443s) 基础大模型商品化速度极快，非商业 API 的‘暗代币’使用占比突破 50%，AI 产业链核心价值正加速流向底层算力基建与顶层垂直应用。
- [01:07:07](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4027s) Google 2026 年资本支出预计高达 2000 亿美元导致自由现金流首次转负并引发股价下挫，但其自研芯片、云生态和 32% 的长期 ROIC 被视作底层算力护城河。
- [01:17:19](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4639s) 纽约极左翼推动将‘驱逐租客’定性为暴力的激进租管提案，引发嘉宾集体炮轰，指出该政策摧毁私有产权并已造成 5 万套‘幽灵公寓’空置反噬市场。
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=0s) Kimi K3 的突破性表现展现了开源权重追平顶尖闭源模型的能力，推动预测市场上美国封禁开源模型的概率短时间内飙升至 45%。
- [02:00](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=120s) Sacks 认为闭源巨头正在进行‘监管俘获’，若真要杜绝模型蒸馏应强化自身 API 的身份验证（KYC），而非游说政府封禁全美开源生态。
- [10:18](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=618s) Chamath 警告若美国政府通过行政手段封杀开源模型，将迫使本土企业承受比海外高出 50 至 100 倍的‘Token 税’，最终重创美国股市估值。
- [15:12](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=912s) Friedberg 指出模型蒸馏本质上是工程基准对标（Benchmarking），犹如早期 Google 分析 Yahoo 和微软的搜索结果，属于对输出的分析而非盗窃底层私有代码。
- [17:12](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1032s) Sacks 明确区分了‘模型权重’与‘模型输出’，指出巨头抓取全网创作者数据抗辩为‘合理使用’，却对他人使用其输出反咬为‘IP 盗窃’，存在极端双标。
- [20:10](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1210s) 开源模型在物理上表现为可独立下载运行的代码和参数包，对其下达禁令不仅侵犯言论自由且在技术上无法执行，历史上的 Apache 与浏览器开源正是互联网繁荣的根基。
- [24:03](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1443s) 资本市场对基础大语言模型的定价权正在以月为单位被击穿，开源与本地部署的‘暗代币’份额已超过 50%，基础模型将彻底沦为通用大宗商品。
- [30:56](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1856s) 尽管 Anthropic 与 OpenAI 面临开源冲击，但其年化经常性收入（ARR）分别突破 700 亿美元与 400 亿美元，当前高喊生存危机更像是在 IPO 前向政府‘假摔碰瓷’索要双寡头保护。
- [39:31](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2371s) 美国本土顶尖团队（如 Thinking Machines、Cursor）高度依赖公有领域的开源底座进行二次创新，封杀外部开源权重等同于扼杀美国本土的中小创新活力。
- [41:36](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2496s) Friedberg 提出全球长局推演：随着 AI 将‘比特经济’（知识与软件服务）彻底商品化，全球地缘竞争的终极壁垒将转移至由廉价电力与制造产能主导的‘分子经济’。
- [48:25](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2905s) Anthropic 达成 15 亿美元版权和解，源于其直接从盗版电子书库抓取 700 万本图书的硬伤，创作者与媒体正成为 AI 训练数据博弈中的重要收税者。
- [01:00:50](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=3650s) Anthropic 公开将蒸馏定义为‘知识产权盗窃’是一场严重的公关战略失误，按此逻辑直接承认了其自身利用全网内容抓取训练的产品本身也涉嫌侵权。
- [01:09:32](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4172s) Google 尽管因年内 2000 亿美元的 AI 资本开支引发单季自由现金流转负与股价回调，但其 32% 的历史 ROIC 与中立支持 500+ 模型的 GCP 云底座构成了极其坚固的长期复利飞轮。
- [01:14:43](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4483s) 构建‘五个九’（99.999%）超高可用性的云基础设施需要数千亿美元的持续投入，这是苹果未能涉足底层云计算而选择巨额股票回购的核心分野。
- [01:19:03](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4743s) 纽约极左翼推动将合法驱逐房客定义为‘暴力’，是对私有财产权基石的致命侵害，破坏产权秩序最终将导致无政府混乱与极权专制。
- [01:23:25](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=5005s) 禁止驱逐违约租客并限制背景审查不仅使守法弱势租户承受恶邻与治安恶化，更迫使房东采取极端防御，直接催生了纽约 5 万套‘幽灵公寓’的主动空置。
- [01:26:03](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=5163s) 解决住房危机的第一性原理在于放宽审批、增加市场供给，奥斯汀模式与阿根廷废除租管的经验证明，唯有供给增加才能真正降低租金，行政干预只会导致市场全面扭曲。
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=0s) Kimi K3 引发行业震动与白宫政策博弈
节目开场，主持人 Jason Calacanis 宣布这是《All-In Podcast》的第 282 期，并照例调侃这是“播客创作者最爱的播客，也是你妈妈最爱的播客”。本周全球科技圈头号重磅事件正是中国月之暗面（Moonshot AI）发布的开源模型 Kimi K3。这一发布不仅迅速点燃了美国是否应当封禁中国开源大模型的激烈争论，甚至直接惊动了白宫。
回顾背景，Kimi K3 展现出的性能不再落后前沿闭源模型 6 个月或 12 个月，而是直接与 Claude Opus 4.8 以及 GPT-5.6 平起平坐，且推理成本直接便宜了近 50%。这一震撼表现引发了类似 2025 年初“DeepSeek 时刻”的全行业恐慌。白宫随即介入，节目老友迈克尔·克拉齐奥斯（Michael Kratsios）公开表示：“我们掌握的信息显示，月之暗面在开发 K3 模型时蒸馏（distilled）了 Anthropic 的模型。”随后，围绕特朗普政府的应对举措传言四起：Axios 周一爆料称白宫正在考虑出台禁令封杀中国开源模型；而 Wired 周三又报道称商务部相关高层（如 Howard Lutnick 等）并不赞成封禁，反而倾向于激励美国前沿实验室打造更优秀的开源模型。预测市场 Polymarket 上关于“美国政府是否会在 2026 年封禁开源大模型”的概率在短短数天内从 22% 飙升至 45%。
- **性能与成本突破**：Kimi K3 追平 Opus 4.8 与 GPT-5.6 的能力梯队，价格却便宜约 50%。
- **白宫内部的分歧信号**：一边是 Axios 报道的潜在禁令考量，另一边是支持者主张通过反向激励美国本土开源研发来对抗，而非一刀切行政封杀。
- **预测市场波动**：Polymarket 平台关于 2026 年美国政府下达开源模型禁令的押注概率迅速升至 45%。
---
### [02:00](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=120s) 萨克斯直指监管俘获与“倒打一耙”的禁令逻辑
针对局势的发展，David Sacks 开启了本期节目的第一次“胜利巡游”（victory lap）。早在两个月前他就公开预警：市场上鼓吹“大模型必须有护栏”、“开源模型移除了安全护栏因而极度危险”的论调，本质上是闭源头部巨头为了推动封禁开源模型而预埋的事实与舆论铺垫，这在 Anthropic 的官方博客中屡见不鲜。
当被问及白宫目前的真实态度时，Sacks 透露据可靠消息来源，白宫根本没有做出封杀开源模型的最终决定。他指出特朗普总统向来善于倾听各方声音，在科技政策上一直倾向于轻度监管和开放竞争，这也是美国在 AI 竞赛中保持领先的原因。Sacks 严厉批评了“禁止美国开发者使用中国贡献到公共领域的开源模型”的主张，称这是极其荒唐且有害的自残行为（"cutting off our nose to spite our face"）。如果公共领域有成果而全世界都在用，单单绑住美国开发者的手脚只会葬送美国的竞争优势。
Sacks 进一步点名 Anthropic，指责其正在实施经典的“监管俘获”（regulatory capture）。Anthropic 年初的 ARR（年度经常性收入）还在 100 亿美元左右，如今已飙升突破 700 亿美元，堪称人类科技史上规模化扩张最快的公司。一家根本不缺增长、拥有 90% 超高毛利的巨头，却四处制造恐慌寻求政府庇护。Sacks 揭穿了这其中的逻辑漏洞：如果工业级规模的“蒸馏”真的是关乎国家安全的威胁，Anthropic 应该做的是封禁来自外部对其 API 的非法访问，在源头上掐断蒸馏；然而他们为了追求狂飙突进的营收增速，迟迟不肯对客户强制推行严格的身份验证（KYC），任由批量注册的账号在暗网倒卖。Sacks 尖锐地总结道：“这个辩论完全搞反了，本该是 Anthropic 交代为什么防盗做得如此糟糕，现在却要全美开源生态来为 Anthropic 的失职买单。”
```mermaid
flowchart TD
accTitle: 模型蒸馏与封禁争议逻辑链
accDescr: 对比Anthropic呼吁封禁外部模型与Sacks提出的源头治理方案差异
subgraph Problem["灰色蒸馏产业链"]
A["东南亚等地批量注册账号"] --> B["暗网打包倒卖并伪装美国IP"]
B --> C["高并发调用API提取万亿级问答对"]
end
subgraph AnthropicArg["Anthropic 的诉求逻辑"]
C --> D["鼓吹国家安全与护栏失效"]
D --> E["游说政府封禁中国开源模型"]
E --> F["将开源模型排除在美国生态之外"]
end
subgraph SacksCritique["Sacks 与 Chamath 的反驳"]
C --> G["源头治理：实施强制 KYC 认证"]
G --> H["牺牲少量增速换取真实接口防御"]
H --> I["允许美国开发者自由利用公共领域开源成果"]
end
```
---
### [07:35](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=455s) 蒸馏的本质：全行业“蜘蛛侠互指”与模型商品化加速
谈话随后转向企业级市场的现实落地与“蒸馏”的技术本质。Jason 询问 Chamath 在其企业级 AI 软件公司 8090 的销售沟通中听到了什么反馈，企业客户是否迫切希望将这些强大的开源模型部署到本地（on-prem），并请他科普究竟什么是模型蒸馏。
Chamath 直白地解释道，蒸馏的本质非常简单：启动一个高水平的目标模型，向它提出海量问题，记录并观察它的全部输出，然后将这些高质量问答对作为训练数据来训练你自己的模型。当这种操作被成百上千万倍地放大时，本质上就是从原模型中秘密抽取了数以万亿计的问答对。Chamath 赞同 Sacks 的观点，如果厂商真想杜绝蒸馏，完全可以通过绑定实体身份信息、带额度限制的信用卡等 KYC 手段进行防御，但这样做必然会拖慢用户获取与营收增长的步伐。
更为讽刺的是，当前的版权与蒸馏争议就像那个著名的“九个蜘蛛侠互相指责”的表情包：Anthropic 刚刚因为侵权抓取图书数据支付了 15 亿美元的和解金；OpenAI 正面临《纽约时报》的大规模版权诉讼；而中国的前沿实验室又反过来蒸馏 Anthropic 的模型。全行业其实都在彼此借鉴与吸收，没有人真正清白。Chamath 强调，在这场纷争背后，真正值得关注的核心洞察在于：基础大模型的商品化（commoditization）速度远超所有人的预见。一旦任何一家前沿模型公开了最新的性能评测标准，短短几周之内，市面上就会涌现出性能完全抹平差距的其他开源或闭源模型，所谓的持续领先壁垒正在被迅速击穿。
### [10:18](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=618s) 闭源大模型的估值保卫战与“毁灭股市”的监管干预
Chamath Palihapitiya 深入剖析了闭源前沿实验室（如 OpenAI、Anthropic）当前的真实焦虑。他指出，市场上已经涌现出一批开源权重模型，其表现不仅能与闭源模型并驾齐驱，甚至在某些场景下实现了超越。少数几家美国头部 AI 公司猛然意识到：当前赖以生存的巨额价值与高估值，在未来五到十年的周期内根本不可持续。然而，当这些公司向华尔街推销商业模式时，必须展示极高的确定性，否则其估值泡沫就会瞬间破灭。因此，当前游说政府限制甚至封杀开源的做法，本质上是一场“闭源实验室的估值保卫战”（a valuation preservation game）。
Chamath 指出，底层基础模型正在以极快的速度走向大宗商品化（commoditized）。一旦看透这一点，就会明白未来的核心商业价值根本不再属于基础模型层，而是被两端截流：上层的“应用层”（Application layer）以及底层的“算力与基础设施层”（Chips & Cloud infrastructure）。如果没有政府横加干预、把手强行压在天平上，资本自然会流向应用层与基建层。这对于那些定价高出开源替代方案 25 到 50 倍的闭源实验室而言，无疑是致命打击。因此，他们才试图动用政治力量，消灭同等质量却便宜得多的竞争对手。
```mermaid
flowchart TD
accTitle: AI价值链的重构与闭源模型的困境
accDescr: 展示基础模型商品化后，价值向上层应用和底层基建转移，闭源实验室通过监管游说保卫估值的逻辑。
subgraph Value_Shift["AI 产业价值转移趋势"]
App["上层：应用层（Application Layer）<br/>捕获高附加值与工作流"]
Foundation["中层：基础模型（Foundation Models）<br/>快速大宗商品化 / 价格战"]
Infra["底层：算力与基础设施（Chips & Cloud）<br/>英伟达、云计算厂商等垄断核心资源"]
end
subgraph Closed_Labs["闭源前沿实验室（OpenAI / Anthropic）"]
Risk["高昂 Token 定价（溢价 25x-50x）<br/>面临开源权重激烈冲击"]
Lobby["游说政府：以国家安全/蒸馏为由封杀开源"]
end
Foundation -.->|"利润被上下挤压"| Risk
Risk -->|"估值保卫战"| Lobby
Infra -->|"承载算力"| Foundation
Foundation -->|"赋能应用"| App
```
Chamath 进一步发出严厉警告：如果美国政府真的介入并对开源模型下达禁令，最终结果将是“直接砸穿美国股市”（it will tank the stock market）。他以可口可乐（Coca-Cola）为例：如果美国本土企业在尝试用 AI 提效降本时，被法律强制剥夺开源选择权，只能采购由 OpenAI 和 Anthropic 构成的双寡头闭源服务，其 AI 采购成本将比海外竞争对手高出 50 到 100 倍。当 AI 成为未来商业运转的核心要素时，美国企业将被迫承担非理性、非市场化的畸高成本结构，资本市场必然会对这些美国巨头进行全面估值下调（rerated）。与此同时，闭源实验室虽然短期内依靠监管庇护获得收入，但由于这种收入全靠人为扭曲的单一市场支撑，其商业逻辑同样脆弱不堪，“所有道路都将通往市场混乱”。
### [13:38](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=818s) “AI 孤岛”上的 Token 赋税与全球地缘博弈
Jason Calacanis 顺着 Chamath 的逻辑总结道，政府若为 OpenAI 和 Anthropic 建立法定双寡头垄断，无异于向所有美国企业强行征收一笔沉重的“Token 税”（token tax）。当世界其他地方都在自由享受低成本、高性能的开源模型时，美国企业反而将自己封锁在一座“极度昂贵的 AI 孤岛”上。David Sacks 随即讽刺道，这就像可口可乐在海外只卖 50 美分，但在美国国内却因为人为壁垒被迫卖 50 美元一罐，“既然能买 50 美分的，为什么非要买 50 美元一罐的可乐？”
Sacks 进一步强调，Anthropic 等公司提出的封杀开源提案根本无法解决所谓的“模型蒸馏”（distillation）问题。如果蒸馏真是个威胁，就必须在源头上解决；仅仅在美国本土封禁开源，绝不可能阻止世界其他地区继续采用并运行中国的开源模型。更致命的后果在于，海外市场将持续通过海量用户交互积累数据、开展强化学习（RL），从而让非美模型加速迭代进化，最终结果将是“百分之百把美国推向输掉 AI 军备竞赛的深渊”。
随后，Jason 将话题引向地缘政治对抗的宏观层面，援引格雷厄姆·艾利森（Graham Allison）的修昔底德陷阱视角提问：全面封杀中国开源模型无疑是一记极具挑衅意味的狠招，北京和中国最高决策层将如何反制？这是否意味着中美两国在科技天梯上的对抗将全面升级？
### [15:12](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=912s) Friedberg 解构“蒸馏”本质：从谷歌搜索对标到工程反向工程
David Friedberg 接过话茬，表示在这一点上他与 Sacks 和 Chamath 高度一致。Friedberg 认为，开源模型是否由中国团队发布并不是问题的核心，潜在的安全风险完全可以通过技术评估与工程手段来隔离化解。核心争议应当聚焦于对“蒸馏”（distillation）的理解。Friedberg 明确指出，蒸馏并非 AI 领域的专属黑魔法，它在本质上就是工业界极其普遍的工程分析与基准对标（benchmarking）。
Friedberg 详细列举了跨行业的真实案例：
- **汽车制造业**：一家车企拆解并深入研究竞争对手汽车的各项运行表现与机械设计，以此作为设计出更好产品的参考，这是完全合法且通行的做法。
- **早期谷歌的搜索优化**：在谷歌创业早期，团队曾向雅虎（Yahoo）和微软（Microsoft）的搜索引擎发送数以百万计的搜索请求（queries），系统性记录并分析对方给出的搜索结果集合，然后将谷歌的搜索结果与对方横向比对，从而调整与精进谷歌自身的排名算法。“这并不意味着我们在偷窃他们的算法，我们没有黑进他们的服务器盗取软件代码，我们只是分析了他们软件输出的结果，并用它来改进我们自己的软件。”
在知识产权（IP）与版权法的框架下，Friedberg 强调必须严格区分“产出物”（output）与“工程过程”（process）。模型蒸馏看重的是观察外部输出，而非直接窃取底层私有软件。服务提供商确实可以在服务条款（Terms of Service, ToS）中约定禁止此类行为，但这属于合同与条款履约层面的商业博弈，绝非传统法理意义上的侵犯版权或窃取商业机密。
### [17:12](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1032s) 权重与输出的天壤之别：硅谷巨头的双标与 Sacks 的鹰派自白
David Sacks 随即插话补充，指出当前政策制定者与科技行业之间存在巨大认知鸿沟：政策圈完全没搞懂“模型权重”（model weights）与“模型输出”（model outputs）之间不可逾越的技术边界。
- **模型权重（Weights）**：是包含海量数值参数的权重文件，它构成了大模型的核心软件与代码本体。如果中国企业黑进 OpenAI 或 Anthropic 的服务器窃取专有权重，那是毫无争议的盗窃行为（theft）。
- **模型输出（Outputs）**：仅仅是模型运行后给出的文本与数据结果。从输出中总结规律、获取灵感并进行学习，是现代科技研发的标准工作方式。
Sacks 尖锐批评了 OpenAI 和 Anthropic 在该议题上的极端伪善：两家公司始终坚称自己有权在不经授权的情况下，抓取全人类公开创作的所有内容来训练模型，不论原创作者是否同意。例如，《纽约时报》（The New York Times）起诉 OpenAI 的核心指控，就是 OpenAI 违反了纽时的服务条款，暴力爬取其全站新闻报道并用作训练数据。面对诉讼，OpenAI 的抗辩逻辑正是：“我们没有偷任何东西，我们只是获取了《纽约时报》的文本输出，并据此推导出了属于我们自己的模型权重。”然而，当中国模型通过 API 观察美国模型的文本输出并推导出自身权重时，OpenAI 和 Anthropic 却立即反咬对方是“窃取知识产权”。
针对白宫和部分舆论指责的“以工业化规模暗中违反服务条款（ToS）存在严重道德问题”，Sacks 澄清了自己的政治立场与动机。他强调自己绝非在为中国辩护，并回顾了自己的强硬鹰派履历（bonafides）：
- 2025 年 1 月 DeepSeek 异军突起时，Sacks 作为政府高级官员（白宫人工智能与加密货币事务负责人）率先登上劳拉·英格拉汉姆（Laura Ingraham）的福克斯新闻节目，成为全美首位公开向公众深入拆解“蒸馏”机制的高官；
- 他更是白宫《赢得人工智能竞赛》（Winning the AI Race）战略报告的核心联合作者，该报告的核心主旨就是动用一切力量击败中国、确保美国在 AI 领域取得绝对胜利。
Sacks 直言，正是因为自己坚定希望美国获胜，才坚决反对闭源巨头以国家安全为借口绑架产业政策；靠行政手段封杀开源不仅无法遏制中国，反而会直接扼杀美国本土的创新活力。
### [19:37](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1177s) 禁止开源是自断经脉：企业应自行强化服务条款而非依赖政府封禁
嘉宾们首先聚焦于美国如何在这场全球 AI 竞赛中取胜的核心路径。讨论指出，美国要想赢得 AI 竞争，关键在于策略与手段。如果在国内草率禁止开源模型，剥夺所有美国企业利用开源技术的机会，而全世界其他国家却依然可以自由享用，这无异于美国“自断经脉”（shoot ourselves in the foot），将酿成巨大隐患。
对于 Anthropic 和 OpenAI 等闭源巨头抱怨中国模型通过“模型蒸馏”（distillation）抄袭其技术并侵犯其服务条款（TOS）的说法，现场观点十分明确：如果涉及违规，Anthropic 和 OpenAI 应当依靠自身的技术力量去严格执行其服务协议，从技术源头阻止蒸馏行为的发生，而不是一味向政府求援要求一刀切封禁。政府如果能提供合法合规的协助固然可以，但究竟能做什么并不明晰，真正的责任依然在这些大模型公司自身的商业与安全风控能力上。
### [20:10](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1210s) 弗里德伯格的互联网历史课：从 Netscape、Apache 到 AI 财富的普及
大卫·弗里德伯格（David Friedberg）随后系统阐述了他反对封禁开源的另外两大核心论点：言论自由与历史经济规律。弗里德伯格向非软件行业出身的听众普及了开源软件的本质：开源模型在物理层面上本质就是一套可下载的代码和参数包，形式上就像下载一本电子书或一份文本。一旦下载到本地电脑，用户无需联网即可独立运行。正因如此，政府试图限制开源将打开一个可怕的“潘多拉魔盒”——执法部门根本无法有效执行禁令，因为这意味着要去剥夺民众运行已经保存在本地磁盘上的公开免费软件的权利，这在法律程序与宪法言论自由层面缺乏任何可靠先例，实施起来将极为丑陋。如果确实涉及知识产权或版权侵权，完全可以通过正当的法律程序起诉并禁止其传播，或通过精准的贸易制裁去惩罚侵权方，而非全盘扼杀开源机制。
弗里德伯格进一步通过互联网早期的商业历史来说明开源对人类经济的不可替代性。在互联网黎明期，网景（Netscape）开发了专有的商业浏览器和专有的 Web 服务器软件，凭借先发优势成功上市并赚取了暴利。然而，这一专有壁垒迅速被开源运动打破：Mozilla 基金会成立并推出了开源浏览器 Firefox（后来 Google 招募了相关核心人才打造出 Chromium 和 Chrome，同样坚持开源）；Apache 软件基金会推出了首个开源 HTTP 服务器软件。任何拥有普通电脑的人都能免费下载 Apache 搭建 Web 服务器，无需向 Netscape、微软或甲骨文（Oracle）缴纳昂贵的门票。这一变革使得互联网的经济红利没有被少数控制准入门户的软件供应商垄断，而是外溢并孕育出了 Google、eBay、Etsy、亚马逊以及全球数以百万计的中小网站与企业。
```mermaid
flowchart TD
accTitle: 专有垄断与开源生态的历史路径演进
accDescr: 对比早年专有软件垄断的围墙花园模式与开源驱动的广泛价值繁荣生态
subgraph ClosedModel ["专有封闭模式（围墙花园）"]
A["专有底层软件<br/>Netscape / 闭源巨头"] --> B["少数垄断企业控制门户<br/>AOL / CompuServe 式围墙"]
B --> C["巨额利润集中于少数寡头<br/>高额准入门槛阻碍创新"]
end
subgraph OpenModel ["开源普及模式（生态外溢）"]
D["开源基础设施<br/>Apache / Linux / 开源 AI"] --> E["算力与模型成为泛在公共品<br/>零边际成本自由接入"]
E --> F["数百万创新企业与应用崛起<br/>经济红利全社会扩散"]
end
```
弗里德伯格强调，如果当年的互联网是封闭的、处处设立专有收费关卡，那么整个产业就会退化成 AOL 或 CompuServe 的局域网翻版，现代互联网经济根本不可能诞生。将这一历史映射到 AI 领域：一旦开源 AI 全面爆发，伯尼·桑德斯（Bernie Sanders）或伊丽莎白·沃伦（Elizabeth Warren）等左翼政客整天担忧的“AI 财富被两三家科技寡头及其背后的亿万富翁股东搜刮殆尽”的噩梦就不复存在。相反，AI 技术将扩散至全球数以百万计的实体企业，AI 将成为促进经济开放、繁荣与创造就业的普惠力量。“那些想控制专有技术收租的 1% 寡头们，他们的好日子到头了，我们甚至根本不需要设立所谓的财富税。”
### [24:03](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1443s) 资本回报神话破灭：数百亿投入与几个月内被蒸发的定价权
查马斯·帕里哈皮蒂亚（Chamath Palihapitiya）随后切入资本市场的残酷现实，直言当前基础模型领域的价值捕获正在以史无前例的速度归零。查马斯回顾道，在他混迹硅谷的 25 年职业生涯中，从未见过任何一个行业在吸纳了数百上千亿美元的资本开支后，原本号称拥有跨越数十年的超强产品定价权，却在短短几个月内被彻底击穿并“蒸发殆尽”。虽然现场有人认为用“蒸发”来形容可能略显极端、或许只是增长放缓或进入平台期，但资本市场对定价权的重新校准已是不争的事实。
为了佐证这一现象，节目展示了两份关键图表：
- **Anthropic 营收增速停滞**：第三方监测数据显示，在经历前期的指数级爆发后，Anthropic 的 ARR（年度经常性收入）曲线在最近两个月出现明显的横盘停滞迹象，面临巨大的增长逆风。
- **“暗代币”（Dark Tokens）吞噬市场份额**：OpenRouter 等路由平台的追踪数据表明，随着以中国模型为代表的开源基础模型性能日益逼近顶尖水准，且在本地服务器私有化部署变得极其简便，越来越多企业选择自建或私有托管。这些不经过公有商业 API 计费的推理用量被称为“暗代币”。在总体推理代币的流向中，这类完全不需要向闭源大厂纳贡的免费开源模型占比已迅速突破 50%，导致商业闭源 API 的收入天花板被剧烈压低。
### [27:37](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1657s) 智能即泛在商品：暗代币崛起与价值向云端基础设施转移
查马斯从专业资金分配者（Capital Allocator）的视角进一步拆解了市场的本质：金融市场极具前瞻性，投资者看穿当下的账面营收，只追问一个底层问题——这项业务在 10 年后会怎样？是收入持续攀升还是下滑？竞争是更剧烈还是形成垄断？它到底是一门独占生意，还是一种大宗商品（Commodity）？如果是大宗商品，最终的市场出清价格（Market clearing price）是多少？
在传统商业周期中，一家企业从专有排他性垄断走向大宗商品化，通常需要经历 5 到 10 年的漫长变迁，给资本留出充足的消化周期。然而在 AI 软件与开源技术的极致压缩下，这个周期被缩短到了以月计。当每一个精明的投资者都意识到基础智能模型将在 5 到 10 年内彻底变成随处可见的大宗商品时，就再也没有人愿意为其未来的远期现金流支付天价溢价。
那么，AI 爆发带来的万亿利润究竟流向了何方？查马斯指出，真正的经济红利正在流入云基础设施服务商（Hyperscalers，如 Google Cloud、AWS）。从各大科技巨头的财报可以看出，云厂商甚至更乐见开源模型的繁荣，因为只要开发者在它们的云端运行最廉价、最高效的开源模型，云平台就能稳稳赚取底层算力、电力与存储的高额基础利润。
讨论最后达成了共识类比：基础大语言模型就像当年的 Web 浏览器，它只是通往数字目的地的载体与基础设施，本身没有持久的专有溢价能力；真正的商业价值将沉淀在两个极端——底层的算力与云服务（如 AWS、Google Cloud 以及马斯克的算力集群），以及上层直接解决用户业务痛点、建立深度场景连接的应用端。随着即时智能的供给变得无穷无尽、边际成本与推理能耗趋近于零，基础模型本身的收费管道化已成定局。主持人们随即准备切入后续环节，由大卫·萨克斯（David Sacks）作全面辩驳，并深入探讨知识产权诉讼及科技巨头的资本开支隐忧。
### [28:48](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1728s) 对立论点：开源与闭源的双赢格局与“恐慌”破除
在节目中充满“激烈共识”之后，David Sacks 提出了不同的反向论点。他认为，开源与闭源最终都将成为市场的大赢家，因为整个 AI 市场足够庞大，二者各有其存在的价值与用途。近期随着月之暗面 Kimi K3 的推出，业内出现了一阵过度恐慌，大家惊呼“中国模型已经全面追赶上来，并且还免费开源，这会彻底摧毁美国领先的前沿实验室”。Anthropic 和 OpenAI 也顺势四处游说，声称“如果中国公司可以随意蒸馏、窃取模型权重，我们就无法继续砸数十亿美元进行研发”，而政府官员也对此作出了积极响应。
但深入剖析数据后会发现事实并非如此：
- **运行成本并无明显优势**：科技博主 Ben Thompson 在 Stratechery 上拆解了 Kimi K3 的成本细节，证实其推理运行成本并没有想象中那么低，不存在显著的成本红利。
- **中国模型并未在全维度赶超**：虽然 Kimi K3 在前端编程与 Web 开发的 Arena 评测排行榜上表现亮眼，但那仅代表单个特定维度；在许多其他核心测试中它的表现并不突出，并没有全面超越或追平美国顶尖前沿模型。
- **前沿实验室仍握有技术代差**：Anthropic 和 OpenAI 的实验室内部仍有领先半年的杀手锏未发布。Sam Altman 预计将在接下来的一周前往华盛顿，展示据称“性能炸裂”的 GPT-6（或下代前沿架构）。Sacks 坚信中国尚未真正追赶上来，美国依然保持着约 6 个月的技术身位领先，目前最重要的是“让自家赛马放开跑”，切忌设置不必要的监管障碍自缚手脚。
---
### [30:56](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1856s) 营收飞轮与“假摔碰瓷”监管：前沿双寡头的商业防线
衡量模型在现实应用中真正实力的唯一硬指标就是营业收入，而在这一点上，Anthropic 与 OpenAI 展现出了科技史无前例的极速规模化增长。针对 Jason Calacanis 拿出的所谓“Anthropic 增长出现停滞小瑕疵”的图表，Sacks 进行了有力反驳：Anthropic 年初设定的内部目标是实现 10 倍增长（从 10 亿美元冲刺至 100 亿美元 ARR），如今仅到年中，其 ARR 就已突破 70 亿美元，达成 100 亿美元年化收入轻而易举。而图表上那点微小波动，很大程度是因为 OpenAI 的重新提速——随着 Codex 编程能力的爆发，OpenAI 在过去一个月重新找回势头，Sam Altman 甚至在 X 上发文称“找回了自身节奏（got our mojo back）”，并将年底 ARR 退出率预测从 60 亿美元上调至 75 亿美元左右。
Sacks 指出，将这两家巨头的数据叠加在一起看作美国前沿模型“双寡头”，就会发现增长没有任何放缓迹象。所谓模型蒸馏（Distillation）早在 2025 年 1 月 DeepSeek 崛起时就已被他指出，在这整个过程中，美方前沿实验室的营收依然在指数级暴增。因此，两家巨头高喊中国模型威胁与“无法生存”，本质上是在借题发挥：
- **IPO 路演中的“假摔碰瓷（Flop & Foul-baiting）”**：正如 NBA 赛场上勒布朗·詹姆斯（LeBron James）等球员夸张倒地以骗取裁判判罚，Anthropic 和 OpenAI 正值 IPO 路演阶段，面对华尔街“未来是否会被开源 AI 商品化吞噬利润”的尖锐质询，最完美的解法就是演一出大戏，“假摔”诱导美国政府介入，赋予其受政策保护的双寡头垄断地位。
```mermaid
flowchart TD
accTitle: 前沿模型实验室的监管套利与商业护城河逻辑
accDescr: 展示前沿实验室如何通过夸大开源威胁以促使政府监管介入，并向应用层延伸构建商业护城河
subgraph Defense ["华尔街质询与假摔策略"]
A["华尔街投资者质询：<br/>开源是否会导致模型商品化？"] --> B["夸大外部开源威胁与权重窃取<br/>（NBA式'假摔碰瓷'）"]
B --> C["游说美国政府介入干预<br/>构筑政策保护型双寡头壁垒"]
end
subgraph Commercialization ["商业变现真实路径"]
D["拥有最顶尖的前沿基础大模型"] --> E["向技术栈上方延伸：<br/>打造终端自研应用与垂直解决方案"]
E --> F["特定尖端模型不对外开放 API<br/>专供自研高价值高壁垒业务"]
end
```
---
### [33:24](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2004s) 向上走还是被开源替代？应用层延伸与初创企业的逃离
Chamath Palihapitiya 赞同两家巨头应该给华尔街一个更自信的答案：往技术栈上游走，直接做应用层（Go up the stack to the application layer）。虽然他们过去在某些自研端到端应用上的做法略显笨拙，但其实力过硬。两家巨头完全可以向资本市场宣告：“我们拥有全球最好的模型，未来我们可能会保留某些超强版本不对外发布 API，全部自留用于自研应用，以此捕获生命科学、网络安全等垂直领域的万亿级企业价值。”
这一观点随即引发了 Jason Calacanis 的激烈反驳，引发了全场最激烈的火药味交锋：
- **客户视角的“反竞争与背叛”**：Jason 情绪激动地打断并喊道：“让老子说完这 30 秒，闭嘴听我说！（Shut the f*** up for 30 seconds）”。他指出，一旦基础模型厂商把最强模型留作自用并与下游企业竞争，像 Lovable、ElevenLabs 这样每个季度向其进贡数十万美元 API 费用的初创企业客户就会立刻视其为死敌并彻底出走。
- **初创企业正集体倒向开源**：作为活跃的初创公司投资人，Jason 直言初创生态正掀起向开源迁移的风暴。对于 95% 的实际日常任务而言，根本不需要最新、最昂贵的闭源前沿模型，上一代模型或本地自部署（Local hosting）的开源模型已经完全能够胜任。
- **IPO 估值面临残酷压缩**：过度高昂的算力资本支出（Capex）搭配逐渐被压缩的毛利率，正在将 OpenAI 和 Anthropic 拖入资本泥潭。初创企业的技术选型往往就是未来企业级市场（Enterprise）效仿的先声，这将对两家巨头的 IPO 估值构成巨大阻力。
- **模型体验已无感知差异**：Chamath 补充道，如果商业模式是把模型的基底调用层当作拥有巨大终值（Terminal value）的资产，在数学逻辑上绝对是个严重错误。Jason 现身说法表示，他在日常使用 Perplexity Computer 等测试框架时，将任务同时交给 Grok、Neotron、GLM 等开源/替代模型与 Claude 相比，输出结果并无肉眼可辨的差距，开源替代品已经足以全面平替。
### [37:41](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2261s) 巨额 ARR 增长与闭源巨头的“监管俘获”争议
嘉宾们就 OpenAI 与 Anthropic 的最新营收估算数据展开剖析。根据流出的第三方行业报告，OpenAI 的年化经常性收入（ARR）出现了惊人的再加速，从 5 月份的约 330 亿美元飙升至 7 月份的 413 亿美元。Sam Altman 也在公开场合自信宣称 OpenAI 已经“重拾魔力（got their mojo back）”，即将迎来史上业绩最辉煌的 12 个月。与此同时，Anthropic 的表现同样令人咋舌：如果今年 1 月有人预测 Anthropic 到年中 ARR 会从 100 亿美元突破到 700 亿美元以上，绝不会有人相信——这无疑是科技史上增长最迅猛的公司。
```mermaid
flowchart LR
accTitle: 开源基础模型到美国本土应用的衍生流程
accDescr: 展示开源模型通过微调与私有数据结合，在本地基础设施安全运行的链路
subgraph Upstream["公有领域与开源模型"]
A["开源基础模型<br/>（如 Kimi K2.5 开放权重）"]
end
subgraph AmericanDev["美国本土团队与衍生创新"]
B["蒸馏与二次预训练<br/>（如 Thinking Machines / Cursor）"]
C["注入专有数据<br/>（如 Composer 2 代码数据）"]
D["本地机房运行<br/>（完全掌控 / 零数据外流）"]
end
A --> B
B --> C
C --> D
```
针对两家公司所谓“正面临被开源彻底摧毁”的危机言论，David Sacks 直言现实数据完全不支持这种恐慌。更何况，生成式 AI 的商业壁垒远不止模型本身，还包含外层封装（harness）、企业系统连接器（connectors）以及深厚的企业客户服务协议。Sacks 严厉抨击了头部闭源公司试图游说政府进行监管干预的行为，称这是“科技史上最成功的企业跑去求政府建立护城河，防范未来的潜在竞争”，其手段既难堪又虚伪。
- **OpenAI ARR 激增**：从 5 月的约 330 亿美元攀升至 7 月的 413 亿美元。
- **Anthropic 业绩神话**：ARR 预估从年初的约 100 亿美元膨胀至年中的 700 亿美元以上。
- **闭源护城河构成**：除了核心基础模型权重，还深度依赖工程封装套件、连接器生态与企业级商务合同。
---
### [39:31](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2371s) 捍卫美国开源生态：从 Thinking Machines 到 Cursor
针对给海外开源模型扣上“知识产权盗窃（IP theft）”帽子的舆论风向，Sacks 强调自己并不是在替中国公司辩护，而是在捍卫需要借助公有领域资源的广大美国开发者。他指出，Mira Murati 的新团队 Thinking Machines 研发出了目前全美顶尖的开源模型，而其训练底层正是基于中国开源模型 Kimi K2.5 进行冷启动蒸馏的；代码编辑器 Cursor 发布的全新产品 Composer 2，也是基于 Kimi K2.5 并在团队自有的专有代码数据上进行了后训练（post-training）。
Sacks 剖析了开源软件运作的基本逻辑：获取公有领域中的公开权重文件（open weights），进行 Fork 分支与微调，融入私有数据打造出全新的衍生产品。这些模型文件直接托管在美国本土的数据中心，运行在自有的硬件上，没有数据包回传，完全由本地开发者掌控。如果 Anthropic 试图在缺乏法律证据的情况下将开源权重污名化，甚至通过行政手段予以封杀，无异于“直接将匕首刺入美国开源生态的心脏”，实质目的仅仅是为了扼杀市场竞争。
- **Thinking Machines 案例**：Mira Murati 打造的美国顶尖开源模型，技术路径上借鉴并蒸馏自 Kimi K2.5。
- **Cursor Composer 2 案例**：采用开源底座并结合自有高质量编程数据完成二次后训练。
- **开源运行本质**：开源权重属于公有领域资产，部署于美国本地机房，完全实现数据主权，不涉及任何海外数据外传。
---
### [41:36](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2496s) Friedberg 的全球长局推演：比特经济的商品化与分子经济的对决
David Friedberg 将讨论提升到全球宏观战略与人类文明演进的视角。过去 50 年中，西方世界通过知识经济与服务经济攫取了数万亿美元的财富，本质是“将一个比特转化为另一个比特”的价值提炼；与此同时，西方将制造业大量外包，催生了东方的制造超级大国。在人类文明的底层逻辑中，技术进步与财富繁衍的终极方程是：**以最低的能量消耗实现分子形态的转化**（molecule conversion）——从制造一把低成本沙发，到生产高级材料与先进半导体，实体世界皆由分子转化驱动。
```mermaid
flowchart TD
accTitle: 全球长局战略比较：比特经济 vs 分子经济
accDescr: 展现知识经济商品化后，全球核心竞争力向能源与实体制造转移的逻辑框架
subgraph EconomyDecline["比特经济（知识与服务）"]
E1["开源 AI 普及与零边际成本"]
E2["知识与代码生成彻底商品化"]
end
subgraph Dominance["分子经济（实体制造与能源）"]
M1["电力产能优势<br/>（中国 8 TW 目标 vs 美国 1 TW）"]
M2["厂房制造规模<br/>（中国 2000 亿平方英尺 vs 美国 100 亿）"]
end
E1 --> E2
E2 -->|"价值压缩转移"| Dominance
```
Friedberg 认为，随着开源 AI 的普及，知识经济与智力服务的价值将被全面压缩并彻底商品化。一旦比特经济的利润空间被抹平，全球经济的主导权将不可逆转地交还给掌握“能源与分子转化”的一方：
| 战略维度 | 美国现状 | 中国现状及长期目标 | 产能对比优势 |
|:--- |:--- |:--- |:--- |
| **发电装机容量** | 约 1 太瓦（TW） | 正迈向 8 太瓦（TW） | 中国具备约 8 倍规模潜能 |
| **工业制造物理面积** | 约 100 亿平方英尺 | 约 2000 亿平方英尺 | 中国拥有约 20 倍物理制造空间 |
| **核心战略聚焦点** | 知识产权、软件服务（比特） | 廉价清洁电力、大规模制造（分子） | 形成全球不可逾越的实体工业依附 |
---
### [44:48](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2688s) 利益屁股决定脑袋：投资圈“对华鹰派”的真实小算盘
随着开源模型在企业部署中的技术门槛迅速降低，各类配套连接工具让私有化部署变得轻而易举。主持人与嘉宾们打趣调侃起风投机构对政策立场的剧烈转向。Chamath Palihapitiya 与 Sacks 指出，硅谷当下充斥着“为自己仓位站台（talking their books）”的投资人——此前大家还在不遗余力地向海外出售芯片赚取巨资，如今部分投资了 Anthropic 和 OpenAI 的风险投资人却摇身一变，打着国家安全旗号成为狂热的“对华鹰派”，呼吁政府封杀开源生态以保护自身的投资组合。
Sacks 辛辣地讽刺道，要求封杀开源之前，各路风投大佬最好先公开自己的股权架构表（cap table）。Chamath 用德州扑克形象地比喻当前巨头投资者的焦虑心态：好不容易在翻牌圈抓到了同花或顺子的绝世好牌，此刻唯一的恐惧就是公共牌面千万别出对子（"Please don't pair the board!"），唯恐开源力量破局砸了自己的超级独角兽估值。在哄堂大笑与日常“带货”的玩笑声中，全场一致认同：无论巨头如何游说，开源 AI 所带来的自由度与技术主权浪潮早已势不可挡。
### [47:06](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2826s) 开源与闭源的分工及主持人们的招商狂欢
讨论延续了对开源 AI 价值的肯定。嘉宾们认为，当前绝大多数开源模型已经能够胜任 95% 的日常与业务任务，企业和开发者根本不需要为了争抢所谓的“绝对第一”而陷入焦虑。开源负责解决那通用的 95%，而剩下的 5% 高价值、极度专业化或需要高级能力的需求，再去付费使用顶尖的专有模型。对于大型企业客户而言，如果需要配套的企业级外壳工具（wrappers）、技术支持和员工权限管理，直接采购 Anthropic、OpenAI、Grok 或 Google Gemini 即可，核心在于选择合适的服务伙伴。
谈到商业合作，节目现场随即进入了充满恶搞意味的“自卖自夸”环节。主持人接连抛出荒诞的赞助促销码和夸张条件：谁要是签下 8090 价值 1000 万美元的企业大单，就能换取几位主持人的 5 分钟 Zoom 通话；如果有人预购弗里德伯格（David Friedberg）旗下 100 万美元的土豆种子，也能解锁专属福利；而杰森·卡拉卡尼斯（Jason Calacanis）更开玩笑称，只要有人向他的 Launch Fund 5 基金注资 2500 万美元，就能获得与“JCal 哥”同台联合主持播客的机会。
### [48:25](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2905s) Anthropic 达成 15 亿美元史上最大版权和解
玩笑过后，话题转向本周重磅法律事件：Anthropic 周一就其 AI 训练侵权诉讼达成和解，赔偿金额高达 15 亿美元。这是美国司法史上金额最大的一笔版权和解案，也是针对 AI 大模型训练版权侵权的首次重大和解。在此案中，Anthropic 被指控从 LibGen 等盗版网站下载了多达 700 万本盗版书籍用于训练 Claude。根据和解方案，约有 50 万本书籍被纳入赔偿范围，原告作者每本书可获赔 3000 美元，代表律师团将拿走 1.01 亿美元的诉讼分成，目前已有 91% 的涉案作者认领了这笔赔偿。
```mermaid
flowchart TD
accTitle: Anthropic 15亿美元版权和解案构成
accDescr: 展示Anthropic从盗版网站下载书籍训练Claude到达成15亿美元和解的资金与赔偿分配流向。
subgraph DataOrigin["侵权源头"]
A["从LibGen等盗版网站<br/>下载700万本盗版书"] --> B["训练Claude模型"]
end
subgraph Settlement["15亿美元和解方案"]
B --> C["达成和解协议"]
C --> D["律师分成<br/>1.01亿美元"]
C --> E["50万本受涵盖图书<br/>每本作者获赔3000美元"]
E --> F["已有91%涵盖作者<br/>认领赔偿金"]
end
```
杰森借此狠狠展开了一番“胜利巡游”（victory lap），并在节目中插播了自己在 2023 年做出的预测回放。当时杰森就断言，大语言模型之间的能力差距会迅速缩小，技术本身甚至可能沦为大宗商品化，而开源最终将大放异彩；因此，AI 浪潮的最大赢家将是《纽约时报》（The New York Times）、Reddit、X（Twitter）和 YouTube 这些掌握核心训练数据的拥有者。杰森情绪激动地表示，内容创作者与版权方必须团结起来一致对外，向 AI 巨头发出通牒：要么支付高昂的许可费，要么彻底禁止爬取索引，绝不能任由 AI 公司无休止地剥夺内容创作者的生存空间。
- 和解金额：15 亿美元，创美国版权诉讼和解历史纪录
- 涉及数据：从影子图书馆（如 LibGen）非法下载约 700 万本图书
- 赔付明细：涵盖 50 万本书籍，作者每本可领 3000 美元；原告律师费占 1.01 亿美元
- 认领比例：目前已有 91% 的涉案作者提交了索赔认领
### [50:48](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=3048s) 萨克斯拆解和解本质：盗版与“双标”的知识产权虚伪
大卫·萨克斯（David Sacks）随即给杰森的兴奋浇了一盆冷水，指出这份 15 亿美元的和解案并没有真正解决 AI 训练与“合理使用”（Fair Use）的核心法律争议。萨克斯指出，Anthropic 本次之所以被迫巨额和解，关键在于他们的手段属于彻头彻尾的“盗版”——他们直接从盗版电子书库抓取，连哪怕一本正版书的钱都舍不得掏。如果 Anthropic 当初花钱购买了每一本书的正版拷贝，他们就能在法庭上以“合理使用”为由抗辩到底，而无需背负直接盗窃盗版的硬伤。
```mermaid
flowchart LR
accTitle: AI巨头在数据训练上的双标逻辑
accDescr: 对比Anthropic等AI巨头对待全网创作者数据与对待自身模型输出时自相矛盾的立场。
subgraph Outward["对待外部世界创作者"]
A["购买单本/公开抓取"] --> B["强行训练大模型"]
B --> C["高呼属于合理使用<br/>创作者无权阻拦"]
end
subgraph Inward["对待自身模型输出"]
D["竞争对手或中国团队<br/>通过API/输出来做蒸馏"] --> E["定性为工业级蒸馏攻击"]
E --> F["指责危及国家安全<br/>却不敢在法庭诉称侵犯IP"]
end
```
萨克斯痛斥了 Anthropic 与 OpenAI 表现出的极度虚伪：在面对全球创作者时，他们主张无论创作者是否同意，AI 公司只要拿到数据就有权进行训练并推导模型权重；但当其他公司或中国团队利用其模型输出进行蒸馏（Distillation）时，他们却立刻换了一副嘴脸，将其视作不可容忍的行为。萨克斯特别提到，Anthropic 在 2024 年 2 月的技术博客中生造了所谓“工业级蒸馏攻击”（Industrial-scale distillation attacks）的概念。而此前，蒸馏在 AI 圈一直属于通行的技术标准。
萨克斯进一步揭露，Anthropic 在公开推销所谓“蒸馏攻击”威胁时，通篇不敢使用“侵犯知识产权”（IP theft）这个词，而是将其包装为“缺乏安全护栏的国家安全风险”。他们之所以不敢在明面上指控中国公司侵犯知识产权，是因为如果他们将“利用输出数据训练模型”定性为侵权，就等同于亲手摧毁自己在《纽约时报》等一系列诉讼中的“合理使用”抗辩基础。查马斯也补充指出，《纽约时报》起诉 OpenAI 本质上指控的就是一场工业级蒸馏——OpenAI 动用爬虫以人类不可能企及的规模抓取报纸全文，反向工程生成了模型权重。萨克斯总结道，这种“我的知识产权是财产，你的知识产权是合理使用”（IP for we but not for thee）的双重标准不仅荒唐，政策制定者更不应该被这些 AI 巨头的游说话术所蒙蔽。
### [55:23](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=3323s) 弗里德伯格的思想实验：AI 学习公开发表评价算不算侵权？
为了厘清合理使用的边界，弗里德伯格向杰森提出了一个层层递进的思想实验。他拿杰森当年撰写的畅销书《天使投资圣经》（*Angel: How to Invest in Technology Startups*）举例：“假设你写了这本书，并明确拒绝将其提供给 AI 训练，也不让搜索引擎抓取全文，任何人想读你的大作都必须自掏腰包花 30 美元购买。”
随后弗里德伯格构建了具体场景：假设有一位真实读者花 30 美元购买并阅读了杰森的《天使投资圣经》，写了一篇详尽的书评发布在公开网络上，网络上陆陆续续积累了上千篇第三方的专业分析与书评元数据。如果 AI 引擎从来没有下载过这本 30 美元的电子书本身，而仅仅是通过公开网络爬取了这 1500 多篇读者评论、摘要和元数据，并借此理解了书的核心观点，随后在用户提问时能够对该书的内容侃侃而谈。弗里德伯格质问杰森：在这种情况下，你是否依然认为自己的图书版权遭到了侵犯？杰森在被问及自己那本“将 10 万美元变成 1 亿美元的投资蓝图”时，陷入了思考并坦承：如果是这种情况，自己似乎确实只能被迫接受。
### [57:00](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=3420s) 知识的弥散与版权的法律边界
嘉宾们在一番打趣中切入讨论。大家调侃起网络上为书籍撰写的虚假好评，甚至戏称某些夸张书评是“用开源 AI 垃圾炮喷出来的”。然而，弗里德伯格（David Friedberg）借此引出了一个核心的法律与哲学命题：人类知识本质上是弥散的，无法被彻底封闭。现行的版权法判例非常清晰——任何人不得直接从一本书中摘抄整段文字并据为己有，这是无可争议的侵权行为；但是，阅读一本书、阅读书评，并将其中扩散的知识融会贯通，在抽象层面上转化为新的内容供他人阅读，这在法律上属于正当转换。弗里德伯格认为，指望将数字化的知识传递与处理判定为侵犯版权，在法理上是很难成立的。
```mermaid
flowchart TD
accTitle: 知识转换与版权法边界
accDescr: 对比直接复制与知识转换在版权法下的定性区别
subgraph DirectCopy["侵权范畴 (Violation)"]
A["原始内容 / 文本"] --> B["直接摘抄与再版"]
B --> C["侵犯版权法"]
end
subgraph Diffusion["合法转换 (Fair Use / Diffusion)"]
D["阅读内容与书评"] --> E["知识抽象与理解"]
E --> F["生成全新衍生内容"]
F --> G["受保护的合法输出"]
end
```
杰森（Jason Calacanis）则对此提出反驳。他指出，市场上虽然一直存在合法的摘要缩写（Cliff's Notes），版权法也有一套成熟的“四要素测试”（Four-Factor Test），但当 AI 模型的应用层直接与原内容创作者展开商业竞争时，性质就完全变了。杰森以音乐产业为例：唱片巨头们就像“恶犬”（rabid dogs）一样在法庭上寸步不让，死磕到底，直到科技公司妥协并达成巨额授权和解；随之确立的判例则会像多米诺骨牌一样套用到下一家公司头上。
- **汤森路透诉 ROSS 案（Thomson Reuters v. ROSS Intelligence）**：杰森着重提及此案，该案涉及利用法律数据库 Westlaw 的输出结果来训练 AI。当训练出的 AI 产品与原告处于同一赛道且直接抢夺市场时，版权法往往会给予原作者极高的排他性保护。
- **直接竞争的临界点**：如果 AI 仅仅是概括一本书讲了什么，尚属合理使用范畴；但一旦生成的工具在市场上直接替代原产品，AI 公司的抗辩就会迅速瓦解。
---
### [01:00:50](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=3650s) 监管游说“搬起石头砸自己的脚”：自相矛盾的知识产权指控
面对版权诉讼的浪潮，杰森建议所有头部 AI 公司成立自律组织，拿出 10% 的营业收入设立版权资金池，持续向媒体与出版业支付版权费并获取更新授权。但萨克斯（David Sacks）立刻指出这一想法过于天真：“10% 根本满足不了那些恶犬，他们想要的是 100%。”萨克斯随即提出了一个直击命门的致命拷问：在合理使用原则尚未盖棺定论、Anthropic 与 OpenAI 正面临资金雄厚的内容创作者发起天价诉讼的当下，他们为了游说美国政府打压中国开源模型，竟然公开将“未经创作者同意蒸馏其模型输出”定性为“知识产权盗窃”（IP theft），这是否犯了一个致命的战略失误？
萨克斯分析认为，Anthropic 这无异于“搬起石头砸自己的脚”（hoisted on their own petard）——这在年轻人的流行语里就是典型的“送人头”（self-own）。如果未经同意使用模型输出是偷窃，那么 Anthropic 和 OpenAI 当初未经许可抓取全网创作者的版权内容来训练模型，按此逻辑岂不等于在法庭和监管机构面前主动承认了“自己的整个产品本身就是偷来的”？这一自相矛盾的表态，直接给那些要求瓜分 AI 公司全部营收的版权诉讼集团送上了最强有力的法庭弹药。
- **更明智的公关路径**：萨克斯指出，Anthropic 本可以采取更审慎的策略，指责某些中国公司利用虚假账户和代理节点绕过其服务条款（TOS），将其定性为“欺诈性商业行为”（deceptive business practice），并在不触碰知识产权定义的前提下请求政府干预。
- **爬虫索引的前车之鉴**：传统媒体曾要求谷歌将搜索引擎爬虫与 AI 爬虫分开，甚至默多克（Rupert Murdoch）当年也曾试图联合报业对抗谷歌抓取，但内容方往往因缺乏统一阵线而在博弈中妥协。
---
### [01:04:13](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=3853s) 硅谷创业生态圈的集体恐慌
由于 Anthropic 将模型蒸馏上升到“知识产权盗窃”的高度，这种强烈的政策游说反冲力迅速波及了整个硅谷创业生态。如果监管层采纳了这套说辞并将中国开源模型贴上“非法盗窃产物”的标签，那么所有基于中国开源模型微调或衍生的下游创业项目都将被视为“受污染的非法产物”。
正因如此，包括 Y Combinator 负责人陈嘉兴（Garry Tan）在内的近 200 位硅谷初创公司创始人联名向特朗普政府致信，紧急呼吁不要封禁中国开源 AI 模型。初创圈清醒地意识到，头部巨头为了维护自身护城河而推行的监管俘获（Regulatory Capture），其代价是把整个美国中小创业者的技术栈与创新空间彻底逼入绝境。巨头们本以为可以通过道德高地卡死海外竞争对手，却未曾料到不仅激活了国内版权方的索赔胃口，还激怒了原本属于自己阵营的整个开发者生态。
### [01:05:34](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=3934s) 华盛顿游说之火与播客虚构书单插曲
讨论继续围绕 AI 巨头试图在华盛顿进行监管寻租展开。查马斯（Chamath Palihapitiya）直言，这些企业如今可能真的走得太远了，“如果你跑去华盛顿与恶犬同眠，就别惊讶醒来时满身跳蚤”（"if you lay down with the dogs, don't be surprised if you wake up with the fleas"），它们主动去招惹政界，很可能是捅了马蜂窝（"poked the tiger"）。
紧接着，杰森（Jason Calacanis）开启了一段轻松幽默的插曲，开玩笑宣布《All-In》旗下虚构的 2027 年图书出版计划，调侃各位主持人的经典人设与八卦：
- 大卫·萨克斯（David Sacks）有两本大作：一本是极尽奢华风的《腕上法拉利：如何赢在人生并买得起 25 万美元的名表》（*Ferrari on my wrist: How to win at life and afford a $250,000 watch*）；另一本则是其政治宏愿《拯救美国之战：摧毁社会主义、缩减债务并赢得 AI》（*The fight to save America: Destroying Socialism, Shrinking the Debt, and Winning AI*）。
- 查马斯则被安排了两本书：一部是他唯一的浪漫虚构小说《查马斯·帕利哈皮蒂亚：性感的意大利之夏》（*A sexual Italian summer*）；另一部则是非虚构写实作品《企业级销售，查马斯的反乌托邦：我的软件初创公司如何影响了我的意大利之夏》（*Enterprise sales. Chamath dystopia. How my software startup affects my Italian summer*），众人捧腹调侃他是“米兰的 IT 男”。
- 此外，节目组未来还将推出布拉德·格斯纳（Brad Gerstner）、比尔·格利（Bill Gurley）以及埃隆·马斯克（Elon Musk）等重磅客座嘉宾的虚构新书。
### [01:07:07](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4027s) 科技巨头财报与暴增的 AI 资本支出：Google 与特斯拉由盈转负
玩笑过后，节目切入严肃财经话题。当天 Google（Alphabet）与特斯拉（Tesla）双双公布了财报，资本支出（capex）规模之庞大引发市场剧烈震荡。虽然 Google 业绩表现极度强劲——Google Cloud 年化营收突破 1000 亿美元大关，增速大幅超预期——但两家公司的股价却出现重挫：特斯拉盘中大跌 14%，Google 则下跌 7% 至 10%。
暴跌的核心推手在于惊人的资本开支与自由现金流（FCF）转负：
- **资本开支激增**：特斯拉 capex 同比暴增 140%，预计全年达 250 亿美元；而 Google 的全年 capex 预期上调至 1950 亿至 2050 亿美元区间，且明年预计更高。
- **自由现金流转负**：两家公司账上的现金不仅没有增加，反而由于巨额支出出现消耗。对 Google 而言，这是其上市以来首次出现单季度自由现金流为负的历史性转折点。
- **SpaceX 上市追踪**：杰森顺带提及刚上市不久的 SpaceX，市值从 2 万亿美元的高点受锁定期（lockups）分阶段解禁及市场情绪影响回落约 30%，目前在 1.5 万亿美元区间震荡寻找底部。
杰森指出，巨头的大举投资主要集中在 AI 基础设施上。与单纯为了向终端用户提供服务的 OpenAI 不同，Google 的资本开支具备双重收益逻辑：一方面自用于驱动搜索、YouTube 与内部基础设施；另一方面通过 Google Cloud（GCP）将其作为算力产品对外转售。这本应是押注未来的积极买入信号，市场却给出了极为负面的抛售反馈。
```mermaid
flowchart TD
accTitle: Google资本支出与飞轮变现机制
accDescr: 展示Google近2000亿美元资本支出如何转化为内部应用与外部云业务双重回报
subgraph Capex["2026年巨额AI资本支出 (~2000亿美元)"]
Silicon["自研芯片 (TPU等)"]
DC["超大规模AI数据中心与算力集群"]
end
subgraph Internal["内部核心业务赋能"]
Search["搜索体验与AI交互升级"]
YT["YouTube个性化与智能广告投放"]
end
subgraph External["外部企业级变现"]
GCP["Google Cloud (企业模型微调与算力租售)"]
Agnostic["支持500+异构模型 (模型不可知论)"]
end
Silicon --> DC
DC --> Internal
DC --> External
Internal --> CashFlow["强劲内生复合回报"]
External --> CashFlow
```
### [01:09:32](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4172s) 查马斯看多 Google：32% 的复利机器与模型碎片化红利
面对市场的恐慌性抛盘，查马斯明确表示自己对 Google 极度看多（"more bullish"）。他在现场让弗里德伯格（David Friedberg）猜测 Google 上市 20 多年以来的平均投入资本回报率（ROIC）：弗里德伯格先后猜了 17%、21%、23%、29% 乃至 35%，最终查马斯公布官方真实数据为 **32%**。
查马斯指出，当一家公司能够以 32% 的年化复利水平持续稳定运转 20 余年时，投资者完全应当给予管理层足够的信任，他们不是在盲目狂飙，而是在极其有条理地巩固核心壁垒。社交媒体上关于 Gemini 真实使用量或者 AI 赋能工作流是否虚高的争论纯属“无稽之谈”（"all malarkey"）。对于 Google 而言，开源生态与模型的极端碎片化反而是最大利好：
- **芯片层与云平台通吃**：就算未来市场上涌现出 500 种不同的模型，Google 凭借自身的硅片设计能力与云平台，只要支持所有模型运行，就能在芯片层和基础设施层赚取极其丰厚的利润。
- **应用层无缝渗透**：依托 AI 技术更精准地进行广告投放并协助创作者产出高质量内容，YouTube 和搜索等核心现金牛将进一步受益。
- **市场恐慌的实质**：查马斯引述瑞安·彼得森（Ryan Petersen，Flexport 创始人）的推文比喻——Google 今年的资本支出规模相当于美国全军军费预算的约 20%。投资者从未见过如此空前规模的资本投入，加上上市以来首次自由现金流转负，短期的抛售更多是情绪上的错愕与震撼。
### [01:11:50](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4310s) 弗里德伯格论 GCP 的护城河与最差情景底线
弗里德伯格对查马斯的观点表示高度赞同。杰森对比了苹果（Apple）的做法——苹果将赚来的巨额利润用于数百亿美元的股票回购或派发股息，而科技巨头如今将资本投入到“即时智能”（on-demand intelligence）这一没有天花板的未来领域，显然是更具远见的战略抉择。弗里德伯格强调，如果投资者仅仅因为想要短期股票分红而反对 Google 进行长期基建投资，“那你就根本不该持有这只股票，自然会有别人来接盘”。
弗里德伯格全面剖析了 Google 的深层价值底盘：
- **多元资产组合**：除了消费级生态、YouTube 和 GCP 外，Google 还持有 10% 的 SpaceX 股份、大比例的 Anthropic 股权，以及自动驾驶领头羊 Waymo。值得注意的是，仅在本季度，Google 对 Anthropic 的按市值计价调整（mark-to-market write up）就高达 1000 亿美元。
- **企业级应用的天生优势**：在企业办公领域，没有哪家云厂商比 GCP 更具优势。企业最核心的数据——包括 Gmail 邮件、Google Drive 内部文档与协作资产——全都在 Google 生态中。企业迫切希望 AI 能访问并理解这些私域数据以大幅提升工作生产力。
- **开放中立与模型不可知（Model-Agnostic）**：与部分深度绑定单一模型架构的云服务商不同，GCP 允许客户自由运行任何第三方模型和任意自定义工作流，拥有极高的微调与定制自由度。
- **最差底线保障（Worst-case scenario）**：弗里德伯格推演指出，哪怕在最极端的最差假设下——Google 所有的应用层产品全部受挫、网络效应失灵、内部自研模型一无所成——Google 依旧坐拥全球成本最低、效能最高的算力基础设施。就像马斯克依托 Colossus 超级算力集群为 Grok 取得惊人回报一样（类似“Elon Web Services”的算力租赁），只要 AI 发展的大趋势不变，Google 仅凭世界顶级的基础设施印钞机，就能在未来数年内持续赚取海量现金。
### [01:14:43](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4483s) 押注 AI 的最佳标的与苹果云计算的“五个九”困境
在审视当前二级市场时，嘉宾们认为如果想直接押注 AI，谷歌（Google）依然是公开市场中最具性价比的选择。投资谷歌不仅能获得其前沿的 AI 布局，还能同时拥有 YouTube 的庞大生态、稳健的 C 端消费者业务以及全套基础设施，而其目前的估值倍数依然显得相当便宜。话题随后转向苹果公司：苹果为何未能效仿亚马逊打造出属于自己的“苹果网络服务（Apple Web Services）”？
从表面上看，苹果与全球开发者保持着极为紧密的联系，坐拥利润丰厚的 App Store 以及忠诚的开发者群体，似乎极具优势。但查马斯（Chamath）指出，涉足云计算基础设施绝非易事。亚马逊历时约 17 年才将 AWS 打磨成熟，谷歌也耗费了十二三年的时间才勉强赶上。作为云服务提供商，企业向客户承诺的核心指标是高达 99.999%（“五个九”）的高可用性与稳定性。达到前两个九（99%）的门槛相对较低，可能只需少量成本就能为制药或国防机构搭建普通托管服务；但推进到第三个九（99.9%）就需要数十亿美元的资本开销，第四个九（99.99%）要砸进数百亿美元，而要确保第五个九的极度可靠，所耗费的资金将高达数千亿美元。这是一场需要海量技术积累和持续资本支出的极重资产游戏，目前市场上实质上只有三家巨头能够在这一级别展开较量。
```mermaid
flowchart TD
accTitle: 云服务可用性与资本支出阶梯
accDescr: 展现从两个九到五个九可用性所需的指数级资金攀升
subgraph CostTiers["云基础设施可用性 vs 资本开支"]
A["99% 可用性 (两 9)<br/>相对低廉 / 基础托管"] --> B["99.9% 可用性 (三 9)<br/>数十亿美元级投入"]
B --> C["99.99% 可用性 (四 9)<br/>数百亿美元级投入"]
C --> D["99.999% 可用性 (五 9)<br/>数千亿美元级巨额护城河"]
end
```
- **苹果的资本分配选择**：在过去十年中，苹果通过股票回购返还了约 7550 亿美元，加上约 1400 亿美元的分红，累计向股东返还了近 9000 亿美元。
- **野心与保本的平衡**：杰森（Jason）感慨，如果苹果当年能将这笔巨款中的一半用于雄心勃勃的新品研发、推进自研汽车项目，或是战略投资 Anthropic、SpaceX 等前沿公司，技术格局可能会大不相同。但大卫·萨克斯（David Sacks）与查马斯反驳称，苹果采取的是“不作恶 / 绝不犯错（do no harm）”的资本分配策略，虽然保守，但切实支撑了股价；随着工程背景更浓厚的高管团队走上前台，外界正在关注其未来是否会打破原有的稳健脚本。
---
### [01:17:19](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4639s) “社会主义专栏”启动：驱逐房客等于暴力？
讨论随后切入全新常驻板块——“社会主义专栏（Socialism Corner）”。杰森直指纽约市政界推行的激进住房干预政策。以纽约市极左翼政治力量为代表的佐兰·马姆达尼（Zohran Mamdani）在下东区廉租公寓博物馆（Tenement Museum）举办了一场所谓“租房欺诈”听证会，并在会后推出了一套针对房东的严苛限制方案。
这套法案如果获得通过，将全面禁止房东向租客收取信用调查费用；在租客审核标准上，房东只能在“信用记录调查”或“年收入达到租金 40 倍的标准”中二选一，不得同时要求两项达标；此外，法案还将从法律上正式承认租客工会（tenant unions），并配合长达一年的全市租金冻结令。听证会上一名头戴夸张帽子、佩戴严密防疫大号口罩的左翼活动家公开发言，直接将房东收回房产的行为定性为暴力，宣称：“马姆达尼政府正在赋予我们力量，让我们不再把驱逐房客这种‘暴力行为’当作习以为常的商业惯例来容忍。”
这套言论和打扮引得全场啼笑皆非。萨克斯和查马斯调侃这位活动家的打扮宛如经典动画《胖子阿伯特（Fat Albert）》里的角色，甚至惊叹“难道如今疫情还没过去吗”，现场火药味与嘲弄感十足。
---
### [01:19:03](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4743s) 弗里德伯格长篇檄文：私有财产权是美国自由的基石
面对将“驱逐欠租租客”等同于“暴力”的论调，大卫·弗里德伯格（David Friedberg）展开了一场极具历史纵深感的系统性批判。弗里德伯格追溯至 1787 年约翰·昆西·亚当斯（John Quincy Adams）在其著作《美利坚合众国政府宪法答辩》中的名言：“一旦社会承认财产不再像上帝的法则一样神圣，并且没有法律和公共正义的力量来保护它，无政府状态与专制暴政便随即拉开序幕。”亚当斯随后在 1791 年的文章中再次强调：“必须捍卫私有财产，否则自由绝不可能存在。”
```mermaid
flowchart LR
accTitle: 从侵蚀财产权到专制暴政的演变链条
accDescr: 弗里德伯格阐述的道德污名化至无政府再到专制权力的演变路径
A["道德批判与污名化<br/>（声称产权拥有者作恶、驱逐即暴力）"] --> B["剥夺私有财产权<br/>（政府限制处置、冻结租金、强行没收）"]
B --> C["滑向无政府混乱<br/>（规则崩塌、群体互抢利益）"]
C --> D["催生极权暴政<br/>（强势团伙掌权，建立寡头管控）"]
```
弗里德伯格深入剖析了美利坚立国之本的核心逻辑：早期移民之所以背井离乡远赴北美，正是为了逃离欧洲君主专制与封建领主社会。在旧大陆，个人毫无私有财产保障可言，领主与专制官僚可以随时没收农民的土地与收获，民众不过是依附于封建领主的农奴。而美利坚合众国之所以伟大，根本在于确立了“个人可以拥有无可剥夺之财产”的权利。
弗里德伯格发出强烈警告，社会主义意识形态往往习惯披上崇高的道德伪装：
- **道德污名化先行**：先将拥有私有财产的房东或富人塑造成“掠夺者”和“不义之人”，把合法的商业履约或驱逐违约租客歪曲为“对无产者的暴力犯罪”；
- **通过混乱滑向极权**：以这种道德优越感为借口，公权力与激进团体便顺理成章地侵占个人私产。无政府状态永远只是一种短暂的过渡阶段，在缺乏产权与法律秩序的互抢混乱之后，必然凝聚出手段最狠辣的强势集团，最终演变成彻底的政治寡头或专制暴政。
- **微小倒退的连锁反应**：弗里德伯格强调，每一个看似琐碎、荒谬的微小监管，合在一起就是对私有财产权这一立国基石的逐步剥蚀；哪怕只是退让毫厘，也会引发无法挽回的连锁崩溃。
萨克斯随即接棒加入论战，指出这些美国民主社会主义者（DSA）宣称“驱逐是暴力”并企图全面终结驱逐程序，表面上是在践踏房东对不动产的支配权益，但更深远的危害在于，它彻底摧毁了同一栋建筑中遵纪守法的其余守法租户的居住权益与社区秩序。
### [01:23:25](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=5005s) “奢华信念”与恶邻之害：租管政策对守法租户的次生灾害
大卫·萨克斯（David Sacks）指出，当公寓楼里聚集了大量赖账不缴租金的租客，房东便失去了维护和翻新大楼的现金流，导致整栋建筑加速衰败破损。更为残酷的现实是，这些本该被驱逐却无法被赶走的“霸房者”（squatters）和违约租客，往往是最糟糕的邻居。萨克斯援引一位管理公寓的朋友的切身经历：楼内经常有人半夜放震耳欲聋的音乐、砸墙砸门、散发难闻恶臭、霸占破坏公共区域，或是酗酒寻衅滋事。对于那些靠租金管制在此生活了几十年的低收入守法老人来说，他们无力搬迁，只能被动困在居住环境恶化的绝境之中。
```mermaid
flowchart TD
accTitle: 驱逐禁令与租管政策的恶性循环
accDescr: 展示禁止驱逐与过度监管如何摧毁公寓生态并转嫁成本给守法租户
subgraph 租客与邻里冲击
A["禁止驱逐失德与欠租租客"] --> B["恶邻滋生扰民与治安恶化"]
B --> C["守法老弱租户生活品质崩塌"]
end
subgraph 资本与物业崩溃
A --> D["房东收入中断、无力维护修缮"]
D --> E["建筑破败/沦为恶性保障房"]
E --> C
end
subgraph 市场防御与反噬
D --> F["翻新合规成本过高"]
F --> G["房东宁愿空置（幽灵公寓）"]
G --> H["有效供给萎缩、起租门槛暴涨"]
end
```
这种政策偏向暴露出进步派思维方式的根本弊病。萨克斯犀利地批评道，美国民主社会主义者（DSA）大多是高学历且家境优渥的精英阶层，他们之所以能够毫无顾忌地持有这些关于公共空间的“奢华信念”（luxury beliefs），是因为他们根本不用去坐地铁、搭公交或去公共公园。当这些公共场所被吸毒者和流浪汉占领时，富裕阶层总是在捍卫吸毒者的所谓“权利”，而非维护工薪家庭的安宁。查马斯·帕里哈皮蒂亚（Chamath Palihapitiya）与杰森·卡拉卡尼斯（Jason Calacanis）对此深表认同，讽刺那些住在加州马林县（Marin County）高档社区或纽约上东区的人，既不用生活在旧金山田德隆区（Tenderloin），也不用在送孩子上学时目睹有人在地铁随地大便或注射毒品。这种道德优越感被廉价地消费着——正如萨克斯所言，“住在封闭式门禁社区里的私募股权合伙人太太们会觉得自己品德高尚，因为她们成功阻止了驱逐”，但真正承受代价的，恰恰是楼里走投无路的平民百姓。
### [01:26:03](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=5163s) 第一性原理与供给法则：奥斯汀模式对决左派堡垒
从第一性原理（First Principles）出发，解决住房危机的逻辑其实简单明了：任何懂基本经济学的人都知道，只要增加供给，价格就会下降。杰森表示，此时身处东京的他感受尤为明显——东京早在几十年前就搞清楚了这一点，只需不断向上建造高楼、增加单位供给；德克萨斯州、佛罗里达州和内华达州同样印证了这一逻辑。不论新增的是豪宅、多户型公寓还是单户住宅，只要配套交通跟得上，新供给的涌入就会自然平抑整体租金。然而，唯独纽约、洛杉矶和旧金山这些精英自由派把持的城市，始终在最基础的概念上原地打转。
查马斯指出，德克萨斯州奥斯汀（Austin）的真实数据给出了最强有力的实证：一旦地方政府放宽建筑许可限制，新开工项目就会成倍增加，而每一个进入市场的新房源，都会直接推动租金向下走。如果政治人物真的在乎“可负担住房”，就应该动用政治意愿推行许可审批制度改革，而不是去强推那些剥夺房东基本权利的管制法案。纽约州众议员佐兰·马姆达尼（Zohran Mamdani）等极左政客所耗费的政治资本，若能用于松绑审批，对社会的益处将远超破坏市场机制的极端提案。
### [01:27:31](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=5251s) 私有产权的瓦解与反常识监管：当租金成为“自选项”
谈及私有产权的本质，查马斯强调戴维·弗里德伯格（David Friedberg）此前提出的警示至关重要。如果按照极左派的逻辑推演到极致，荒谬程度堪比：“你正开着车，突然有人跳进副驾驶座说‘我现在在这里了，你无权把我赶下去’；或者你开门去拿个联邦快递（FedEx）或亚马逊（Amazon）包裹，陌生人直接冲进你家坐在沙发上，法律却规定你不能赶他走。”
如果政府剥夺物理财产所有者进行信用背景审查、无法筛选租客并禁止驱逐，必然导致更剧烈的市场反弹。查马斯甚至戏谑地提议，干脆就让纽约通过这项法律，然后和奥斯汀进行一场严谨的“科学 A/B 测试”，让所有人看看灾难性的后果。萨克斯对此表示悲观，认为社会主义者从来不会从失败的实验中汲取教训：
- 阿根廷总统哈维尔·米莱（Javier Milei）废除了租金管制后，当地住房供给迅速激增，实际租金显著下降；
- 芝加哥等城市已经在盲目管制中走向下坡路，但左派依然视若无睹；
- 在纽约这样体量最大、最复杂的舞台上搞这种激进社会实验，只会演变成一场代价惊人的公开惨剧。
### [01:29:35](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=5375s) 5万套“幽灵公寓”与房东的反抗：严管之下的市场扭曲
纽约当前的监管政策正在将市场推向死胡同。官方不仅试图彻底封死驱逐通道，还禁止房东对潜在租客进行信用审查、犯罪背景调查甚至核实收入来源。杰森感叹道：“我们已经完全脱离了常理。”如果房东既不能核实租客是否有支付能力，租客违约后又无法驱逐，租金在事实上就变成了“自选项”（optional），租客按时交租的动力直接降至为零。
面对这种系统性道德风险，理性的房东只能采取极端的避险策略：
- **租金暴涨与高门槛预付**：为了消化违约租客带来的坏账损失，房东必须将初始挂牌租金提高三到四倍，并要求租客在签约时直接电汇预付整整一年的租金，再通过漫长的市场试探慢慢下调；这种严苛的对冲手段最终彻底摧毁了低收入人群的租房希望。
- **5万套“幽灵公寓”的诞生**：纽约繁琐严苛的翻新法规要求公寓在改造时必须满足海量建筑规范，单套装修成本动辄耗费数十万美元。面对无法通过合理租金收回成本的困境，房东宁愿将腾出来的空房彻底封锁闲置，也不愿投入巨资装修后亏本出租。据相关报道，纽约市内已出现多达5万套处于长期空置状态的“幽灵公寓”（ghost apartments）。
- **短租禁令的挤压**：杰森提醒，这绝非 Airbnb 造成的现象，因为纽约市此前早就通过法案事实上取缔了 Airbnb 短租。房东在既不能合法维权、又无法盈利的绝境下，唯一的出路要么是低价抛售资产迁往佛罗里达或德州等自由辖区，要么就是锁死大门干耗到底。
### [01:31:55](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=5515s) 驱逐困境导致的“幽灵公寓”与节目尾声及峰会推介
针对纽约等大城市对租金与驱逐制度的严苛管控，主持人们探讨了房东在极端政策压力下的无奈对策。当房东无法自由调整租金、无法合法驱逐违约违规的租客，甚至在签约前连基本的租客背景调查（diligence）都受到重重限制时，理性的经济行为往往走向反面：只要手头背负的债务压力不算太大，房东宁可选择把整栋大楼或单元空置出来观望等待，也不愿承担租客欠租不走又无法驱逐的巨大风险。这种荒谬的政策扭曲直接导致了大量“幽灵公寓”（ghost apartments）的出现，反而进一步加剧了市场上可租房源的短缺。
在结束沉重严肃的社会议题讨论后，节目切入每年一度的 All-In 峰会（All-In Summit 2026）宣传。Jason Calacanis 热情号召观众今年务必到场参会，强调本届峰会拥有世界级的演讲嘉宾阵容、顶级活动与派对体验。参会者中高达 80% 均为创始人、投资人或高级管理者，今年峰会还将更加聚焦于高端社交与人脉连接。其他几位 Besties 也纷纷插话补充现场的豪华待遇，从精致的美食美酒，到历届令人难忘的重磅音乐表演嘉宾（如 Grimes 与 Diplo 等）。Jason 呼吁大家访问官网抢票，并宣布本期“传奇”级别的 All-In 播客正式圆满收官。
节目的最后伴随着几位好友之间经典的搞笑互损、片尾标志性的口号（“Let your winners ride”）以及幕后花絮式的荒诞调侃与日常闲聊，在热烈而幽默的气氛中落下帷幕。
- **租房市场扭曲**：房东因无法有效背景调查租客且面临驱逐禁令，选择主动空置房源，“幽灵公寓”现象频现。
- **All-In 峰会亮点**：参会人群中创始人、投资人与核心高管占比达 80%，今年重点强化深度 Networking，并继续配备顶级餐饮及音乐演出嘉宾。
## Notable quotes
> [02:00](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=120s) 这个辩论完全搞反了，本该是 Anthropic 交代为什么防盗做得如此糟糕，现在却要全美开源生态来为 Anthropic 的失职买单。 — David Sacks
> [10:18](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=618s) 如果美国政府真的介入并对开源模型下达禁令，最终结果将是直接砸穿美国股市。 — Chamath Palihapitiya
> [13:38](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=818s) 既然能买 50 美分的，为什么非要买 50 美元一罐的可乐？ — David Sacks
> [15:12](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=912s) 这并不意味着我们在偷窃他们的算法，我们没有黑进他们的服务器盗取软件代码，我们只是分析了他们软件输出的结果，并用它来改进我们自己的软件。 — David Friedberg
> [20:10](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=1210s) 那些想控制专有技术收租的 1% 寡头们，他们的好日子到头了，我们甚至根本不需要设立所谓的财富税。 — David Friedberg
> [33:24](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=2004s) 让老子说完这 30 秒，闭嘴听我说！ — Jason Calacanis
> [01:00:50](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=3650s) 如果未经同意使用模型输出是偷窃，那么 Anthropic 和 OpenAI 当初未经许可抓取全网创作者的版权内容来训练模型，按此逻辑岂不等于在法庭和监管机构面前主动承认了‘自己的整个产品本身就是偷来的’？ — David Sacks
> [01:05:34](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=3934s) 如果你跑去华盛顿与恶犬同眠，就别惊讶醒来时满身跳蚤。 — Chamath Palihapitiya
> [01:11:50](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4310s) 那你就根本不该持有这只股票，自然会有别人来接盘。 — David Friedberg
> [01:19:03](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=4743s) 一旦社会承认财产不再像上帝的法则一样神圣，并且没有法律和公共正义的力量来保护它，无政府状态与专制暴政便随即拉开序幕。 — David Friedberg
> [01:23:25](https://www.youtube.com/watch?v=wcV0SRPFK9s&t=5005s) 住在封闭式门禁社区里的私募股权合伙人太太们会觉得自己品德高尚，因为她们成功阻止了驱逐，但真正承受代价的，恰恰是楼里走投无路的平民百姓。 — David Sacks
## Mentioned
Jason Calacanis, David Sacks, Chamath Palihapitiya, David Friedberg, Michael Kratsios, Howard Lutnick, Donald Trump, Ben Thompson, Sam Altman, Mira Murati, Bernie Sanders, Elizabeth Warren, Ryan Petersen, John Quincy Adams, Zohran Mamdani, Javier Milei, Garry Tan, LeBron James, Rupert Murdoch, Elon Musk, Brad Gerstner, Bill Gurley, Moonshot AI, Kimi K3, Kimi K2.5, Anthropic, Claude Opus 4.8, OpenAI, GPT-5.6, GPT-6, Codex, DeepSeek, Polymarket, 8090, Google, Google Cloud, Gemini, Chromium, Chrome, Tesla, SpaceX, Waymo, Apple, Amazon, AWS, Microsoft, Yahoo, Netscape, Mozilla, Firefox, Apache, OpenRouter, Thinking Machines, Cursor, Composer 2, LibGen, Perplexity Computer, Grok, Neotron, GLM, Lovable, ElevenLabs, The New York Times, Westlaw, ROSS Intelligence, Airbnb, Democratic Socialists of America (DSA)
## Video description
(0:00) Bestie intros!
(0:18) The fight to save open source AI: Kimi K3 panic, Anthropic/OpenAI regulatory capture
(27:38) Anthropic/OpenAI historic growth rates, China's long game
(48:29) Anthropic's $1.5B piracy settlement and the great IP theft hypocrisy
(1:07:12) Google and Tesla stocks tank on surging AI capex
(1:17:19) Socialism Corner: "Evictions = Violence" and the threat to private property rights
Apply for All-In Summit 2026:
https://allin.com/events
Follow the besties:
https://x.com/chamath
https://x.com/Jason
https://x.com/DavidSacks
https://x.com/friedberg
Follow on X:
https://x.com/theallinpod
Follow on Instagram:
https://www.instagram.com/theallinpod
Follow on TikTok:
https://www.tiktok.com/@allin
Follow on LinkedIn:
https://www.linkedin.com/company/allinpod
Intro Music Credit:
https://rb.gy/tppkzl
https://x.com/yung_spielburg
Intro Video Credit:
https://x.com/TheZachEffect
Referenced in the show:
https://artificialanalysis.ai/models/kimi-k3?intelligence=artificial-analysis-intelligence-index#intelligence-comparisons
https://www.axios.com/2026/07/20/ai-us-china-open-source-kimi
https://www.wired.com/story/the-white-house-is-trying-to-figure-out-what-to-do-about-chinese-ai
https://x.com/mkratsios47/status/2079933645888880708
https://polymarket.com/event/us-government-bans-an-open-source-ai-model-in-2026-20260703221501747
https://fortune.com/2025/01/29/deepseek-openais-what-is-distillation-david-sacks
https://stratechery.com/2026/whos-afraid-of-chinese-models
https://x.com/CommerceGov/status/2080341953086886387
https://x.com/tickerplus/status/2080123562560504240
https://blog.tickertrends.io/p/anthropic-vs-openai-arr-tracking
https://x.com/sama/status/2077817060068057493
https://www.reuters.com/world/us-judge-approves-anthropics-15-billion-settlement-copyright-lawsuit-2026-07-20
https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks
https://techstartups.com/2026/07/22/nearly-200-silicon-valley-startups-urge-trump-not-to-ban-chinese-ai-models-warn-it-could-kill-innovation
https://x.com/typesfast/status/2080339052398891244
https://www.thefp.com/p/why-new-york-city-has-50000-ghost
#allin #tech #news
[Source](https://www.youtube.com/watch?v=wcV0SRPFK9s)
Source quotes
[2:00] David Sacks: 这个辩论完全搞反了，本该是 Anthropic 交代为什么防盗做得如此糟糕，现在却要全美开源生态来为 Anthropic 的失职买单。
[10:18] Chamath Palihapitiya: 如果美国政府真的介入并对开源模型下达禁令，最终结果将是直接砸穿美国股市。
[13:38] David Sacks: 既然能买 50 美分的，为什么非要买 50 美元一罐的可乐？
[15:12] David Friedberg: 这并不意味着我们在偷窃他们的算法，我们没有黑进他们的服务器盗取软件代码，我们只是分析了他们软件输出的结果，并用它来改进我们自己的软件。
[20:10] David Friedberg: 那些想控制专有技术收租的 1% 寡头们，他们的好日子到头了，我们甚至根本不需要设立所谓的财富税。
[33:24] Jason Calacanis: 让老子说完这 30 秒，闭嘴听我说！
[60:50] David Sacks: 如果未经同意使用模型输出是偷窃，那么 Anthropic 和 OpenAI 当初未经许可抓取全网创作者的版权内容来训练模型，按此逻辑岂不等于在法庭和监管机构面前主动承认了‘自己的整个产品本身就是偷来的’？
[65:34] Chamath Palihapitiya: 如果你跑去华盛顿与恶犬同眠，就别惊讶醒来时满身跳蚤。
[71:50] David Friedberg: 那你就根本不该持有这只股票，自然会有别人来接盘。
[79:03] David Friedberg: 一旦社会承认财产不再像上帝的法则一样神圣，并且没有法律和公共正义的力量来保护它，无政府状态与专制暴政便随即拉开序幕。
[83:25] David Sacks: 住在封闭式门禁社区里的私募股权合伙人太太们会觉得自己品德高尚，因为她们成功阻止了驱逐，但真正承受代价的，恰恰是楼里走投无路的平民百姓。
Source quotes
[3:04] David Sacks: If stopping distillation was their primary objective, Anthropic would push to ban Chinese access to American models, not American access to Chinese models.
[8:13] Chamath Palihapitiya: If the United States government intervenes, it will tank the stock market. Period. Not debatable.
[20:15] David Friedberg: If open-source AI takes off, then all the worries that Bernie Sanders and Elizabeth Warren and all the socialists are harping and larking about are not going to be the case anymore... AI becomes a power for good for creating an open economy.
[24:03] Chamath Palihapitiya: I've never seen it in my 25 years in Silicon Valley where a sector of the economy can absorb hundreds and hundreds of billions of dollars... and economic pricing power effectively evaporates in months.
[39:28] David Sacks: You take things that are in the public domain, you fork them, you make them your own... An American company has taken open-source contributions in the public domain, made it their own, and then developed their own model. And if you say that American companies can't do that or that somehow it's tainted with IP theft, you are basically going to put a dagger through the heart of the entire American open source ecosystem.
[50:47] David Sacks: They believe it is fair use to train their models and derive their own weights based on fair use. However, they say that the one type of content that you should never be able to train on is *their* output... IP for we, but not for thee.
[62:50] David Sacks: If you go to Washington and you lay down with the dogs, don't be surprised if you wake up with the fleas... they may have poked the tiger.
[69:32] Chamath Palihapitiya: When you are a machine and a business model that compounds money at 32% over a 20-year average, you give these guys the benefit of the doubt. These are not people that are flying fast and loose.
[77:36] Activist: emboldening us so that we no longer tolerate the violence of evictions as a matter of business as usual.
[78:57] John Quincy Adams: the moment the idea is admitted into society, that property is not as sacred as the laws of God... anarchy and tyranny commence.
Original transcript
[0:00] All right, everybody. Welcome back.
[0:02] Episode 282 of the world's greatest
[0:05] podcast.
[0:06] It's your podcasters's favorite podcast.
[0:08] It's your mom's favorite podcast. It's
[0:10] the All-In podcast. With me again, David
[0:13] Sax up to you, David Freeberg. It was a
[0:15] big week. It was a big week. Uh the
[0:18] continuing number one story in the world
[0:20] is Kimmy K3. It's sparked a debate about
[0:24] banning Chinese open-source models here
[0:27] in the United States and it's gone all
[0:30] the way to the White House last Friday.
[0:31] We talked about it here. China's
[0:33] Moonshot AI released Kimmy K3 open
[0:36] source model. Obviously performance on
[0:38] par on par not 6 months behind not 12
[0:41] months behind but now on par with models
[0:44] like Opus 4.8 and GPT 5.6 six, which in
[0:48] and of itself is extraordinary, but
[0:50] about 50% cheaper and uh this has
[0:53] created a bit of a panic similar to the
[0:55] Deep Seek moment that we had here back
[0:57] in early 2025.
[1:00] The White House, David Saxs, has gotten
[1:03] involved. Michael Katzio, friend of the
[1:05] show, said, quote, "We have information
[1:09] that Moonshot AI distilled anthropics
[1:12] fable for the development of its K3
[1:14] model. Here's how the Trump
[1:16] administration has reacted so far.
[1:18] Monday, Axios reported the White House
[1:20] was considering banning Chinese open
[1:22] source models. A couple of weeks ago,
[1:24] David, I I think it was three weeks ago,
[1:25] I I gave that to you as a hypothetical,
[1:28] and here we are. On Wednesday, Wired
[1:30] reported that Howard Letic,
[1:33] friend of the show, does not want to ban
[1:35] Chinese models. So apparently palace
[1:37] intrigue there might be different
[1:38] opinions inside Trump's white house and
[1:41] instead they want to incentivize more US
[1:43] frontier labs to develop better open
[1:46] source models. Poly market says 45%
[1:48] chance US government bans an open source
[1:50] model in 2026. Uh that was a brand new
[1:53] market started just it was at 22% a
[1:56] couple days ago. Saxs um you called it
[2:00] two months ago. Our first victory fap of
[2:03] the episode. I think where it's all
[2:05] leading to is an effort to ban
[2:07] opensource models. There's a lot of
[2:09] breadcrumbs leading here. If you look at
[2:11] a lot of the rhetoric around how models
[2:13] need to have guard rails and that with
[2:15] open source models, the guardrails can
[2:17] be removed and therefore they're
[2:19] dangerous. You see this rhetoric already
[2:20] in Anthropics blog posts. Any threat
[2:23] that they describe, they kind of go out
[2:25] of their way to take that shot at open
[2:27] source models. I think again they're
[2:29] trying to create ideas or put predicate
[2:33] facts in the public record to justify an
[2:36] action later on. I think it's just a
[2:38] matter of time before they feel like
[2:40] they're at a position where maybe they
[2:41] can push for that type of ban directly.
[2:44] >> All right, there it is. Sachs,
[2:48] what's going on at the White House? What
[2:50] is the administration's position here?
[2:52] Why are we getting multiple? Is the
[2:55] White House testing and and probing to
[2:58] figure out what their position is here
[3:00] or is it just this is a super dynamic
[3:02] situation? What's going on?
[3:04] >> Well, look, I mean, I have it on good
[3:05] authority that there is no decision by
[3:07] the White House to ban open- source
[3:09] models and I think they want that known.
[3:12] I think there's an ongoing conversation
[3:14] happening around what to do about
[3:18] Chinese distillation and we should talk
[3:20] about that. But no decision has been
[3:22] made and the president listens to a
[3:24] course of voices. He wants to get advice
[3:26] from as many people as possible. And I'm
[3:29] confident that if everybody weighs in
[3:32] that the president will make the right
[3:33] decision as he always has with these
[3:35] tech issues. I think his instincts have
[3:37] been absolutely impeccable on this and
[3:39] he's always supported a let's say
[3:42] lighter regulation more open approach
[3:45] and that's why I think the US is winning
[3:47] the AI race. So I think that's kind of
[3:50] where where things stand. Um
[3:51] >> where do you stand? Where do you stand
[3:53] Sax? That's what everybody wants to
[3:55] know.
[3:55] >> Yeah, I think it's important for me to
[3:56] make my opinion known in the spirit of
[4:00] again contributing my voice so the
[4:02] president hears all perspectives and
[4:03] then can make the best decision. Look, I
[4:06] think it would be a tragic mistake if
[4:09] the government were to take action
[4:10] against the open source ecosystem. That
[4:14] would do nothing but hurt America's
[4:16] position in this AI race. It would
[4:18] backfire badly. And I think that the key
[4:22] point here is that regardless of what
[4:24] you think about distillation, you cannot
[4:27] punish American developers for it. So,
[4:31] you know, you can't say that American
[4:33] companies and American developers can't
[4:35] use Chinese contributions to the public
[4:38] domain. That's just cutting off our nose
[4:40] to spite our face. I mean, obviously,
[4:42] >> American companies have to be able to
[4:44] use everything that's in the public
[4:46] domain because the rest of the world
[4:48] will be using those things. And let me
[4:50] just say, I've said for a while that
[4:52] Anthropic is guilty of regulatory
[4:54] capture, of attempts to seek government
[4:56] protection. This is a company that is
[4:58] the fastest growing tech company at
[5:00] scale in history. They started the year
[5:02] at 10 billion of ARR. They're now over
[5:04] 70 billion of ARR. This is not a company
[5:06] that needs government protection. This
[5:08] is not a company that is under threat
[5:10] from competitors or or whether they're
[5:12] Chinese or otherwise. And yet they have
[5:15] been very successful at trying to panic
[5:17] everybody into thinking that they need
[5:19] some sort of government protection. And
[5:22] the tell on this, the way that you know
[5:24] that this whole distillation thing is
[5:27] fake is because if stopping distillation
[5:31] was their primary objective, Anthropic
[5:34] would push to ban Chinese access to
[5:36] American models, not American access to
[5:38] Chinese models.
[5:39] >> Yes, they could. And that is that is
[5:41] achievable. They could block it. So
[5:43] >> they're the ones in the best position to
[5:45] block it. If distillation, if industrial
[5:47] scale distillation is a national
[5:50] security threat, they're the ones who
[5:52] need to stop it because that is the
[5:54] place where distillation occurs. You
[5:56] have to stop it at the source. Once you
[5:58] allow Chinese companies to distill, the
[6:00] horse is out of the barn.
[6:01] >> Yeah.
[6:01] >> And the reality is that I think that
[6:04] what's happening here is that in their
[6:07] lust for growth, Anthropic has done a
[6:10] very poor job at stopping distillation.
[6:13] I mean they're saying that distillation
[6:15] is occurring at industrial scale. Okay,
[6:17] if it's industrial scale, it must be
[6:19] pretty obvious to see. So is waves of
[6:23] accounts being created by students
[6:25] rolled up and sold on the dark web, you
[6:29] know, in those kind of channels. So you
[6:31] got people in Manila, in the
[6:32] Philippines, I understand, and India
[6:34] signing up for all those accounts and
[6:36] then sending them to the dark web and
[6:38] selling them using IP addresses from
[6:40] America. Yeah. So that that's how it
[6:41] occurs. But the more the more industrial
[6:43] scale it is, the more obvious it is to
[6:45] see. And Tamatha has been saying for a
[6:47] while, why don't you KYC your customers?
[6:50] Well, they know that if they KYC their
[6:53] customers, it'll slow their growth. So
[6:54] instead, what they're saying is, hey,
[6:56] ban our competitors. Well,
[6:58] >> that's ridiculous. I mean, they're in
[7:00] the best position to stop the
[7:02] distillation. I think that they're
[7:04] negligent about doing that. or I mean if
[7:06] they really think it's that big a
[7:07] threat, they should use a few points of
[7:09] their 90% gross margins to do that. What
[7:12] you don't do is then say that American
[7:14] developers cannot use everything that's
[7:17] in the public domain. So it seems it
[7:20] seems to me that this debate is all
[7:22] backwards that the question should be on
[7:25] Anthropic to explain why it's doing such
[7:27] a bad job, not on the whole American
[7:30] open source ecosystem to be punished for
[7:33] Anthropic's failure. All right, Freberg,
[7:35] I have a really good question for you,
[7:36] but before we do that, Shimath, can you
[7:38] give me
[7:40] maybe a little bit of what you're
[7:42] hearing on your sales calls for 8090?
[7:44] You're talking to enterprises. They're
[7:46] hearing all these reports, whether it's
[7:48] you, Daario, this pod, other places
[7:50] talking about, hey, open source is
[7:52] ready. This is the moment. Get control,
[7:54] AI sovereignty, etc. They must be
[7:56] calling you up and saying, hey, okay,
[7:58] we're ready. Like, how do we get these
[8:00] things on prem? How do we do it? So
[8:01] what's the what's happening on those
[8:03] calls you're doing with the enterprise
[8:05] and then can you maybe give people an
[8:07] idea of what distillation is just and
[8:11] and why it's so important here.
[8:13] >> Let's start with the second thing.
[8:14] Distillation
[8:16] is when you
[8:18] fire up a model and you ask it a
[8:21] question and you observe it and you take
[8:26] its output and you use that in training
[8:28] of your own model. Now multiply that
[8:30] behavior by tens of millions
[8:34] and what you exfiltrate is essentially
[8:37] trillions of questions and answers.
[8:40] And Sax is right. If you really care
[8:44] about distillation,
[8:47] you implement KYC. You force people to
[8:51] make an account, not just with a
[8:54] username and a password, but with some
[8:56] form of identification,
[8:58] maybe with a bounded credit card.
[9:00] There's all kinds of steps that you can
[9:02] take that would frankly slow things down
[9:05] in terms of revenue traction, but would
[9:08] solve the distillation problem on its
[9:10] face. So, that isn't really a thing.
[9:13] It's a bit of a red herring. The other
[9:15] thing on distillation is everybody has
[9:18] at some point distilled. The question is
[9:21] who is distilling from whom? And it
[9:23] looks like that funny meme where there's
[9:26] like nine Spider-Man all pointing at
[9:27] each other. That's what this is because
[9:31] Anthropic has distilled from all of
[9:35] these publishers. They just pay a $ 1.5
[9:38] billion fine. Apparently, Open AI
[9:40] distilled from the New York Times.
[9:41] There's still an ongoing lawsuit. the
[9:44] Chinese labs have distilled from
[9:45] anthropic.
[9:46] >> It's wholesale stealing everywhere.
[9:48] Yeah.
[9:48] >> Well, I don't want to call it stealing
[9:50] because it's not clear who actually owns
[9:52] the copyright in the first place. But
[9:54] here should be the important
[9:56] observation.
[9:58] These models are getting commoditized
[10:01] much faster than anybody thought.
[10:04] And how do we know this? Because there
[10:07] is no meaningful sustained advantage
[10:09] once a model publishes their performance
[10:12] criteria. What you see is literally
[10:14] within weeks other models some open some
[10:18] closed some open weight who are able to
[10:20] match and in some cases exceed the
[10:22] performance. So I think what's happening
[10:25] here is a handful of American companies
[10:27] have realized whoa this value that we
[10:32] are seeing today may not be sustainable
[10:35] in a 5 and 10 year period and when you
[10:37] go and present a business model to Wall
[10:38] Street you need to have that certainty
[10:41] otherwise it impacts your valuation and
[10:43] so I think a lot of what's happening
[10:44] right now Jason is a valuation
[10:46] preservation game by the closed frontier
[10:50] labs because if you actually understood
[10:53] how commoditized these things are
[10:54] becoming and the velocity at which it's
[10:58] happening. You see that the real
[11:01] business model is not in the
[11:03] foundational model anymore. It's at the
[11:05] application layer above and it's in the
[11:07] infrastructure below whether that's the
[11:09] cloud or whether that's chips. And so I
[11:12] think in the absence of regulatory
[11:14] intervention and in the absence of the
[11:17] United States government stepping in to
[11:18] put their thumb on the scale, what will
[11:21] happen is that as people learn about how
[11:24] value is changing, they're going to put
[11:26] more value in the application layer and
[11:28] more value in the infrastructure layer.
[11:29] That is bad for closed frontier labs,
[11:31] especially when they're mispriced 25 to
[11:34] 50x the open alternative. And so this is
[11:37] an attempt to stop a competitor that is
[11:40] of the same quality but just much
[11:42] cheaper. Now there's another important
[11:45] thing here which is if the United States
[11:48] government intervenes it will tank the
[11:50] stock market.
[11:51] >> Okay?
[11:52] >> Period. Not debatable. Now you can
[11:55] debate which companies get tanked and we
[11:58] we can probably play that scenario out.
[12:01] But for example, if they said no more
[12:03] open source, American companies cannot
[12:06] use open source. Okay, let's just take
[12:08] at it from a stock perspective. Let's
[12:09] take an average normal company,
[12:11] Coca-Cola. Hey, Coca-Cola, you're trying
[12:14] to use AI to improve your business. You
[12:17] know what? You can only use these two
[12:19] options. And those things cost 50 to 100
[12:22] times more than your other best
[12:24] alternative that you may use otherwise.
[12:27] that will eventually show up in your
[12:29] costs because AI is supposed to be this
[12:32] incredible thing that just kind of
[12:34] solves every problem and does everything
[12:36] for you. And so this incredibly
[12:38] important input into your cost model is
[12:42] now orders of magnitude multiples
[12:44] greater than your competitors that are
[12:46] outside the United States simply because
[12:48] you're in the United States. So what
[12:51] would the capital markets do? They're
[12:52] going to say, "Wow, you have a crazy
[12:54] cost structure. This doesn't make sense.
[12:56] you're forced to absorb costs that
[12:58] aren't rational nor market driven. So
[13:01] then Coca-Cola has to get rerated. But
[13:03] then you look at the people who are
[13:05] selling those tokens and this is where
[13:06] anthropic and open AAI need to
[13:08] understand. If the government comes in
[13:10] and actually tells you that there's no
[13:12] open source, their valuation will
[13:14] crater. Why? Because all of that revenue
[13:17] is artificially being propped up.
[13:20] >> It's not being driven by market demand
[13:22] where you're being forced to compete.
[13:23] It's because of regulatory capture where
[13:26] you now get an artificial constraint,
[13:28] but it only works in one market. And so
[13:31] anyways, all roads lead to market chaos.
[13:34] >> Love it. Yeah.
[13:35] >> If anybody gets involved, so we should
[13:37] just not get involved. What it sounds
[13:39] like is you're saying that American
[13:40] enterprises will pay a token tax if the
[13:44] government gives anthropic and open AI a
[13:47] government enforced duopoly
[13:49] >> and enterprises are no longer free to
[13:51] use open source like the rest of the
[13:53] world. Yes.
[13:54] >> Yeah. The markets with the cap.
[13:55] >> We will put ourselves on an island.
[13:57] Yeah.
[13:57] >> We'll be on an island of overly
[13:59] expensive AI.
[14:01] >> You can have CocaCola or Pepsi or
[14:03] Coca-Cola or Pepsi.
[14:04] >> Yeah. Well, it's not your beverage
[14:06] choices. You have two beverage choices,
[14:08] but they cost 50 times more than Coke
[14:11] outside of America. This is the point.
[14:13] We already have Coke everywhere. So,
[14:16] >> you can buy 50 cent coke or $50 Coke.
[14:19] Why would you buy $50 Coke when you can
[14:20] buy 50 cent coke?
[14:21] >> Yeah.
[14:21] >> All right. Let's get Let me get free
[14:23] here. Yeah.
[14:24] >> Okay. I was going to say just just just
[14:25] one thing on this. This goes back to my
[14:27] point of these proposals to ban open
[14:30] source that are coming from anthropic.
[14:32] They don't solve the distillation
[14:33] problem. If dissolation is a problem,
[14:35] you have to stop it at the source. In
[14:37] other words, if you want to ban open
[14:39] source in America, the rest of the world
[14:42] will still be using Chinese open models.
[14:44] We want to solve that problem.
[14:46] >> Yes. And that means they'll get the data
[14:48] and they'll get the reinforcement
[14:49] learning and then we lose the AI race
[14:51] guaranteed. Freedberg,
[14:53] let's take it from a Graham Allison and
[14:56] level up the discussion here. Would be
[14:58] quite provocative to ban the Chinese
[15:01] models. How does Xi Jinping respond to
[15:03] that? How does the CCP respond to that?
[15:05] This is a crazy chessboard. It would
[15:07] seem like a pretty escalatory and we'll
[15:09] be going up the ladder. Yeah, Freeberg.
[15:12] >> Yeah, I look I don't I don't think it's
[15:14] as relevant that the open source model
[15:16] is published by China yet that there may
[15:18] be security risks, but those can be
[15:20] estimated and addressed. I think on the
[15:22] three things that are worth highlighting
[15:23] on this this issue, I'm I'm very much
[15:25] aligned with Sax and Chim and I think
[15:27] the three things are really around
[15:28] distillation. And I don't think
[15:29] distillation
[15:31] is just about AI. You know, distillation
[15:34] is a process whereby you look at the end
[15:37] product that someone else has produced
[15:40] in thinking about and learning about how
[15:42] to engineer your product. It is a common
[15:46] technique that is used across every
[15:48] product category in every industry. One
[15:51] car maker will look at how the other car
[15:52] maker's car operates and they will use
[15:54] that to help them design a better car.
[15:56] You know, at Google in the early days,
[15:59] we would submit millions of search
[16:01] queries to Yahoo and Microsoft search
[16:04] engines to see what the result sets were
[16:06] and we would compare our results against
[16:08] their results as a way of improving our
[16:10] search engine rankings and our
[16:11] algorithm. It was a very common
[16:13] technique. It doesn't mean we were
[16:14] stealing their algorithm. We didn't go
[16:16] into their servers and steal their
[16:18] software. We looked at the output of
[16:20] their software and use that to improve
[16:21] our software.
[16:22] >> It was called benchmarking, right? It
[16:24] was benchmark. You can call it there's
[16:26] been a million terms. Exactly. Right.
[16:27] And so I don't think that this matters
[16:28] as much. I think the question around
[16:30] copyright infringement or IP
[16:32] infringement
[16:34] >> Sax is right. There's a terms of service
[16:36] question here, but that's on the service
[16:37] providers to fix the terms of service
[16:39] blocking people from doing this if they
[16:41] so chose.
[16:42] >> But the copyright argument, the IP
[16:44] argument is really about did they steal
[16:45] the software? They didn't steal the
[16:47] software and they just looked at the
[16:49] output.
[16:50] >> That's not IP infringement. That's not
[16:52] copyright infringement in the classical
[16:53] sense. And so I do think from a
[16:55] distillation IP argument perspective,
[16:58] it's the output, not the process that
[17:00] matters. So the question is, are they
[17:02] taking copies of copyrighted software
[17:04] and using it or are they looking at the
[17:07] output? And so I think output not
[17:09] process of engineering is what you
[17:10] really need to assess here.
[17:12] >> Can I insert something free? comment on
[17:13] it is I think a lot of people in the
[17:15] let's say policym community don't
[17:17] understand this distinction but I think
[17:18] everybody in tech does and I think it's
[17:20] a big part of why there's a disconnect
[17:22] on this is there's a huge difference
[17:24] between model weights and outputs right
[17:27] so the weights are the it's the file of
[17:31] numbers it's the numerical parameters in
[17:34] the software code that's the code
[17:35] >> that's the software that's the code and
[17:37] if Chinese companies were to steal
[17:40] >> weights proprietary weights from anthrop
[17:42] ropic or open AI that would be theft.
[17:45] Okay, but that's not what we're talking
[17:47] about here because no one's accused
[17:48] that. What we're talking about here is
[17:51] taking model outputs and then trying to
[17:54] learn from them and
[17:55] >> using other people's software to learn.
[17:57] >> Yes. And it's exactly the situation you
[17:58] said with like the the Google searches
[17:59] or whatever. And here's the thing that's
[18:01] so hypocritical is that OpenAI and
[18:04] Anthropic have both argued that they are
[18:07] free to train on all the world's output
[18:10] regardless of whether the creator wants
[18:11] them to or not.
[18:12] >> That's right.
[18:13] >> That is their current position. That's
[18:15] like Chimath mentioned the New York
[18:16] Times lawsuit. The New York Times is
[18:18] suing suing OpenAI right now
[18:20] >> for going onto the New York Times
[18:22] website in violation of the New York
[18:24] Times terms of service, scraping all the
[18:27] information and training on it.
[18:29] >> And Open AI's argument is look, we're
[18:31] not stealing anything. We're taking the
[18:33] output of the New York Times and we are
[18:35] deriving our own model weights. And that
[18:38] is exactly what these Chinese models are
[18:40] doing is is they are taking the output
[18:42] of American models and then they're
[18:44] deriving their own weights. They're
[18:46] learning from it.
[18:46] >> Yeah. And so the accusation though, just
[18:48] so we're clear here, is that
[18:50] >> there's ethical issues around the
[18:53] industrial scale covert breaking of
[18:56] terms of service. That's what the White
[18:57] House has been terms of service as well.
[19:00] So just so people understand like there
[19:02] is a bit of recognition of that.
[19:04] >> Look, let me let me be clear that I'm
[19:06] not defending China in this. In fact,
[19:09] you know, look at my my bonafides. I was
[19:11] the first administration official to
[19:13] even talk about distillation. I did it
[19:15] in January of 2025 when deep sea came
[19:17] out. I went on Laura Ingram and I think
[19:19] I was probably the first person in the
[19:20] government to even explain this concept
[19:22] publicly to people and moreover, you
[19:25] know, I was a co-author of the winning
[19:26] the AI race report in which the whole
[19:28] premise of it was that we want to win.
[19:30] We want to beat China. So, you know, I'm
[19:33] definitely not someone in this camp that
[19:35] doesn't want the US to win. I want the
[19:37] US to win. The question is how. And if
[19:39] we shoot ourselves in the foot by
[19:41] banning open source, which is to say not
[19:44] letting all of our American companies
[19:45] take advantage of open source when the
[19:47] rest of the world is able to, then that
[19:50] is a huge problem. Now, I'm fine with
[19:53] Anthropic and Open AI enforcing their
[19:55] terms of service. They need to do a
[19:57] better job, stop the distillation from
[19:58] occurring in the first place. If there
[20:00] are things that the government can do to
[20:02] help them, okay, but I'm not sure what
[20:04] those things are. What needs to happen
[20:06] is those companies need to do a better
[20:07] job enforcing their terms of service.
[20:09] >> Yes.
[20:10] >> Okay. Now, Freeberg, you had you I think
[20:12] you said you had three points to make. I
[20:13] think you made one. I want to get the
[20:14] other two out of you.
[20:15] >> Yeah. So, the other one was just on the
[20:16] free speech argument, which is look,
[20:18] what is an open-source model? And I
[20:20] think the audience needs to understand
[20:21] this if you're not from the software
[20:22] industry, but open source is a
[20:24] downloadable package of software. You
[20:26] can think about it as downloading a a
[20:28] text, a book. You just got all the code.
[20:31] Once you get the code, you've got it on
[20:33] your computer. You don't have to be
[20:34] connected to the internet. You can just
[20:35] run it and use it. And I think part of
[20:38] the challenge that's going to be faced
[20:40] here, if there is any attempt at
[20:41] restricting open source, it's going to
[20:43] open a can of worms on how do you
[20:45] actually enforce restrictions on open
[20:47] source because you're basically telling
[20:48] people once they've downloaded and
[20:50] gotten a copy of this free publicly
[20:52] available software, they're not allowed
[20:54] to use it. And that becomes a real
[20:56] challenge. I don't think we have a lot
[20:58] of great precedent for that. It's going
[20:59] to be very ugly to try and stop open
[21:00] source. I do think there's a question on
[21:02] like copyright action but if there is
[21:04] there's a legal due process to go
[21:06] through to make that case and stop that
[21:09] open source from being available and
[21:10] then my third point is just open source
[21:12] is better for the world to Chimath's
[21:15] point this is 100 times cheaper that is
[21:18] better for the industry for enterprise
[21:21] the beneficiaries are going to be the
[21:22] economy the consumer fundamentally if
[21:25] you look back on the internet in the
[21:27] early days Netscape made a proprietary
[21:29] browser and a proprietary server
[21:31] software, the Netscape software, and
[21:33] that they went public, and they were the
[21:34] first to do this, and it was super it
[21:36] was super valuable and profitable. That
[21:38] company ended up getting crushed because
[21:40] of open- source. The Mozilla Foundation
[21:42] was formed to create an open-source web
[21:45] browser called Firefox, and then Google
[21:47] ended up hiring everyone and made it
[21:49] Chrome, but it was still open source.
[21:51] The Apache Foundation set up the first
[21:53] HTTP server as an open-source product.
[21:56] Rather than having to pay Netscape or
[21:58] Microsoft or Oracle for their server
[22:01] software, anyone with a computer could
[22:03] download the Apache software and make a
[22:06] web server and be on the internet and
[22:07] create a website. And what ended up
[22:09] happening is the value acrewed to the
[22:12] internet. It didn't acrue to the small
[22:14] number of software providers that
[22:16] controlled the gate and the portal of
[22:18] the internet. Basically, everything got
[22:20] open sourced and Google took off and
[22:22] eBay took off and Etsy and Amazon and
[22:24] all the millions of small websites and
[22:26] all the millions of small businesses and
[22:28] everyone that benefited from an openly
[22:30] accessible open-sourced internet. If the
[22:33] internet was closed and there was
[22:34] proprietary software gates and portals
[22:36] throughout the internet that everyone
[22:38] had to pay to get through, the internet
[22:39] would not have taken off and the economy
[22:41] wouldn't have grown and all these jobs
[22:43] would have been called AOL and Coffee
[22:44] Serve like and now now when we look at
[22:48] this analogy, the analogy here is if
[22:50] open-source AI takes off, then all the
[22:52] worries that Bernie Sanders and
[22:54] Elizabeth Warren and all the socialists
[22:56] are harping and larking about are not
[22:58] going to be the case anymore because
[23:00] you're not going to see all the value of
[23:01] AI acrue to two or three or four
[23:03] companies and their small group of
[23:05] billionaire shareholders. What will
[23:07] happen is AI proliferates and a million
[23:10] AI integrated enterprises all over the
[23:14] world will benefit. Everyone will
[23:15] benefit. The economy will grow, jobs
[23:18] will be created and AI becomes a power
[23:20] for good for creating an open economy.
[23:22] So the I know that some people that are
[23:24] listening to this in the government and
[23:26] that are on the other side are going to
[23:27] say but but but Chinese open source
[23:30] versus American open source let it all
[23:32] proliferate. And frankly if the Chinese
[23:34] are violating copyrights, stealing
[23:35] software, go after them for that. Put in
[23:38] place trade sanctions. Do all the you
[23:39] can do to stop that from happening. But
[23:41] fundamentally open- source AI will
[23:43] transform the global economy and it will
[23:45] ensure that the economic value of AI
[23:48] will diffuse to everyone and not be held
[23:51] captive by a small number. Why the 1%
[23:53] that controls open source if they have
[23:56] it the oligarchs are going to get their
[23:59] clocks rung and we don't need to have a
[24:00] wealth tax. Chamathy we're going to add
[24:02] to this.
[24:03] >> We have to acknowledge that it's
[24:06] incredible how fast the value capture at
[24:10] this segment of the market has basically
[24:12] evaporated. I've never seen it in my 25
[24:16] years in Silicon Valley where a sector
[24:18] of the economy can absorb hundreds and
[24:20] hundreds of billions of dollars
[24:24] and then you think that there's going to
[24:26] be economic pricing power many decades
[24:29] into the future and it effectively
[24:32] evaporates in months. Months.
[24:34] >> It took off in months. Remember remember
[24:37] and it's evaporated in months.
[24:38] >> No, I got this is an area where I
[24:39] disagree with you guys.
[24:40] >> Evaporate is strong term. It it does
[24:42] look like it could slow down or it could
[24:45] or or it could plateau.
[24:48] >> Go ahead, Sax. Make your argument and
[24:49] I'll give you my argument.
[24:50] >> Yeah, but let me pull up the anthropic
[24:52] revenue chart before you go there, Saxs,
[24:53] because we can this will help mitigate
[24:55] it here and educate the audience. So, as
[24:57] you can see here, we got a little bit of
[24:59] a stall in anthropics revenue uh in the
[25:02] last couple of months. And it seems and
[25:05] and this is third party tracking.
[25:08] So it's it's not perfect data, but we do
[25:10] see that this is a a clear headwind. And
[25:13] then do you have the second chart I had?
[25:15] Um I talked on the pod just uh it was
[25:18] last week or the week before when we're
[25:20] having the discussion about open source
[25:22] and like I think it was um Brad
[25:25] Gersonner uh from alimter was talking
[25:27] about hey tokens are still growing. Well
[25:29] there are routers that track open
[25:31] router. the better an open source model
[25:33] does and the easier it is to implement
[25:34] on your own servers and take it in-house
[25:36] etc. Those are dark tokens. They're not
[25:39] recorded
[25:40] and you're not going to see them show up
[25:41] on a revenue chart anywhere because
[25:43] they're free essentially. You just need
[25:45] to have servers and energy to to do
[25:46] them. And here you see that now well
[25:48] over 50% and um a lot of these are
[25:51] coming from Chinese.
[25:53] >> By the way, Sax, I want to be clear
[25:55] before you give the counter. I'm not
[25:57] saying that these companies won't make
[25:59] money. That's not what I'm saying.
[26:01] But what I am saying is that markets are
[26:04] very savvy in looking through current
[26:07] earnings and asking a very specific
[26:09] question which is what does this revenue
[26:11] look like 10 years from now? Does it go
[26:13] up? Does it go down? Is there more
[26:15] competition or is there less
[26:16] competition? Is it effectively
[26:18] monopolistic or is it more of a
[26:20] commodity? If it's a commodity, how many
[26:22] people can price this good? At what
[26:25] price is the market clearing price of
[26:26] that good in 10 years? And all I'm
[26:28] saying is normally those variables get
[26:32] exposed. Those cards get turned over
[26:35] relatively slowly. And so you have 5 10
[26:38] year cycles to transition from being an
[26:42] exclusive provider of a good to
[26:44] effectively a commodity provider of a
[26:45] good. And all I'm observing is it's so
[26:48] unique that only technology could create
[26:51] a market where that cycle could get
[26:53] compressed into a few years because it
[26:55] is very hard if you're an allocator of
[26:57] money to sit there and look at this data
[27:02] and not wonder to yourself why it's not
[27:04] a commodity in 5 to 7 to 10 years. And
[27:08] when when they get to that conclusion,
[27:10] which every capital allocator will
[27:12] because it'll be pretty negligent to
[27:15] not,
[27:17] it's very hard to assign huge future
[27:19] premiums. And where the real money is
[27:22] going, by the way, and we saw it in
[27:23] Google's earnings, which I'm sure we'll
[27:24] talk about, it's going to the cloud.
[27:26] It's going to the infrastructure. So, by
[27:28] the way, there's another cohort of
[27:30] people that don't want to see the end of
[27:31] open source because they want to serve
[27:33] the cheapest models possible because
[27:35] they know that's where all the margin
[27:37] capture is.
[27:37] >> What's your Where's this going, Sax? Is
[27:39] it both open? We're going to just see a
[27:41] proliferation. I mean as far as I'm
[27:44] concerned there is an unlimited appetite
[27:46] for ondemand intelligence and there's
[27:48] going to be I don't think there's an
[27:49] upper bound for how much intelligence
[27:52] you can tap as long as it's continues to
[27:54] get be better which then the thesis
[27:56] would be yeah it's a commodity and the
[27:58] prices keep going down but consumption
[27:59] keeps going up. What's your thoughts on
[28:01] you have to assume that this is exactly
[28:03] the example that you guys used. It's a
[28:05] web browser and in a web browser it's a
[28:08] mechanism to get to a place and so the
[28:11] apps that are actually the places are
[28:14] where the value is captured and I think
[28:16] if you assume for a second that
[28:18] intelligence becomes completely
[28:20] ubiquitous. It's widely available. The
[28:23] marginal cost of it is effectively zero.
[28:25] The energy to generate it is effectively
[28:27] zero which I think is an accurate
[28:29] assumption. It's very hard to make the
[28:32] case of why this isn't like the browser
[28:35] >> and it becomes just gets subsumed into
[28:37] the Amazon web services cloud
[28:39] businesses, Elon's web service, etc. Sax
[28:42] where do you think this is going? Then
[28:43] we're going to go to our second story
[28:44] which is the IP story. Third story is
[28:46] going to be the markets and Google
[28:48] specifically and Tesla and SpaceX.
[28:49] >> Look, there's been a lot of violent
[28:51] agreement on this show so far. So, let
[28:53] me just make the counterargument. I
[28:55] think that both open source and closed
[28:57] source will be big winners in this. I
[28:59] think the market is huge and they each
[29:00] serve their purpose. What happened is
[29:02] with the introduction of committee K3,
[29:05] there was a little bit of a panic in
[29:07] which everybody said, "Oh my god, all
[29:09] the Chinese models have caught up and
[29:10] they're giving them away for free and
[29:11] they're going to destroy our leading
[29:13] American frontier labs and Anthropic and
[29:15] Open AI are running around saying,
[29:17] listen, we can't continue to invest
[29:18] billions of dollars if Chinese companies
[29:21] can just steal our weights." Right? So
[29:23] that's the argument that they're making
[29:25] that government officials are responding
[29:26] to. The truth of the matter is that look
[29:29] when Kimmy K3 first launched I was
[29:31] concerned about it because I was like oh
[29:33] has China caught up are they now able to
[29:35] produce a much cheaper Frontier model.
[29:38] Then the details started coming out. So
[29:40] Ben Thompson on his blog went through
[29:42] some of the cost numbers and it turns
[29:45] out that Kimmy K3 is not that much
[29:48] cheaper to run. There's not a
[29:50] significant cost advantage to it. So
[29:51] that was that's point number one. Point
[29:54] number two is that China has not caught
[29:57] up. It's true that Kimmy K3 scored
[29:59] really well on the arena battleground
[30:02] for front-end coding for web
[30:04] development, but that's just one test.
[30:06] That's just one dimension. There are
[30:08] areas where it scores really well, but
[30:10] there's lots of other areas where it
[30:11] doesn't score that well. So, it is not a
[30:14] clear advance or a clear catch-up to the
[30:16] leading American models. Moreover, you
[30:19] still have stuff in the labs by
[30:21] Anthropic and Open AAI that is way ahead
[30:24] of this and the reports are that Sam is
[30:27] going to Washington over the next week
[30:29] to go talk about GPT 6.0 which is
[30:32] blowing the doors off. So, I don't
[30:34] believe that China has really caught up.
[30:36] I think there's
[30:38] >> Sam's going to go see daddy and what is
[30:40] he going to ask for? I think there's
[30:41] incredible unreleased models in the
[30:43] pipeline and we need to let our horses
[30:45] run here and not slow them down with a
[30:48] bunch of unnecessary hoops and if we do
[30:50] that I think we're going to be just
[30:51] fine. I don't believe that China has
[30:53] caught up. I still think we are 6 months
[30:54] ahead. And then just the final point on
[30:56] this if you look at revenue which is the
[30:59] test of real usage in the real world
[31:02] anthropic and open AI are blowing the
[31:04] doors off. They are by far the fastest
[31:07] growing tech companies at scale that
[31:09] we've ever seen. Jason, you showed this
[31:12] chart that supposedly shows a hiccup in
[31:14] anthropic. Listen, their internal
[31:16] forecast was to 10x this year from 10 to
[31:18] 100. We're in the middle of the year.
[31:20] They're already over 70 billion of ARR.
[31:23] They're easily going to get to 100
[31:24] billion. And we don't really know what
[31:26] this little blip is here. I think one
[31:29] thing it might be is that open AAI, if
[31:32] you superimpose the OpenAI chart on
[31:34] this, they have reacelerated over the
[31:36] past month. So I think that if you were
[31:38] to add Open AI,
[31:40] >> Codex is excellent. Codex is
[31:41] >> Codex is excellent and I think they're
[31:43] taking a little bit of share and Sam is
[31:45] out there tweeting that we've got our
[31:46] mojo back and they're taking their
[31:48] forecast up. I think they were expecting
[31:50] to end the year at 60 billion of ARR and
[31:53] I think they're forecasting more like 75
[31:55] billion of exit ARR. So my guess is that
[31:57] if you were to superimpose open AI and
[32:00] anthropic and looked at them together
[32:02] and you were basically just to say that
[32:04] you know let's call it the frontier
[32:06] model duopoly in the US you do not see
[32:08] any slowdown you don't see any blip
[32:10] they're taking their forecasts up and
[32:12] the reality is that if distillation is
[32:14] going on it's been a thing for you know
[32:17] again I pointed it out back in January
[32:18] of 2025 with deep seat so it's been a
[32:21] thing this entire time that they've been
[32:23] growing exponentially so I just don't
[32:25] believe that these guys are actually
[32:28] suffering in any way. I don't think they
[32:30] need government protection. I think
[32:31] they're growing exponentially still.
[32:33] This is a little bit of a case of um you
[32:35] know what do you call it uh in
[32:37] basketball when a player flops you know
[32:40] that you did a foul flop or whatever.
[32:42] >> Yeah. It's a flop. Yeah.
[32:44] >> Well, or you you basically act super
[32:46] dramatic after a foul in order to draw
[32:48] the charge. You're foul baiting when
[32:50] you're trying to get a foul and then
[32:51] you're flopping which is just making
[32:53] like an exaggerated thing like LeBron
[32:54] does
[32:55] >> and that's exa Yes. It's a LeBron
[32:58] >> it's a flare flop, you know, whereop
[33:00] these guys are trying to they're trying
[33:02] to draw the foul. They're trying to get
[33:03] the government to intervene. Now, why
[33:05] are they doing this? Because they're in
[33:07] the middle of road shows right now. And
[33:09] Chamas, to your point, I do think they
[33:10] get the legitimate question about why
[33:12] won't you be commoditized over time by
[33:15] open source? And by far the best
[33:17] response to that would be that if they
[33:18] can lure the government into giving them
[33:21] a government protected duopoly, then
[33:23] that would be incredible.
[33:24] >> The best answer is what you gave and
[33:26] anthropic and open AAI should own this
[33:29] because they're good at it, which is
[33:31] they're going to go up the stack to the
[33:33] application layer.
[33:34] >> Well, they already done
[33:36] >> I know, but they've done it in this way
[33:37] which is a little ham-handed in some
[33:39] cases,
[33:40] >> but they're excellent at it. these
[33:42] enduser apps are really good and they
[33:45] should just own that. And that should be
[33:47] their answer to Wall Street, which is
[33:48] guys, we have the best model. We will
[33:50] eventually go up the stack. And if I
[33:52] were them, I'd actually practice the
[33:54] following answer. There may be a version
[33:56] of a model that I don't release and just
[33:57] keep for myself and I'll just use in my
[33:59] own applications. How about that?
[34:00] >> Yeah. Well, that would be
[34:01] >> that's the real answer. That's
[34:03] anti-competitive.
[34:04] >> This No, it's not. They're allowed to
[34:06] build a model and not release it and use
[34:08] it for themselves.
[34:08] >> Hold on. Hold on. Let me answer that. it
[34:10] it would be anti-competitive
[34:12] in the eyes of their customers. It might
[34:14] not be in the governments, but if you
[34:15] are using them as a customer and you're
[34:17] lovable, which I talked to, and they're
[34:20] paying a ton of money or you're 11 Labs
[34:22] and you're paying them a ton of money
[34:23] and they say, "Hey, we got our latest
[34:25] and greatest. You can't use it cuz we're
[34:26] going to compete with your company."
[34:27] They would stop using it and they go to
[34:29] open source. I'll tell you why so I
[34:30] think you're wrong on this issue. I
[34:32] think open source is having its moment.
[34:34] I work with startups. They are all
[34:36] moving off of these and they're using
[34:38] open source and they're using it at much
[34:39] cheaper rates because a lot of the jobs
[34:42] don't need the latest models. They can
[34:44] use these uh the last generations models
[34:47] and people are moving them local.
[34:48] They're hosting them themselves. This is
[34:51] I think a nonzero chance that this is
[34:53] going to derail anthropic and open eyes
[34:57] uh IPOs and
[34:58] >> they're going to they're going to
[34:59] they're going to be fine. They're going
[35:00] to be fine as long as No, hold on. Let
[35:04] me finish guys. I'm making my point.
[35:06] Shut the [ __ ] up for 30 seconds. I
[35:08] believe that this is going to derail
[35:10] their IPOs. I'm taking it from the top.
[35:12] It's going to derail their IPOs. It's
[35:13] going to be headwinds against it because
[35:15] I think that they're going to have
[35:16] massive margin compression. I believe
[35:18] they're spending so much money that I
[35:20] think they're going to get caught in a
[35:22] trap. I think this could be a trap for
[35:24] them. They overspend. They don't have
[35:26] the same profitability and it doesn't
[35:28] pencil out. And startups are the future.
[35:32] The startups are what eventually the
[35:34] enterprise copies. I think you're wrong,
[35:36] Sax. Go ahead.
[35:39] >> Well, I'll come back, but let your mouth
[35:41] respond. Raise your hand.
[35:42] >> You're wrong. You're totally wrong.
[35:45] >> I think you're wrong. I think they have
[35:47] caught up for 95% of the jobs. This is a
[35:50] crazy head. I'm not saying that the IPOs
[35:51] are off, but I think that their market
[35:53] caps and the headwinds are coming.
[35:56] There's a non-zero chance that they're
[35:57] going to get slowed down.
[35:58] >> I think the answer is slightly
[35:59] different. It's not that 95% of the
[36:01] bleeding edge tasks can be done by
[36:03] everybody. It's that 95% of the tasks
[36:06] can be done by many different models.
[36:08] That's the actual answer. And that's
[36:10] okay. And by the way,
[36:11] >> that's exactly right, Chimath. I agree.
[36:13] >> It's also okay for Anthropic to say, you
[36:15] know what, I have this next generation
[36:16] class of model. I'm going to
[36:17] instantiate, I don't know, a life
[36:19] sciences program, a cyber security
[36:21] business. That's where they can capture
[36:24] as Saxs was saying before over time
[36:28] trillions and trillions of enterprise
[36:29] value because I do think they have
[36:31] excellent models and they have excellent
[36:33] engineers and they have momentum. But if
[36:35] you're going to build a business model
[36:37] that tries to ascribe this layer as
[36:40] having a lot of terminal value, I think
[36:42] that that is a mathematical mistake.
[36:45] That's it. You can't do it. Open source
[36:47] is going to be a spec a headwind to
[36:48] these companies specifically because
[36:49] because
[36:51] all of their best customers and I talked
[36:54] to them whether it's lovable or 11 labs
[36:56] or the startups that were spending
[36:58] hundreds of thousands of dollars with
[36:59] them every quarter just last you know
[37:01] six months ago they have all moved on
[37:03] mass GLM52
[37:06] making their own models the cat's out of
[37:08] the bag they are going to start losing a
[37:11] lot of customers to open source and
[37:14] Google and AWS and Elon Web Services are
[37:17] going to host them. And how do I know
[37:19] this? Every time I do a job, I'm using
[37:21] Plexity Computer. Not not a paid
[37:23] partnership or anything. It just happens
[37:24] to be the best harness that I found. And
[37:27] I start with Grock, Neotron, GLM52, and
[37:30] I also put it into Claude and the
[37:33] results are as good or better. I in
[37:34] other words, I can't even tell the
[37:36] difference between these.
[37:37] >> Let's superimpose the the OpenAI numbers
[37:39] on top of the anthropic numbers because
[37:41] I think it supports the the point I'm
[37:42] trying to make.
[37:43] >> These are estimates, by the way. This is
[37:44] not like literally from the companies.
[37:46] Just want to make sure people know.
[37:47] >> So according to this company that I mean
[37:50] look, who knows how they derive this. We
[37:52] don't know that they're totally true.
[37:54] They're basically showing that OpenAI's
[37:57] run rate, which is ARR. It rose from 33
[38:00] billion in May to 41.3 billion in July.
[38:04] So they're seeing acceleration. Like I
[38:06] said, I mean Sam is out there saying
[38:07] they got their mojo back. They're going
[38:09] to have their best 12 months forward
[38:10] looking ever. And look, Anthropic is
[38:13] still growing really fast. We've talked
[38:14] about this on a previous show. It's not
[38:16] physically possible to grow 10x year
[38:18] over year forever. You'll run out of
[38:20] compute. You'll run out of energy.
[38:22] You'll run out of everything. There's
[38:23] just no Yeah. There's just no way to do
[38:25] things.
[38:25] >> But anthropic, I mean, look, if any if
[38:27] anybody had
[38:29] >> if anybody had said Anthropic would be
[38:31] at over 70 billion in the midpoint of
[38:34] the year back in January and they're at
[38:36] 10 billion,
[38:36] >> nobody would have said it.
[38:37] >> You would have said this is the fastest
[38:39] growing tech company of all time. The
[38:41] idea that they're at risk of getting
[38:44] their entire franchise destroyed, it's
[38:46] not in the data yet is what I'm trying
[38:48] to say. And moreover, Jamas, to your
[38:50] point, it's not only about the model.
[38:52] It's also about the harness. It's about
[38:54] the connectors. It's about the
[38:56] enterprise agreements. There's a lot of
[38:58] things here that you need in order to
[39:00] grow a business like this. And you know,
[39:02] this idea that they need to race to get
[39:05] government protection because of a
[39:07] competitive risk that might happen in
[39:09] the future, I think is just kind of
[39:11] unseammly and gross. I mean, this is
[39:14] literally the most successful tech
[39:16] company of all time and they're racing
[39:18] to the government to basically say, "You
[39:21] need to protect us against our
[39:22] competitors." Great argument.
[39:24] >> Not just our Chinese competitors, our
[39:26] American competitors, our potential
[39:27] future competitors.
[39:28] >> They should hire Lena Khan. Okay. And
[39:30] frankly, it's gross. And just let me
[39:32] give a couple of examples cuz I know
[39:33] people don't have a lot of sympathy for
[39:34] Chinese companies. That's fine. I'm not
[39:36] defending Chinese companies. I'm
[39:37] defending American developers who need
[39:39] to be able to use everything in the
[39:41] public domain. And let me give you an
[39:42] example. Ma Maratti, her new model,
[39:45] >> thinking machines.
[39:46] >> Thinking machines. They currently have
[39:48] the best American open-source model. You
[39:51] know how it was trained? It was
[39:52] bootstrapped. It was distilled off a
[39:55] Chinese model, Kim K 2.5. Now, if you
[39:58] say that Chinese model is based on IP
[40:01] theft, what is thinky? It's a derivative
[40:05] work off a model that anthropic is
[40:08] trying to taint as being IP. By the way,
[40:11] there's been no evidence of this.
[40:13] There's no evidentiary process. They are
[40:15] simply trying to paint with a very broad
[40:17] brush here and say that now that model
[40:21] is tainted. Let me give you another
[40:23] example. So cursor rolled out its new
[40:25] product composer 2. They were able to
[40:28] post-train that model using Kimik 2.5 on
[40:32] their own proprietary coding data. Okay,
[40:34] so think about this. They started with a
[40:36] Chinese open- source model and then they
[40:39] used their own data and they came up
[40:40] with a new derivative product. This is
[40:43] the way that open source works. You take
[40:44] things that are in the public domain,
[40:47] you fork them, you make them your own.
[40:49] And by the way, once it's in the public
[40:51] domain, it's not a Chinese model
[40:53] anymore. It is open weights that are
[40:55] freely available to anyone. It's a file,
[40:57] okay? And you take that, you fork it,
[40:59] you run it on your own hardware in an
[41:01] American data center. No packets are
[41:03] going back to China. No data is going
[41:05] back to China. Nothing's going back to
[41:07] China. An American company has taken
[41:09] open- source contributions in the public
[41:11] domain, made it their own, and then
[41:13] developed their own model. And if you
[41:15] say that American companies can't do
[41:17] that or that somehow it's tainted with
[41:20] IP theft, you are basically going to put
[41:22] a dagger through the heart of the entire
[41:24] American open source ecosystem. And that
[41:26] is exactly what Anthropic wants because
[41:29] they do not want to have the
[41:30] competition.
[41:31] >> Freeberg, maybe you can uh close this
[41:34] out here and then I'll put my
[41:36] >> I'll just zoom out for a second and I'll
[41:38] say it. Think about
[41:41] the strategy as well for China. If you
[41:45] think about the global economy of the
[41:46] last 50 years, the US has acrewed so
[41:48] much value by being at the core of the
[41:51] knowledge economy and effectively a
[41:54] services economy. And in that sense
[41:57] through the development of intellectual
[41:59] property of IP of knowledge and then the
[42:02] conversion of one bit to another bit
[42:05] we've been able to derive trillions of
[42:07] dollars in GDP. Meanwhile we outsourced
[42:10] manufacturing and created a sleeping
[42:13] giant in China where they have this
[42:16] incredible manufacturing capacity. And
[42:18] at the end of the day, if you think
[42:19] about the course of like human
[42:21] technology evolution and human
[42:23] prosperity, it's largely driven by our
[42:26] capacity to convert molecules from one
[42:29] form to another and use the least amount
[42:32] of energy possible to do that. That's
[42:34] that all of technology ultimately leads
[42:36] to that simple equation. Molecule
[42:37] conversion. Making that beautiful couch
[42:39] behind you at the lowest cost possible.
[42:41] Making materials, making semiconductors,
[42:44] making all this stuff. Everything in our
[42:45] world is driven by molecule conversion.
[42:48] So at the end of the day, if the
[42:50] knowledge economy and the services
[42:51] economy gets compressed, much like AI
[42:53] and open source AI in particular,
[42:56] effectively flattens that value because
[42:59] all of that value is now open source.
[43:02] It's free and it's simply a function of
[43:04] turning on a switch and running it. The
[43:06] value of the US and the western economy
[43:09] has been largely degraded. And what's
[43:11] left is the value of the molecule
[43:13] economy. the ability to convert
[43:15] molecules and use energy to do that.
[43:17] When you look at the juxtaposition of
[43:19] China versus the United States today, we
[43:21] have one terowatt of electricity
[43:22] production capacity in the US and
[43:24] they're on their way to having eight. We
[43:27] have about 10 billion square ft of
[43:29] manufacturing capacity. They have 200
[43:33] billion square ft of manufacturing
[43:36] capacity. So they have 20x the
[43:38] manufacturing capacity, 8x the
[43:40] electricity production plus all of their
[43:42] other sources of energy. And I think
[43:45] that's the the long game for China over
[43:47] a decade, two, three decade process is
[43:50] by compressing the knowledge economy and
[43:52] the services economy, commoditizing it
[43:55] completely, they are left holding all
[43:58] the value in the global economy because
[43:59] they can make stuff and they can make it
[44:01] cheaper than anyone because they have
[44:03] the most electricity production. That's
[44:05] a very simple rubric for kind of how I
[44:07] look at the the long game that they're
[44:09] trying to play here.
[44:09] >> You believe they're trying to
[44:10] commoditize this very important space
[44:13] just like they did for
[44:14] >> global cars,
[44:16] >> global knowledge and global services.
[44:18] The creation and movement of bits gets
[44:20] commoditized and what's left over is the
[44:23] creation of electricity and the creation
[44:25] of molecules both of which they have
[44:27] this very difficult to surmount
[44:30] advantage that's going to make them the
[44:32] core dependency for the world. That's
[44:34] what I think is kind of the long game
[44:36] here.
[44:36] >> And just so you know, this data comes
[44:39] from reports in places like the
[44:41] information or other sources or leak
[44:45] numbers and they try to just make charts
[44:47] based on it.
[44:48] >> Yeah. And by the way, I you know, I have
[44:50] my own sources too. I've talked to
[44:51] investors in these companies and I'm
[44:53] just telling you that both Anthropic and
[44:55] Open AI are taking their estimates up
[44:57] right now, their forecasts up. So I mean
[45:01] look I think in the future it may be the
[45:03] case that open source takes share fine
[45:05] it's because the market's so big and
[45:07] there is always a market for open
[45:09] because open is more controllable it's
[45:11] more customizable you can own it you get
[45:14] the data sovereignty you get the
[45:15] sovereignty but it's also more work so
[45:18] there are different use cases there's
[45:20] different parts of the market
[45:21] >> that's actually very interesting the the
[45:23] more work part there Sachs is super
[45:25] interesting like 3 to 6 months ago it
[45:27] was so much work to stand these up and
[45:30] Now there's so many intermediaries
[45:31] building the harnesses that default to
[45:33] it that that is getting worked out. But
[45:35] that has always been the issue with open
[45:37] source for sure is the amount of work it
[45:39] takes to implement. Uh if you do want to
[45:41] implement this inside of your
[45:43] organization, please make a call to 8090
[45:46] and try the software factory. Use the
[45:47] promo code JAL to get a free
[45:49] consultation at 8090. All right.
[45:53] Can they get a free consultation?
[45:55] >> Everyone's talking their books.
[45:57] Everyone's talking their books actually
[45:59] including I'd say the anthropic and uh
[46:03] open AAI investors. It's amazing how
[46:05] many of these
[46:05] >> Brad was on the show two weeks ago when
[46:07] you on the episode Brad was LIKE LET ME
[46:09] TELL YOU WHY this is going to he's
[46:12] holding on he's holding on to this
[46:14] actually
[46:17] >> he's like
[46:18] >> I actually give Brad a lot of credit
[46:20] because I do think that he's objective
[46:21] about public policy or as objective as
[46:23] you can be given that he does own all
[46:25] these companies. But look, let me just
[46:28] tell you that you know I see a lot of
[46:30] folks who are suddenly China hawks and
[46:33] saying we need to stop China. WE STOP
[46:34] LIKE THEY'RE CHINA HOGS. EVERYBODY
[46:36] WANTED TO SELL THEIR CHIPS.
[46:37] >> How about disclosing first whether
[46:39] you're on the cap table of anthropic?
[46:41] Okay.
[46:41] >> Absolutely. Are you guys on any of these
[46:43] cap tables?
[46:45] >> No, I'm not.
[46:46] >> Not directly.
[46:46] >> No, not directly. Yeah, that's
[46:49] indirectly. Maybe got a little access.
[46:50] You know what it's like, Chimat? It's
[46:52] like when you you got that great flush
[46:53] or straight and you're like, "Please
[46:55] don't pair the board." Brad's like,
[46:57] "DON'T PAIR THE BOARD. PLEASE don't pair
[46:59] the board."
[47:01] All right. Here we go.
[47:02] >> I love open source. I love I love this
[47:04] open source AI stuff.
[47:06] >> I think it's so awesome.
[47:07] >> The value punk rock.
[47:10] >> It's it's just like there's so much to
[47:11] be done with it. It's just exciting and
[47:13] awesome and yeah, Chimoff's point is
[47:15] exactly right. Most of the models can do
[47:18] 95% of the tasks. And if that's the
[47:20] case, then it's not like everyone needs
[47:22] to scramble to get the best open source
[47:24] model. You just need open source to do
[47:26] 95% of what you want to do with AI. And
[47:28] then the other 5% you get specialized or
[47:31] high value or you pay a premium. And by
[47:32] the way, if you're a big enterprise and
[47:34] you need to have rappers and support and
[47:35] all these other tools for your
[47:36] employees, buy Anthropic or OpenAI or
[47:38] Gro tools or Gemini, like just make sure
[47:41] you have a good partner,
[47:43] >> use the promo code Jal. You get a Zoom
[47:45] call with your mom. Okay.
[47:48] A Zoom call with A ZOOM CALL WITH YOUR
[47:50] take a selfie with him on Zoom.
[47:52] >> I'm sorry.
[47:53] >> Anyone's going to sign a $10 million
[47:55] contract with 8090, you get to have uh
[47:57] three of the four of us for a fiveminute
[47:59] Zoom call.
[48:00] >> Absolutely. You get to call
[48:02] 10 million. You get to GUESS WHO'S ON
[48:04] THE POD.
[48:05] >> And anyone ready to make a pre- purchase
[48:06] on a million dollars of potato seed, you
[48:08] also get to have
[48:10] >> use the promo code Sax Poo. Yes. And if
[48:13] you'd like to guest host, the first
[48:15] person to put a $25 million check into
[48:17] Launch Fund 5 gets to guest moderate an
[48:20] episode with your boy J Cal. All right.
[48:22] Uh Promo City this week. All right.
[48:25] Anthropic Copyrights. Here we go.
[48:28] Anthropic has settled their AI copyright
[48:30] lawsuit for 1.5 billion
[48:34] billy on Monday. Largest copyright
[48:36] settlement Freedberg in the history of
[48:39] the United States of America.
[48:42] First major AI training lawsuit to
[48:44] settle. There are many more in the
[48:46] pipeline. Anthropic downloaded 7 million
[48:49] books from pirated websites to train
[48:51] Claude. And that alone it may or may not
[48:55] be a crime. This has been adjudicated a
[48:58] little bit in the courts. They had ruled
[49:00] previously that AI on copyrighted books
[49:02] is legal under fair use, but there's
[49:04] going to be some future cases. So, this
[49:06] is a settlement. They didn't go to the
[49:08] bat. Lawyers getting 101 million.
[49:10] Authors get 3,000 a book. 500,000 books
[49:13] were covered in it. Thus far, 91% of the
[49:16] covered authors have claimed their
[49:18] share. Tons of other ones are on the
[49:21] way. And here's your second victory p
[49:24] fap of the episode. Content providers as
[49:28] a group need to get together and fight
[49:30] for their rights in unison. New York
[49:32] Times Met for the right to party.
[49:34] >> No. Fight for the right to get paid and
[49:36] to survive.
[49:38] TBT and say as a group either give us
[49:41] these terms or don't index us. They are
[49:43] interfering with their ability to
[49:45] leverage their own content. is
[49:47] profoundly unfair and those magazines
[49:49] and newspapers need to what's that?
[49:52] >> You're going to get steamrolled.
[49:53] >> It's possible. YouTube is a great
[49:55] example. That's what's going to happen
[49:56] here. There'll be a settlement where
[49:58] they are going to be able to claim their
[50:01] I will bet any amount against your your
[50:04] your premonition here Jal. This is like
[50:08] >> I am going to go with for my biggest
[50:09] winner for
[50:11] >> training data owners like the New York
[50:13] Times, Reddit X, Twitter, YouTube etc. I
[50:16] think what we learned in 2023 was that
[50:19] the language models are starting to hit
[50:21] parody very quickly and that the real
[50:24] value is going to be in and it may even
[50:27] become commodities and open source may
[50:28] win the day. So then I think the winner
[50:30] is folks who have the training data.
[50:34] >> You know the best thing about this is
[50:35] watching Jason's reaction to Jason.
[50:39] Did you do a picture and picture of me
[50:40] just be like go Jac
[50:43] did we make a bet here? Did we make a
[50:45] bet? I don't know. He's gone.
[50:47] >> Well, actually, this I'm this settlement
[50:50] I don't think quite proves exactly what
[50:53] you want it to prove. Jal,
[50:55] >> go ahead.
[50:56] >> Explain. Can I can I make a nuance here?
[50:59] >> Of course. Of course.
[51:00] >> So, okay. Look, and you know, obviously
[51:02] I'm not a huge fan of anthropics. I
[51:04] think they're potentially destroying the
[51:05] whole ecosystem for their own purposes
[51:07] of regulatory capture. But let's just be
[51:09] very clear about what Let's just be very
[51:12] clear about show anytime, Daria.
[51:13] >> Yeah. Let's just be very clear about
[51:16] what this judgment was and was not. So,
[51:19] okay, what Anthropic did is they pirated
[51:22] all these books from LibGen and they
[51:25] trained on them. And the reason why they
[51:27] got in trouble is cuz they basically
[51:30] took stolen books. They didn't even pay
[51:32] for one copy of them. But if they had
[51:34] paid for just one copy of each book,
[51:38] they could not have been nailed for
[51:40] piracy. they would have been potentially
[51:42] under fair use, which I understand Jal
[51:44] is still being litigated in the courts,
[51:46] but that would have been their defense.
[51:47] So, the reason why they got nailed with
[51:49] this $ 1.5 billion judgment is they
[51:52] wouldn't even buy one copy. It is still
[51:54] Anthropic's position and it's OpenAI's
[51:57] position that they should be able to
[51:58] train on all these books under fair use
[52:01] if they buy one copy. And that issue has
[52:04] not been resolved yet. Now you should be
[52:07] able to see the total hypocrisy of their
[52:10] point of view relative to the previous
[52:12] issue which is they believe they should
[52:15] be able to train on every creator's
[52:17] output in the world. You know as long as
[52:19] I guess they bought one copy of it
[52:21] against the will of those creators
[52:23] whether those creators like it or not.
[52:25] They believe it is fair use to train
[52:28] their models and derive their own
[52:30] weights based on fair use. However, they
[52:34] say that the one type of content that
[52:36] you should never be able to train on is
[52:38] their output. That is currently their
[52:40] position. It's completely hypocritical.
[52:42] And actually, if you go back to
[52:44] anthropics blog post in February where
[52:47] they defined this concept of industrial
[52:49] scale dissolation attacks for the first
[52:51] time, they coined that expression. And
[52:53] this is, you know, I worry that people
[52:54] in the government policy makers don't
[52:56] understand that this is all part of a
[52:58] anthropic op. No one used the terms
[53:01] distillation and attack together until
[53:03] Anthropic wrote that blog post.
[53:05] Distillation was simply an industry
[53:07] standard practice. But then Anthropic
[53:09] coined this idea of industrial scale
[53:11] dissolation attacks. In any event, if
[53:14] you go to that blog post, search for the
[53:16] words IP theft. It's not in there.
[53:18] Anthropic did not claim even though they
[53:21] were trying to coin this new concept and
[53:23] brand this idea of industrial scale
[53:26] dissolation attacks. They did not have
[53:28] the hutzbah to claim that it was IP
[53:31] theft.
[53:32] >> The coahjones the hypocrisy the kutzbah
[53:34] to claim that it was IP theft. Why?
[53:37] Because they maintain that it is their
[53:40] right to train their models on all the
[53:43] world's output even if the creators
[53:45] don't want them to. IP for we but not
[53:48] for thee.
[53:48] >> Exactly. So, so Jal, I don't even want
[53:50] to get into whether you're right or not
[53:52] on the fair use question. Maybe you are
[53:53] right. I don't know. Okay. But my point
[53:55] is about the hypocrisy and they
[53:58] themselves never claimed that this was
[54:01] IP theft by the Chinese companies. What
[54:02] they tried to claim was that it was a
[54:04] national security threat because what
[54:06] would happen is these Chinese companies
[54:08] would distill off them, create their own
[54:10] models, and those models would not have
[54:12] guard rails. So they were making a
[54:14] different kind of of argument. That
[54:16] argument never that argument though
[54:18] never found purchase with policy makers
[54:21] because I think that they could see that
[54:24] yeah look guardrails are important but
[54:26] you know it never really found purchase
[54:28] until anthropic started claiming oh this
[54:30] is IP theft but they have not been
[54:33] willing to make that argument publicly
[54:35] because they know that it would poison
[54:37] all of their fur use lawsuits that are
[54:39] happening. And Jimoth, like you
[54:40] mentioned, the New York Times is
[54:42] currently suing Open AI for basically a
[54:46] industrial scale distillation attack. I
[54:48] mean, Open AAI went on the New York
[54:50] Times website, used scrapers, slurped up
[54:53] all of their information at a scale that
[54:55] no human could achieve, and then they
[54:57] used that as training data and reverse
[54:59] engineered the model weights
[55:00] effectively. So my point is that even
[55:05] Anthropic and Open AAI won't publicly
[55:07] admit that what China is doing is IP
[55:09] theft because they are doing it
[55:11] themselves. It's totally hypocritical
[55:13] and I don't think policy makers should
[55:15] be making arguments that these companies
[55:17] themselves won't make because they know
[55:19] that they will lose all these court
[55:20] cases.
[55:22] >> Freeberg, any thoughts here on um and
[55:24] obviously this is still being litigated
[55:27] as we've discussed. It's 150 major
[55:29] cases. New York Times is one of the
[55:31] music industry.
[55:43] >> So you write this book. Okay.
[55:44] >> Harper Collins
[55:45] >> and you opt to not submit it to AI
[55:48] because there's a restrict. You keep it
[55:50] closed. No one can read your book.
[55:52] >> We did for Google search. You can't be
[55:53] on Google.
[55:54] >> No one can read your book. You're not
[55:56] letting anyone read your book. you want
[55:58] to pay for your book if you want to read
[55:59] it
[56:00] >> 10 bucks
[56:01] >> and someone pays $30 for Genule the
[56:03] great American novel and they read it
[56:05] and then they write a review and they
[56:07] publish their review on the internet now
[56:10] on the internet the reviewer talks about
[56:11] your book and describes your book gets
[56:14] webcrolled by an AI engine and the AI
[56:16] engine learns from that learns about
[56:18] your book and now there's some
[56:20] commentary made about your book when
[56:21] someone asks a question about your book
[56:22] in the AI engine do you feel like your
[56:24] copyright was violated in that sense
[56:27] Um,
[56:28] >> so the book the AI never ingested your
[56:30] book. It ingested metadata about your
[56:32] book. It ingested reviews about your
[56:33] book. It ingested third party analysis
[56:35] about your book. All of which was on the
[56:37] open internet. And it didn't just copy
[56:39] that stuff, but it used it to learn
[56:40] about your book. So you're saying these
[56:42] 1500 reviews for Angel, how to invest in
[56:44] technology stars, timeless advice from
[56:46] an angel investor turned 100,000 into
[56:48] 100 million. These reviews would then be
[56:50] the basis of the AI. So I guess I would
[56:53] have to be okay. But this like, you
[56:56] know, this eloquently brash blueprint
[56:58] for angel investing, that verified uh
[57:00] review. Yes, I would be fine with that
[57:03] >> fivestar review being in there.
[57:05] >> Really?
[57:06] >> You now have no knowledge. You now have
[57:07] no knowledge of the above
[57:12] 1498 reviews did you not write? Three.
[57:14] Uh, actually, I'll tell you the secret.
[57:17] When you guys actually get asked to
[57:18] write a book or any of you have the
[57:19] capability of completing a book. Yeah,
[57:21] because I because I'm 97 years old
[57:23] living in the
[57:24] >> What AI agent did have you used to what
[57:26] bot what bot did you use in 2017?
[57:29] >> Now we know where he pointed that stupid
[57:31] open source AI slop cannon that he's
[57:33] built.
[57:36] >> I verified the purchase too.
[57:39] Freeberg, you make a great point. You
[57:41] make a Oh, there it is. Jennifer, great
[57:42] American novel. Practical wisdom for
[57:45] mastering ambition, building resilience
[57:48] and winning at life work and innovation.
[57:50] >> Master virtue signaling master virtue
[57:53] signaling.
[57:55] >> Incredible. Incredible.
[57:56] >> That's beautiful. That is beautiful. But
[57:58] JL, I mean, this is my point. Knowledge
[58:00] can't be contained. It's diffuse. And so
[58:02] the form of copyright is very clear. The
[58:04] case law and copyright's very clear. I
[58:06] cannot lift text out of your book,
[58:08] reprint it, and claim it as my own. That
[58:11] is a violation of copyright. But my
[58:13] reading of your book, my reading of the
[58:15] reviews of your book, the diffusion of
[58:17] the knowledge that arises from your
[58:19] book, that is ultimately going to lead
[58:22] to some abstract transformation of
[58:24] knowledge into a new output that someone
[58:26] might read. And I think it is very
[58:28] unlikely that we will find ourselves in
[58:31] a place where the idea that knowledge
[58:33] transferred digitally, processed
[58:35] digitally, and turned into other content
[58:38] is going to end up violating copyright
[58:40] in I understand your position. There are
[58:42] workarounds. Obviously, we've always had
[58:44] cliff notes, right? So, if a book became
[58:45] good enough, somebody could write the
[58:46] cliff notes of it. You can't stop that.
[58:48] There's a four-part test for this. We
[58:49] we've talked about this for three years
[58:50] here on the pod. What I'd say is
[58:52] American companies should take 10% of
[58:54] their revenue, if they're building these
[58:56] models, and do splashy cashy and do
[58:58] settlements. And that's exactly what's
[59:00] happening. If you're a copyright owner,
[59:01] you should study what the music industry
[59:03] does. They are raid dogs and they will
[59:06] fight tooth and nail and keep you in the
[59:08] courts until you submit and make a
[59:10] settlement and agree that you're
[59:12] licensing it and then that gives them
[59:14] that case law and that settlement to go
[59:17] to the next person and the next person
[59:18] and the next person and that's why
[59:19] they've been able to successfully defend
[59:21] it and then if you're competing with me
[59:23] that becomes the issue. So if you said
[59:25] hey what's his book about and what do
[59:27] people think about it what are the best
[59:29] parts of it and that comes from the
[59:30] reviews okay fine fair enough. The
[59:33] problem is and I and I'll I'll share
[59:35] with you there's been some other
[59:36] lawsuits here that are making the way
[59:38] through courts. What's going to be the
[59:39] problem is the application level layer
[59:42] that you talked about Shimoth as they go
[59:44] into the application layer and they use
[59:45] this. There's a there's a big court case
[59:49] here. Obviously you know about some of
[59:51] these but there are now other cases.
[59:55] There are music cases. There's a New
[59:56] York Times case. The one that's kind of
[59:58] interesting is Thompson Reuters versus
[59:59] Ross. This is a uh final judgment on AI
[60:02] uh training copyright. There's a company
[60:05] called Weslaw. They're like Lexus Nexus
[60:07] and people have been trying to claim
[60:10] that they can train on the outputs of
[60:12] something like uh Westlaw. And when
[60:15] you're in the same business as me that
[60:18] has a special place in copyright law
[60:20] because I have you're you're infringing
[60:22] on my ability to use my copyright.
[60:25] And your argument I think Fraber holds
[60:28] up that it wouldn't be it wouldn't stop
[60:31] somebody from buying the book. But the
[60:33] second you are actually competing with
[60:36] me like directly that's when these
[60:37] things have problems. And that's why I
[60:39] think the music industry is going to win
[60:40] and and some other places are going to
[60:42] win. But listen, this is
[60:44] >> brand new territory. These lawsuits are
[60:46] brand new territory and IP law does not
[60:49] >> actually have the nuance yet. So we're
[60:51] going to as a society have to make a
[60:52] decision here on what is fair. I suggest
[60:55] just like the self-regulatory group that
[60:57] we talked about last week, all the AI
[61:00] companies should get together, take 10%
[61:01] of your revenue, put it in a pool, and
[61:03] keep paying the people and getting
[61:05] permission from them so you can get
[61:07] updates on the content so you get the
[61:08] next book so you get the next New York
[61:09] Times story, the next Reuters story.
[61:11] >> 10% is not going to satisfy the rabbit
[61:13] dogs, let me tell you. They're going to
[61:14] go for 100%. Now, JK, let me here's a
[61:16] question I want to ask you. Yes. Given
[61:19] that the fair use doctrine is not a
[61:22] decided matter yet, given that anthropic
[61:25] and open AI are embroiled in huge
[61:27] lawsuits against very well financed
[61:30] content creators and those communities
[61:33] and the outcome is indeterminate and
[61:35] there's billions of dollars at stake,
[61:37] maybe even their entire product at
[61:39] stake. Do you think that they have
[61:40] potentially made a fatal mistake by
[61:43] arguing that distilling content against
[61:46] the wishes of its creator is IP theft?
[61:51] >> Do you see what I'm saying?
[61:52] >> Yes.
[61:52] >> Like, is this potentially a fatal
[61:54] mistake? See, here's what they could
[61:55] have done.
[61:55] >> Well, you would bring that to the
[61:56] Supreme Court and you say, "Hey,
[61:57] listen."
[61:58] >> But here's what they could have done.
[61:59] What Anthropic could have done is they
[62:01] could have said, "Listen, we have these
[62:02] Chinese companies are creating fake
[62:04] accounts and they're using proxies to
[62:06] basically use our product in violation
[62:09] of our terms of service." Now,
[62:11] >> using those model outputs is not IP
[62:13] theft because it's fair use. However,
[62:15] it's a deceptive business practice for
[62:17] these guys to lie about who they are
[62:19] when they set up accounts at scale. And
[62:22] so, they're engaged in a deceptive
[62:23] business practice. and we're going to do
[62:25] everything we can to stop that, but we'd
[62:27] like the government's help in stopping
[62:28] that, too. However, we not we're not
[62:30] saying anything about IP theft.
[62:32] >> Yeah.
[62:32] >> Wouldn't that be the more nuanced
[62:34] approach? Of course, because I think
[62:35] that they're on the verge of being
[62:36] hoisted on their own petard here.
[62:38] >> Yeah. I mean, it's a cell I think that
[62:40] kids call it a cell phone, right? Like
[62:41] you basically
[62:43] >> it's just a classic cell phone. And you
[62:46] know, in most of these cases,
[62:47] settlements happen. So again, if you
[62:48] just look at the music industry, the
[62:50] newspapers, the magazines, and some of
[62:51] those folks and book authors, they've
[62:53] tend to be very meek. There's a new
[62:56] trend happening now, Friedberg, they're
[62:58] a lot of content providers are saying to
[63:00] Google, take us out of the index because
[63:02] Google has one bot and that bot does the
[63:05] Google crawl and that bot also does the
[63:08] AI crawl. And what the industry is
[63:10] saying is, hey, split that up. I want to
[63:12] be in Google, but I don't want to be
[63:13] indexed in AI. So people are now saying,
[63:16] hey, we'll take you out of the index.
[63:17] which Rupert Murdo got right. If all the
[63:19] newspapers said collectively, "Do not
[63:21] index us, Google. We're no indexed, that
[63:23] would have made Google come to the table
[63:25] and give them a royalty and give them
[63:27] some money for being indexed."
[63:28] >> I did that, dude. There was a there was
[63:29] a deal that happened, but it didn't it
[63:31] go the way you're describing.
[63:33] >> Well, cuz they didn't have a united
[63:34] front. Now, I think
[63:35] >> these guys wanted they wanted Google's
[63:37] user base. So, they ended up doing a
[63:38] deal where they had this like payw wall
[63:41] like exclusion rule whereas like you
[63:43] could show a certain number of free
[63:44] articles. There was a whole negotiated
[63:45] settlement. Well, I'm talking even long
[63:46] before that when the first index
[63:47] happened. But here's I you know I I I
[63:50] think we are on the cusp of a some type
[63:53] of a settlement getting done here. Uh
[63:55] and I think that would be good for
[63:57] America to take a leadership position in
[63:59] that because you do want to keep getting
[64:01] that. But all of these content
[64:03] companies, they should be meeting with
[64:04] each other, the music industry, the New
[64:06] York Times, all of them to stop their
[64:09] content from getting used without their
[64:10] permission and from competing with them.
[64:13] That's my point. You know, you you only
[64:15] get a settlement when both sides can
[64:18] agree. And it seems to me that if you're
[64:20] one of these content creator lobbies and
[64:23] you see that Anthropic has just told the
[64:26] government that training on a creator's
[64:30] output without their consent is IP
[64:32] theft.
[64:32] >> Yeah.
[64:35] >> They have now basically confessed to
[64:38] their entire product being stolen. And
[64:39] it seems to me that why wouldn't the
[64:41] content creators now assert that they're
[64:43] entitled to own 100% of anthropics
[64:46] revenue? It seems to me that this could
[64:48] be a bridge too far that you know that
[64:50] that they're so good at regulatory
[64:52] capture. They're so good at making these
[64:54] arguments and getting the government
[64:55] involved to create new regulations to
[64:57] protect them. But I wonder if this was
[64:59] just a little bit too cute. And again,
[65:02] if they just positioned it slightly
[65:03] differently, if they said, "Look, these
[65:05] Chinese companies are creating fake
[65:06] accounts. That's a deceptive business
[65:08] practice. We're not saying this is IP
[65:10] theft, right? But instead they said IP
[65:13] theft and now the whole startup
[65:14] community is activated. You saw that you
[65:17] know Gary Tan and 200 startups wrote
[65:20] that letter. Why? Because they know that
[65:22] if this IP theft thing sticks that all
[65:25] derivative works of Chinese models are
[65:28] tainted now too. So that means the whole
[65:31] startup ecosystem is now at risk.
[65:33] >> Yeah.
[65:34] >> So I just wonder if these guys have just
[65:36] gone it's just all a bridge too far. I
[65:38] mean, if you go to Washington,
[65:40] >> Yeah. and you lay down with the dogs,
[65:42] don't be surprised if you wake up with
[65:43] the fleas. Like, you they decided to
[65:46] engage in this, you know. So, they they
[65:49] may have poked the tiger. Um I think I
[65:51] think it's pretty accurate. Oh, by the
[65:52] way, uh the publishing schedule for uh
[65:55] 2027 was released. We have some new
[65:57] books coming. You heard mine, Jen, you
[65:59] flecting coming to you now. Ferrari on
[66:02] my wrist. How to win at life and afford
[66:04] a $250,000 watch from David Sax.
[66:07] >> Oh my god.
[66:08] >> This is Pompic Sachs.
[66:10] >> We're going to get need to get that
[66:11] done. Here it is. The fight to save
[66:12] America: Destroying Socialism, Shrinking
[66:14] the Debt, and Winning AI by David.
[66:17] >> Yeah, it's a pretty good one. I think
[66:19] this is our new all-in uh here it is.
[66:21] Chama polyatina, a sexual Italian
[66:24] summer.
[66:26] It's a romance novel. This is the only
[66:28] person who decided to do fiction. This
[66:29] is fiction.
[66:31] >> It's fiction. It's fictional. Chimat,
[66:33] >> I got Traumat's uh non-fiction novel
[66:35] also.
[66:35] >> Oh, you have his non-fiction as well.
[66:37] Here it is. Here it is. Enterprise
[66:38] sales. Chimat dystopia. How my software
[66:41] startup affects my Italian summer.
[66:45] >> There you go.
[66:46] >> It's great. This is great. Oh my god.
[66:48] Wow. That's coming from the allin
[66:50] >> the IT guy in Milan.
[66:52] >> Absolutely. This is This is coming from
[66:54] Allin uh books. It's our new publishing
[66:57] label coming in 2027. We'll also have
[66:59] the Brad Gersonner, Bill Gurley, Elon
[67:01] Musk, and other titles coming. So, we'll
[67:03] have those on future episodes. More more
[67:05] titles coming. Breaking topic here.
[67:07] Google and Tesla shared their results
[67:09] today. Had a lot of talk about capital
[67:13] expenditures. Google blew the doors off
[67:17] of their um
[67:18] >> blew the doors off. I mean,
[67:20] >> it was insane and outrageous. And they
[67:23] were down like 7 to 10%.
[67:25] Google Cloud growing
[67:27] >> because of their cash%
[67:29] year-over-year.
[67:30] >> Yeah, because of the cash flow numbers.
[67:32] And now is on a hundred billion dollar
[67:34] run rate. That's but one business.
[67:36] Tesla's capex surged 140% year-over-year
[67:40] and they expect 25 billion in capex.
[67:43] Google's capex forecast from 195 to 205
[67:46] billion this year. So next year will be
[67:49] even higher. Tesla down 14, Google down
[67:51] 7% of taping. Who knows? Um but both
[67:53] reported negative free cash flow. In
[67:55] other words, the amount of cash in the
[67:57] bank went down instead of up. And for
[67:59] Google, that was the first time ever.
[68:03] IPO update, uh SpaceX down 30% from its
[68:06] day one closing price, now trading 1.5
[68:08] trillion. A lot of pressure on the
[68:09] stock. We'll talk about that as well.
[68:12] Obviously, they went public at 2
[68:13] trillion. Got a big pop, uh the Elon
[68:15] pop, and um there could be uh more
[68:19] downward pressure or it could have found
[68:21] a bottom. You know, you never know with
[68:22] these things. This is unprecedented
[68:24] territory. We've never had an IPO this
[68:26] big, but some lockups. Here's the chart.
[68:29] Bunch of lockups are are happening at
[68:31] different staged intervals. So, let's um
[68:35] Chimath
[68:36] talk a little bit. Here's your SpaceX. I
[68:39] guess capex chimoff is being built out
[68:43] obviously for AI and there is the case
[68:46] of open AI spending on capex in order to
[68:50] provide their service but then there's
[68:52] also Google which is making these capex
[68:54] investments to resell it as part of
[68:58] Google cloud incredible product and then
[69:00] on the other side they're using it for
[69:02] their own obviously infrastructure and
[69:04] their business is just growing like
[69:05] crazy whether it's YouTube or Google
[69:07] cloud or even search is still growing So
[69:09] I looked at this and I said, "Well, this
[69:10] seems like a really good use of capital
[69:12] instead of just giving dividends like
[69:14] building out this infrastructure to me
[69:15] sounds like an investment in the
[69:16] future." Seems like a buy signal to me,
[69:18] but the market is obviously
[69:20] disappointed. Why is the market
[69:21] disappointed? Is it a buy signal for
[69:23] you? Is it make you more uh excited
[69:26] about management and what Sundar and
[69:28] Sergey are doing over there? Or does it
[69:30] make you concerned? Yeah,
[69:32] >> I'm more bullish. Do you know what
[69:35] Google's 25y year average return on
[69:40] invested capital has been since going
[69:42] public?
[69:42] >> Take a guess.
[69:43] >> 17%.
[69:44] >> No.
[69:47] 21%.
[69:48] >> No.
[69:49] 23%.
[69:51] >> No.
[69:52] >> 29%.
[69:54] >> Prices, right?
[69:55] >> 35%.
[69:56] >> 32%.
[69:58] >> Jesus.
[69:58] >> Okay. This is when you are a machine and
[70:02] a group of people and a business model
[70:04] that compounds money at 32%
[70:08] over 20 year average. You give these
[70:10] guys the benefit of the doubt. These are
[70:12] not people that are flying fast and
[70:14] loose. They are methodically investing
[70:16] in their edge. And this is I go back to
[70:18] the first conversation. They are going
[70:21] to get massively rewarded. You know,
[70:23] there was a lot of Twitter chatter or ex
[70:25] chatter about Gemini usage and was it
[70:28] real or was it not real and is their
[70:30] revenue growth really coming from AI
[70:32] enabled workflows? It's all malarkey.
[70:36] Google has an incredible search
[70:38] experience. They've seem to be
[70:40] navigating this transition to use AI.
[70:42] It's been done very well. They have an
[70:44] incredible cloud business and they have
[70:46] an incredible silicon business. The best
[70:48] thing that can happen to them is 500
[70:51] different models proliferate and they
[70:54] support all of them because they will
[70:56] make so much money at the silicon layer.
[70:59] They'll make so much money as the cloud
[71:01] provider and they'll find a bunch of
[71:04] apps including YouTube and other things
[71:06] to make money from because you use the
[71:08] AI to target ads better or to help make
[71:10] better content etc etc.
[71:11] >> Fragmentation's good for them.
[71:13] >> Oh, it's great for them. It's a
[71:15] compounding machine. I think the
[71:17] reaction, by the way, is because Jason,
[71:19] I saw a tweet from Ryan Peterson. I
[71:20] don't know if it's true, but he said
[71:22] Google will be spending 20% of this
[71:24] year's military budget in capex. So
[71:26] maybe what people are reacting to is
[71:28] just the scale of the investment they
[71:29] haven't seen. They're free cash flow
[71:31] negative for the first time since going
[71:33] public. So that obviously takes people
[71:36] by surprise,
[71:38] but they're in a huge investment period.
[71:41] And I think it'll pay off dividends even
[71:43] if they, as Freeberg's first guess was
[71:45] half the number. So even if they did
[71:47] half the number, they'd still be
[71:48] overachieving.
[71:50] Yeah. I mean, Freeberg, if you look at
[71:51] Apple, I think they bought back half
[71:53] their stock. They've given hundreds of
[71:55] billions of dollars back in profits. And
[71:58] gosh, it seems to me giving all this
[72:00] money back, buying back in the form of
[72:01] buying back your shares, you know, or or
[72:04] giving tons of dividends.
[72:06] I think it's great that tech companies
[72:08] are now saying, wait, we have something
[72:09] to invest in. The next big thing is
[72:11] ondemand intelligence and there is no
[72:13] upper bound for intelligence or I don't
[72:16] think anybody I certainly don't see you
[72:18] know anytime in the next 10 years people
[72:20] saying I got enough intelligence I've
[72:22] got it I've solved all the problems in
[72:23] the world I think they're going to keep
[72:24] wanting it so what do you think about
[72:27] this incredible change and you know
[72:29] basically a hundred billion $200 billion
[72:32] you know depending on the company just
[72:34] going into capex uh this seems like a
[72:36] savvy move right this is all of us think
[72:39] this is a good I mean, if you want to
[72:41] bet against Google's deployment of
[72:44] capital into infrastructure
[72:47] because you'd rather have them give you
[72:48] cash for your shares today, you
[72:51] shouldn't own the stock and someone else
[72:54] will buy it. I think Google wins in a
[72:57] lot of different ways. There's just so
[72:58] much to Google. There's the consumer
[73:00] business, which is a lot of stuff.
[73:02] There's also YouTube. There's also GCP.
[73:04] There's also this portfolio of other
[73:06] bets, which by the way includes 10% of
[73:08] SpaceX.
[73:10] and a good chunk of Anthropic that they
[73:12] own and so on. Whimo worth took a
[73:15] hundred billion dollar write up on
[73:16] Anthropic in the quarter. So they had a
[73:18] hundred billion dollar mark to market on
[73:20] on Anthropic just in one quarter and
[73:23] they own a piece of all these
[73:25] businesses. So there's a lot to like
[73:27] about Google,
[73:29] but just on GCP,
[73:31] I think that there's probably no better
[73:34] suited
[73:36] enterprise layer than GCP to take
[73:39] advantage of capturing value with AI for
[73:42] that enterprise setting. I think
[73:44] >> I think you're just so much better
[73:45] because you have so much of your
[73:46] enterprise data, all your email, your
[73:49] drive, a lot of information that you
[73:51] would want to have AI have knowledge of
[73:53] and have AI have access to to improve
[73:55] workplace productivity. And then they're
[73:57] model agnostic. I mean, you can run any
[73:59] model you want and you can run any
[74:01] workflow you want and you don't have to
[74:02] be tied in. A lot of other cloud service
[74:05] providers, cloud SAS, they're kind of
[74:07] model dependent. It's run in a certain
[74:09] way. With Google, you can better tune
[74:11] your system how you want to tune it. And
[74:12] what's the worst worst worst case
[74:14] scenario? The worst worst worst case
[74:16] scenario is they have the lowest cost
[74:17] infrastructure in the world to run other
[74:19] people's models as a service like Elon
[74:21] did with Grock with the Colossus
[74:23] >> Elon Web Services seems to be doing
[74:25] pretty great.
[74:25] >> And look at the return Elon's making on
[74:27] the Colossus install. So I think if
[74:30] Google in the worst case scenario, none
[74:32] of their application layer stuff works,
[74:34] none of their network effects work and
[74:36] they don't have any good models, they're
[74:38] still going to have the world's best
[74:39] infrastructure they can print cash on
[74:41] for years if you believe in AI. So if
[74:43] you want to bet AI, I think the best
[74:45] public market stock to own is Google.
[74:47] And by the way, you also get YouTube,
[74:49] you also get the consumer, you also get
[74:50] everything else. The multiple is kind of
[74:52] ridiculous right now.
[74:53] >> We talked about this on a previous
[74:54] issue. I I think Chimath and you and I
[74:57] were talking about like what Apple
[74:58] should do next and I I think we both
[75:01] came to the conclusion it's like why not
[75:02] have Apple web services they have such
[75:04] great relationships with developers they
[75:06] have the app store they have this deep
[75:07] developer it's not so easy because you
[75:10] have to build a cloud service provider
[75:11] you have to build some critical
[75:12] infrastructure it's taken Amazon
[75:15] >> call it 17 years to perfect it
[75:18] >> it's taken Google
[75:20] call it 12 or 13 years to mostly catch
[75:24] up. But the minute that you sign up to
[75:27] be a web provider, Jason, and a cloud
[75:29] service provider, what you're really
[75:30] signing up for is 59s of reliability and
[75:33] uptime. And that is just extremely
[75:36] expensive. Getting to the first two
[75:37] nines, you know, 99% uptime. If you're
[75:40] hosting something for a pharma company
[75:42] or a defense company, you can probably
[75:44] do it for relatively cheaply. Getting to
[75:46] the third nine, 99.9 probably cost you
[75:50] in the billions. Getting to the fourth
[75:53] nine costs the tens of billions, but
[75:54] getting to that fifth nine costs
[75:55] hundreds of billions. And that takes a
[75:58] real investment and real technical
[76:00] skill, and there's only three games in
[76:02] town.
[76:03] >> Yeah, I think Tim Cook's not the guy to
[76:05] do it, but this new CEO might be uh
[76:06] since he's an engineer, but they've
[76:08] returned in the last decade.
[76:11] >> They could buy 900 billion, 755 in
[76:14] buybacks and 140 billion sachs in
[76:17] dividends. Just let that sink in. 900
[76:20] billion if they had invested that in
[76:22] Anthropic and SpaceX and other things.
[76:24] They just could have found some good
[76:25] uses for that money. But Sax, any
[76:27] thoughts here on what's happen?
[76:28] >> Who's to say? Sorry, but who's to say
[76:29] that the investors that got that money
[76:31] didn't find a good use for it?
[76:33] >> Oh, yeah. So, on society level, but I
[76:35] just think Apple could have been more
[76:36] ambitious if they just spent half that
[76:38] money instead of on buybacks and
[76:39] dividends on creating new products and
[76:42] actually releasing their car, maybe
[76:43] buying some interesting companies. I
[76:45] think they
[76:45] >> I think it was very much a do no harm
[76:47] capital allocation strategy which worked
[76:49] for the stock.
[76:50] >> It's going to be really interesting to
[76:53] see if John Turners flips the script.
[76:55] >> I think he does. I think he's going to
[76:56] be like engineer guy and yeah. All
[76:58] right, enough on the markets. The
[77:00] markets are going to do what markets do.
[77:01] We will talk about Iran and other stuff
[77:04] like that when there's more news for
[77:06] venture capitalists to comment on. Right
[77:07] now, it's just on on I know everybody
[77:09] keeps asking. I I don't think there's
[77:11] much for us to say on it. I do think
[77:12] there's a lot for us to say uh in
[77:14] socialism corner our new reoccurring uh
[77:18] theme here Freedberg every week there's
[77:20] more news coming out of socialism corner
[77:23] uh producer
[77:23] >> you can show my uh my videos on
[77:25] socialism going back six years
[77:26] >> let's add to the Freeberg socialism
[77:28] rants earlier this month
[77:31] city dictator mayor Zohan Mani
[77:36] >> don't don't misuse that term please come
[77:37] on
[77:39] >> a rental ripoff hearing at New York
[77:42] City's Tenement Museum. After the
[77:44] hearing, he introduced a rental ripoff
[77:46] report. If passed, bars landlords from
[77:48] charging applications for credit checks.
[77:52] It's going to let landlords require a
[77:54] credit check or the 40x rent income
[77:56] standard, but not both. Legally
[77:58] recognizes tenant unions and more. He's
[78:00] obviously frozen the rent for a year. At
[78:02] the hearing, an activist uh wore a COVID
[78:04] mask and referred to evictions as the
[78:06] violence of evictions. Here's your 22nd
[78:09] clip. The Mandani administration is
[78:12] emboldening us so that we no longer
[78:14] tolerate the violence of evictions as a
[78:17] matter of business as usual.
[78:21] >> What the What were we just watching? Is
[78:23] that the sax? Do you remember the guy
[78:25] from Fat Albert who had the hat like
[78:27] that?
[78:28] Can we pull that guy up FROM YOU
[78:31] REMEMBER THE GUY FROM FAT ALBERT who had
[78:34] the hat? What's his name? Oh my god.
[78:36] >> Is CO still happening? I thought CO was
[78:38] over. No, no, she just got a
[78:41] >> That was like a super duper mask. That
[78:43] was a super duper mask. That just wasn't
[78:44] like a little cloth one. That was one of
[78:46] these
[78:46] >> those big ones.
[78:48] >> I mean, combined with the hat, it was uh
[78:51] that was pretty There it is. I remember
[78:53] that guy from uh
[78:56] What is going on?
[78:57] >> Wait, let's we're going to give Freeberg
[78:58] a chance to do a rant because I mean I
[79:00] could do I could do a rant on this.
[79:01] >> No, no. Let's give Freeberg his This is
[79:03] Freeberg rare meat. Freeberg in 1787. In
[79:07] 1787, John Quincy Adams,
[79:09] >> there he goes,
[79:10] >> published a work called a defense of the
[79:12] constitutions of governments of the
[79:14] United States of America. And in that
[79:15] work, he had a comment. The moment the
[79:19] idea is admitted into society, that
[79:22] property is not as sacred as the laws of
[79:24] God, and that there is not a force of
[79:27] law and public justice to protect it,
[79:31] anarchy and tyranny commence. And then
[79:34] in 1791 he made the statement publicly,
[79:37] "Property must be secured or liberty
[79:40] cannot exist in an essay series."
[79:42] >> Can you unpack it and explain why he
[79:44] said that? Do you think
[79:45] >> fundamental to the foundation of the
[79:46] United States of America was this idea
[79:48] of private property rights? Because if
[79:50] you think about where everyone that came
[79:52] to America was coming from, there were
[79:54] these tyrannical governments, monarchies
[79:57] or whatever, where some overlord or some
[80:00] cabal could decide at any point to take
[80:04] the things that you have. You had no
[80:06] private property rights as an
[80:08] individual. They could come in, they're
[80:09] like, "That farm is my farm. You're
[80:11] actually a surf. I'm the lord. That
[80:13] thing is my thing. You have a right to
[80:15] use it because I vest you that right to
[80:17] use it. I am the all powerful. I am the
[80:20] tyrannical overseer of these lands. And
[80:22] it was that stasis that drove so many to
[80:25] come to the United States and say we
[80:27] want a place where individuals, one
[80:29] person can say, I own something and no
[80:32] one can take it from me. Private
[80:33] property rights are the foundations of
[80:35] liberty in America.
[80:37] this idea that you can then claim acts
[80:41] of violence, that you can then claim
[80:43] circumstances of of extraordinary,
[80:45] extravagant wealth and say to that
[80:47] individual, I now have a right, the
[80:50] government now has a right to take your
[80:52] private property ultimately leads to
[80:54] this tyrannical form. And it starts out
[80:56] as being an anarchctic because, and
[80:58] remember, anarchy is a temporary state.
[81:00] It's always in between one state and
[81:03] another. All anarchies end up in
[81:05] tyranny. Groups of people fight each
[81:06] other. They're all stealing [ __ ] from
[81:08] each other. Everyone just goes and takes
[81:09] and gets what they want. And eventually
[81:11] people coales. They form groups. And
[81:13] those groups become the more powerful
[81:15] groups. And the powerful groups end up
[81:16] winning and they become the tyranny over
[81:18] the mass. And that is why all anarchies
[81:20] eventually evolve into tyranny. So
[81:21] anarchy and tyranny are one and the
[81:23] same. And fundamentally what's going on
[81:25] with these socialist principles is that
[81:28] we are taking your private property and
[81:29] you no longer have rights on your
[81:31] private property. Whether you are a
[81:33] landlord or whether you are a wealthy
[81:35] person that we've deemed to have too
[81:37] much wealth, we now will have the rights
[81:39] to come in and take your property and
[81:41] control it and take it from you. And it
[81:43] always starts with this framing of
[81:45] moralistic intent. We are good. You are
[81:48] bad for the following reasons. You have
[81:50] committed violence against the people
[81:52] that live in your building. You have
[81:53] taken too much wealth and none of us
[81:55] have wealth. We have a right to go and
[81:56] take your wealth from you. It has always
[81:58] started from this point of view that you
[82:00] are bad. So the first framing is that
[82:02] the private property owner is evil and
[82:05] that the private property owner has
[82:06] committed an act of injustice against
[82:08] those who don't have the private
[82:10] property and that is the justification
[82:12] for taking away their private property
[82:14] rights and it is the beginning of this
[82:16] transition towards a tyrannical system
[82:19] which will ultimately be what I call
[82:20] this kind of great American polit or
[82:23] whatever socialist framework gets set up
[82:24] by the the cabal of the socialists and
[82:27] and what they're trying to put together.
[82:28] So all of these little acts while
[82:30] seeming ridiculous and insane and
[82:31] inappropriate in aggregate are the same
[82:34] thing. They're a transition away from
[82:36] private property rights which is the
[82:37] foundation of the United States of
[82:39] America. That's why I think we should
[82:40] all be so shocked. And to John Quincy
[82:42] Adams point we need to vehemently defend
[82:45] those rights. As soon as those rights
[82:47] start to slip away even in the tiniest
[82:49] way it is a cascading effect and
[82:52] everything will become tyrannical and it
[82:53] will be very ugly in the United States
[82:54] of America. Saxs, uh, what are your
[82:57] thoughts here on um, Comrade Mandami?
[83:00] >> Yeah, I just want to add a layer to this
[83:02] this idea. You know, the these DSA
[83:05] socialists say that evictions are are
[83:07] violence and they basically want to stop
[83:09] them. I think Freeberg's making the
[83:12] point that this deprivives the landlord
[83:13] of their property, and that's true, but
[83:16] I think we also have to stop and
[83:18] consider what this means for the other
[83:20] residents in these buildings. I mean,
[83:22] first of all, if the landlords aren't
[83:25] making income because they got a bunch
[83:26] of delinquent tenants in the building,
[83:29] they can't now pay for upkeep and
[83:30] maintenance. And so, these buildings
[83:32] become more dilapidated, and that
[83:33] affects the the other tenants. But also,
[83:36] I have a friend who manages these
[83:38] apartment buildings, and he makes the
[83:39] point that it's often these squatters,
[83:42] these delinquent tenants who should be
[83:44] evicted but can't, who make the worst
[83:46] neighbors. So, when you think about the
[83:49] problems in a in an apartment complex
[83:51] where you've got people creating noise
[83:54] at night, maybe they're playing loud
[83:55] music or you have people punching walls
[83:59] or there's disgusting smells coming from
[84:02] apartments or they're misusing common
[84:05] areas or you have drunken and disorderly
[84:08] behavior. I mean, this is all the kind
[84:10] of stuff that happens in these rent
[84:12] controlled apartment buildings. And you
[84:14] have to remember that there's
[84:16] long-standing residents, a lot of old
[84:18] people who can't afford to find a new
[84:20] place. They depend on the rent control.
[84:23] And when you can't evict that unruly
[84:25] tenant, yes, it it affects the landlord,
[84:28] but it affects the neighbors even more
[84:31] because now they're stuck in a downward
[84:32] spiral. And I think this is a problem
[84:34] with the progressive mindset across the
[84:36] board is that these DSA types are always
[84:39] these like highly educated and often
[84:43] affluent types and they can afford to
[84:44] have luxury beliefs about public spaces
[84:48] because they never use them, right? They
[84:49] don't use the bus or the subway or
[84:52] parks. And so when they get taken over
[84:54] by homeless drug addicts, they always
[84:56] defend the addicts as opposed to the
[84:59] middle class.
[85:00] >> An easy thing to do if you don't live in
[85:01] the tenderloin.
[85:02] >> Right. Exactly. And it falls the hardest
[85:04] on the working class because they
[85:05] actually need these amenities. They need
[85:07] the parks for their kids or they need to
[85:09] use a subway, right? And it's really a
[85:11] problem when you get people shooting up
[85:14] or doing drugs or, you know, defecating
[85:16] in a subway.
[85:18] >> You're walking your kids to school like
[85:19] the person in Maring County who has
[85:21] these luxury beliefs or on the upper
[85:22] east side, they just don't they're
[85:24] they're abstracted from this. They don't
[85:26] need to deal with it, right?
[85:27] >> We talked about this.
[85:28] >> And it's it's no different with these
[85:30] rent control departments. I think that's
[85:31] the important point here is that if over
[85:34] a period of several years you can't
[85:36] evict anyone no matter how problematic
[85:39] they are, you effectively turn these
[85:41] apartment buildings into the equivalent
[85:43] of housing projects. And you know that
[85:46] really affects decent people of modest
[85:49] income who have nowhere else to go and
[85:51] their quality of life suffers. And look,
[85:53] you know, the private equity wives in
[85:55] their gated communities will still feel
[85:56] good about themselves because, you know,
[85:59] they prevented these evictions, but it's
[86:00] the people in the building who are going
[86:02] to suffer the most.
[86:03] >> And well, these and they're Shimat,
[86:05] these people are not just thinking from
[86:06] first principles. If you want to solve
[86:08] the housing problem,
[86:11] anybody with any basic understanding of
[86:14] economics would just say, well, increase
[86:16] the supply and the price will go down.
[86:19] And it actually doesn't matter which
[86:20] supply you add. It doesn't matter if
[86:22] it's luxury units or multif family or
[86:24] single family. As long as there's more
[86:27] uh housing and there's transportation to
[86:29] get to it and to move people in and out,
[86:31] it'll be fine. As we've as I like I'm
[86:33] sitting here in Tokyo, like they figured
[86:35] this out a long time ago. Just build up
[86:37] and put more units in. They figured it
[86:39] out in Texas, Florida, Nevada. The only
[86:41] people who can't seem to figure this
[86:42] out, you know, is like New York, LA, and
[86:45] San Francisco just happens to be liberal
[86:48] elite
[86:50] enclaves. And it's not hard on a
[86:53] conceptual basis. Just allow people to
[86:55] build some more units and different
[86:57] types of units and then the price will
[87:00] go down. Chimath, any any thoughts here?
[87:02] The data from Austin says once you relax
[87:06] the permitting constraint, you'll get
[87:08] more units built. And for every unit
[87:10] that comes online, it literally drives
[87:14] down the rent. So if you want low rent,
[87:16] you need to have more units. If you want
[87:18] more units, you need to permit more
[87:20] aggressively. That's it. So, it's just a
[87:22] decision. The same political will that
[87:25] it would take Mandani to pass this law,
[87:27] he could actually pass some permitting
[87:29] reform and it would do a lot more good.
[87:31] The thing on private property, the
[87:32] reason I asked Freebrook to explain it
[87:34] is I really believe what he's saying is
[87:36] really important. It's like funny to me
[87:38] this idea that at the limit, let's just
[87:40] say you're driving your car and all of a
[87:42] sudden somebody jumps in it and they're
[87:45] like, "Well, you know, now I'm here. You
[87:46] can't kick me out." or you know, you go
[87:48] outside, you leave your door open to go
[87:50] get a FedEx package or Amazon box and
[87:52] somebody runs in and sits on your couch
[87:54] and now all of a sudden they can't you
[87:56] can't kick them out. And it sounds so
[87:59] ludicrously dumb.
[88:01] If you force the owners of physical
[88:04] property to not be able to credit check
[88:06] and differentiate who they rent their
[88:08] apartments to, what's going to happen is
[88:10] rents will go up even more. So, I
[88:11] suspect that what they should do is they
[88:15] should pass this law and they should
[88:18] observe what the outcome is and then you
[88:19] can do a pretty scientific AB comparison
[88:22] between New York City and Austin and
[88:23] you'll know what works and what doesn't
[88:25] work.
[88:26] >> Yeah.
[88:28] >> I mean, here's the problem is is the
[88:30] socialists never learn. I mean, we
[88:31] already have
[88:32] >> got rid of rent control in Argentina.
[88:34] Rents went down.
[88:36] >> Yeah.
[88:36] >> Yeah. Yeah, I mean this is the problem
[88:37] with Jimoth is that look, if the
[88:39] socialists ever learned from their
[88:40] failed experiments, you wouldn't have
[88:43] Chicago, you know, it's just you like we
[88:46] don't need New York to go down the tubes
[88:48] to know this isn't going to work cuz
[88:50] it's happened in so many other places
[88:51] already. But somehow it just never seems
[88:53] to stop.
[88:54] >> They never seem to learn. So they should
[88:55] run the experiment and learn. I mean
[88:57] what's crazy is you'll be learning and
[88:59] it'll be a failure on the grandest stage
[89:01] possible. You're talking about the
[89:03] biggest, most complicated New York City.
[89:04] My gosh.
[89:05] >> I mean, and and Sachs, what happens if a
[89:08] landlord
[89:09] >> cannot raise the rent reasonably? Well,
[89:13] they have no incentive to invest in new
[89:16] units. They have no incentive to upgrade
[89:18] new units. And there's this trap in New
[89:19] York City specifically. They have made
[89:22] this regulations for apartments such
[89:24] that if you do a renovation, it has to
[89:27] hit certain codes. And there are a ton
[89:29] of codes. Okay. So that means it's
[89:31] incredibly expensive. So now they have
[89:35] you have ghost apartments.
[89:37] >> Yes. So now the housing stock becomes
[89:38] dilapidated. And look, they're doing
[89:40] something even worse now, I think, which
[89:41] is they are banning landlords from doing
[89:44] credit and background checks on
[89:45] potential tenants. And they they say
[89:48] that you can't look at their income, so
[89:50] you can't vet whether they can actually
[89:52] pay the rent. So they're banning
[89:54] >> landlord. No more landlord.
[89:55] >> They're they're banning eviction. And
[89:57] then they're and then they're preventing
[89:58] you from doing the diligence to see if
[90:00] this is even a tenant who will pay the
[90:02] rent.
[90:02] >> We've lost the script.
[90:03] >> So what are you supposed to do?
[90:04] >> The interesting question is if you were
[90:06] forced to live under these rules, the
[90:08] landlord, what would you do? And the
[90:10] obvious answer is you'd start rent three
[90:12] or four times higher and you force
[90:15] people to sign up to a multi-month
[90:17] prepayment and you'd slowly ease those
[90:20] conditions until you find a market
[90:21] clearing price. That's the only way to
[90:23] do it. So rents will not go down. Rents
[90:26] will go up. So run the experiment
[90:28] >> and let's just observe what happens.
[90:30] >> That's a really good point, Jimoth. Like
[90:32] let's say that you know X percent of
[90:34] tenants are going to become delinquent,
[90:36] right? Because and by the way, what's
[90:37] their incentive to pay when they know
[90:38] they can't be evicted?
[90:39] >> Zero. Zero.
[90:40] >> So like actually a pretty significant
[90:42] percentage of people could just decide
[90:44] I'm going to make rent optional.
[90:46] >> And so now the landlord has to absorb
[90:48] those losses.
[90:49] >> Exactly.
[90:50] >> And that means they have to pass on a
[90:51] higher rent to everybody else. You're
[90:53] going to set the rent 3x higher and
[90:55] you're going to slowly meander it down.
[90:56] And like I said, you're going to have to
[90:58] wire in the first full year of rent.
[91:01] >> Well, good luck. How is that affordable?
[91:03] >> It's even worse, Sax. Not only do the
[91:06] landlords
[91:08] keep the dilapitated apartments in some
[91:11] cases, it's better for them to just
[91:14] leave apartments, housing stock, empty.
[91:17] So, somebody leaves, they are forced to
[91:20] renovate it and it costs more, you know,
[91:23] hundreds of thousands in uh renovations.
[91:26] So, they say, "You know what? I'll just
[91:27] leave it empty for now." And so, you
[91:30] have 50,000 ghost apartments according
[91:32] to reports in New York City.
[91:34] >> Well, that's an Airbnb problem. That's
[91:36] like a
[91:37] >> No, they banned Airbnb in New York. You
[91:38] cannot get an Airbnb.
[91:40] >> Oh, really?
[91:41] >> Yes. So, now it's like
[91:42] >> That's really interesting. Yeah. Look,
[91:44] if you're if you're a landlord, okay, I
[91:46] mean, I guess one thing you would do is
[91:47] just sell and go to another
[91:48] jurisdiction. Another thing you could do
[91:50] is just wait this out
[91:51] >> because it's not profitable to run an
[91:54] apartment building. You can't raise your
[91:55] rents. You can't evict people. You can't
[91:57] diligence the tenants. So, maybe you
[92:00] just leave the building empty and you
[92:01] wait this out. I mean, that's assuming
[92:03] you don't have too much debt on it,
[92:04] right?
[92:04] >> Or ghost apartments. Yeah.
[92:06] >> And then you have ghost apartments.
[92:07] >> Yeah. All right. Listen, you've been
[92:09] thinking about coming to the All-In
[92:10] Summit. This is your year. speakers are
[92:12] world class. The events, the parties,
[92:15] the networking, 80% of the people there,
[92:18] founders, investors or high level
[92:19] operators this year. Greater focus on
[92:21] network,
[92:22] >> the musical performances.
[92:23] >> Oh yes. I mean, in the past, we've had
[92:25] >> the food, the drink,
[92:27] >> Grimes,
[92:28] >> Diplo,
[92:28] >> Diplo. We've had so many incredible
[92:31] people. Go to the all-inssummit.com. All
[92:33] right, everybody. It's another amazing
[92:36] alltime legendary episode of the All-In
[92:39] podcast.
[92:41] Love you boys. Bye. Byebye.
[92:45] >> Let your winners ride.
[92:52] >> And it said
[92:53] >> we open sourced it to the fans and
[92:54] they've just gone crazy with it. Love
[92:56] you.
[93:05] >> Besties are gone.
[93:08] That is my dog taking an driveway.
[93:13] >> Oh man, my habitasher will meet me up.
[93:15] >> We should all just get a room and just
[93:17] have one big huge orgy cuz they're all
[93:18] just It's like this like sexual tension
[93:21] that we just need to release somehow.
[93:29] >> We need to get mercy.
[93:38] I'm going all in.
