Nvidia Signs $50 Billion Texas Lease as Big Tech Resumes Hiring
TBPN
Corporations are accelerating headcount expansion alongside AI tools as early fears of widespread job displacement give way to productivity gains and rising demand.
## Background
Published by TBPN on YouTube on July 28, 2026, this episode of Diet TBPN delivers a 30-minute summary of today's live tech show hosted by John Coogan and Jordi Hays. According to the page description, the show covers daily tech news—including big company hiring trends, Anthropic's open-weight stance, and Zuckerberg's perspective on AI—and features high-profile tech leaders like Mark Zuckerberg and Sam Altman.
## TL;DR
The episode examines a major shift in the AI jobs narrative as major corporations resume hiring human workers alongside AI productivity tools, defying early predictions of catastrophic job losses. The hosts discuss recent corporate moves and market shifts, including Black Sheep's guerilla truck campaign targeting Google, Recursive Superintelligence's $410 million AWS compute deal, and Nvidia's massive $50 billion data center lease in Texas. They also break down Dario Amodei's position on AI safety, open-weight models, and chip controls against China, alongside Mark Zuckerberg's op-ed defending open-source AI. Finally, they cover Apple's new Klarna hardware leasing program and playfully reflect on their own show production setup.
## Atmosphere & style
The hosts maintain a sharp, fast-paced, and banter-filled tone, blending savvy Silicon Valley analysis with irreverent tech comedy. The dynamic feels like a lively, candid conversation between plugged-in insiders who balance serious commentary on multi-billion-dollar deals with humorous visual gags, such as reviewing 'business villain' trading cards and joking about upgrading studio gongs. The overall atmosphere is engaging and lighthearted yet grounded in a deep understanding of corporate strategy and modern AI economics.
## Key takeaways
- [00:01](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1s) Major corporations across tech, defense, and other sectors are resuming hiring alongside AI tools, refuting initial fears of widespread job loss and demonstrating Jevons paradox.
- [00:01](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1s) Executive teams previously used AI as a public relations excuse to cover up overhiring and underlying business issues during layoff cycles.
- [00:48](https://www.youtube.com/watch?v=QmVHay-kPEc&t=48s) AI inside lean teams is primarily replacing low-complexity freelance work rather than replacing core human employees.
- [00:48](https://www.youtube.com/watch?v=QmVHay-kPEc&t=48s) Lattice and Cloudflare highlight that tech companies are actively hiring entry-level talent and software engineers again to integrate them into legacy teams to accelerate AI adoption.
- [07:58](https://www.youtube.com/watch?v=QmVHay-kPEc&t=478s) Guerilla marketing tactics like Black Sheep's fleet of 25 LED billboard trucks around Google's Chelsea office show how physical stunts are used to publicize corporate disputes over ad spending glitches.
- [14:10](https://www.youtube.com/watch?v=QmVHay-kPEc&t=850s) Ilya Sutskever's Safe Superintelligence (SSI) clarifies that scaling research means testing architectural flywheels on larger compute clusters rather than preparing an immediate public product launch.
- [15:54](https://www.youtube.com/watch?v=QmVHay-kPEc&t=954s) Recursive Superintelligence signed a $410 million compute deal with AWS to build open-ended self-improving systems through non-circular financing.
- [17:35](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1055s) Nvidia signed a $50 billion lease for Hut 8's 1-gigawatt data center in Texas to house hundreds of thousands of GPUs, pushing Nvidia's market cap past $5 trillion.
- [20:16](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1216s) Anthropic CEO Dario Amodei opposed blanket bans on open weights while proposing strict export controls on AI chips to China, crackdowns on industrial-scale distillation, and mandatory safety testing.
- [26:18](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1578s) Mandatory safety testing risks creating FDA-style regulatory bottlenecks for smaller AI startups while tech giants skip ahead in approval queues.
- [31:17](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1877s) Mark Zuckerberg published an op-ed in The Wall Street Journal arguing that open-source software and open-weight models provide the best long-term framework for safety, security, and economic opportunity.
- [33:16](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1996s) Apple partnered with Klarna to launch 'Apple Upgrade,' a hardware leasing and subscription program for iPhones, iPads, Macs, and Apple Watches in response to rising device costs.
## Detailed walkthrough
Here is the detailed walkthrough of this portion of the transcript:
[00:01](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1s) The AI Jobs Narrative Flips: Big Companies Resume Hiring
The hosts discuss a major shift in the ongoing debate around AI, labor markets, and hiring, highlighted by a front-page article in *The Wall Street Journal*. Early predictions from AI lab leaders foresaw catastrophic job losses, but current realities show productivity boosts, slower technology diffusion, and new capabilities rather than direct "drop-in replacements" for human co-workers. After roughly a year of cautious, restrained hiring across tech, transportation, defense, and other sectors, major corporations are actively expanding their headcount to work alongside AI tools—an outcome framed as an example of Jevons paradox, where increased efficiency drives up overall demand and total labor needs.
```
┌──────────────────────────────────────────┐
│ AI Efficiency Increases │
└────────────────────┬─────────────────────┘
│
▼
┌──────────────────────────────────────────┐
│ Demand & Volume Escalates │
└────────────────────┬─────────────────────┘
│
▼
┌──────────────────────────────────────────┐
│ Overall Need for Human Workforce Grows │
└──────────────────────────────────────────┘
```
The discussion turns to how corporate leaders previously cited AI when announcing major layoffs over the past year. The hosts critique this tactic, calling it PR spin that allowed executive teams to cover up overhiring or underperformance rather than admitting structural business issues. They praise leaders who offer total honesty, referencing a recent candid statement from the new CEO of Xbox. Additionally, companies with tens of thousands of employees often use layoff cycles to prune low performers—who might excel at 90th-percentile levels at a different firm—and bring in fresh talent suited for updated strategic goals.
* **Firms Signaling Renewed Hiring:** CSX, Alphabet, ServiceNow, Snap-on, and Booz Allen Hamilton.
* **Jobless Claims Data:** Recent US weekly jobless claims dropped to near-decade lows, pointing to a resilient labor market.
* **Lattice Insights:** Sarah Franklin, CEO of HR platform Lattice, noted that initial expectations that AI agents would replace entry-level workers have faded. Lattice is seeing clients resume hiring for junior roles and software engineers alike.
* **Robert Half Perspective:** CEO M. Keith Waddell described AI's impact on employment as "more benign than some had feared," with hiring demand steadily improving.
[00:48](https://www.youtube.com/watch?v=QmVHay-kPEc&t=48s) Practical Uses of AI vs. Core Talent
Reflecting on their own operations as a niche daily tech show, the hosts examine how AI actually functions inside a lean team. Rather than replacing core employees, AI replaces low-complexity, Upwork-style freelance work—such as generating a joke song via Suno or spinning up a quick, funny landing page. Core business functions still require dedicated, accountable humans leveraging AI tools to increase their output.
They also debate differing strategies on hiring entry-level talent:
* **The "AI Backfill" Approach:** Investor Bryce Roberts highlighted a text exchange stating a company was backfilling departing roles with AI rather than hiring, commenting "RIP new grads."
* **The Counter-Argument:** Cloudflare CEO Matthew Prince argued against halting new grad hiring. His recommended playbook is to hire entry-level talent and embed them directly into legacy teams to accelerate AI adoption. Cloudflare notably hired roughly a thousand workers recently to support its trajectory.
> "I'm actually closer than ever to hiring a full-time Celtic punk band to play music to recreate songs... where that was not even on the roadmap a few years ago."
[07:58](https://www.youtube.com/watch?v=QmVHay-kPEc&t=478s) Business Villain Trading Cards and Black Sheep vs. Google
The hosts view a satirical viral video reel from *Good Work*, showcasing custom-printed, high-production "business villain" trading cards. The physical cards feature notable figures including Bernie Madoff (from an 80s-themed set with physical glove relics), Sam Bankman-Fried (crafted on balsa wood), Elizabeth Holmes, and a "vintage '05" Mark Zuckerberg card complete with an embedded piece of underwear material. They remark on how business comedy has solidified a distinct canon around these figures, similar to high-profile athlete collectibles.
```
┌─────────────────────────────────────────────────────────────────────────┐
│ "BUSINESS VILLAIN" TRADING CARDS │
├───────────────────┬─────────────────────────────┬───────────────────────┤
│ Figure │ Physical Card Material/Relic│ Vintage Style │
├───────────────────┼─────────────────────────────┼───────────────────────┤
│ Bernie Madoff │ Fabric Glove Piece │ 1980s Retro │
│ Sam Bankman-Fried │ Balsa Wood │ Modern │
│ Mark Zuckerberg │ Fruit of the Loom Patch │ 2005 Draft │
│ Elizabeth Holmes │ Standard Multi-Card Set │ Modern │
└───────────────────┴─────────────────────────────┴───────────────────────┘
```
The conversation then moves to an aggressive, real-world guerilla marketing stunt taking place in Manhattan targeting Google's Chelsea office. Eyewear company **Black Sheep** deployed a fleet of 25 mobile LED billboard trucks driving around the headquarters carrying messages like *"Shame on you, Google. Return our $80,000."*
According to Black Sheep—a factory-direct optical seller whose $8 sunglasses beat a $350 pair in NBC lab tests—their organic traffic surged after an *NBC Today Show* segment. During the spike, Google's organic search engine indexed a broken 404 error page for their brand, routing user traffic away from organic links and forcing the company into spending heavily on Google's paid search listings.
* **Financial Impact:** Black Sheep claims Google's search engine glitch drained **$77,000** (or nearly $80,000) in ad spend over a 30-hour period.
* **The Guerilla Campaign:** Deployed 25 LED billboard trucks to surround Google's office, generating viral attention that likely exceeds $77,000 in promotional value.
* **Host Takeaways:** While Black Sheep framed the trucks as "minimalist," the hosts note the stunt was maximalist in scale, calling mobile truck billboards an underrated surface area for corporate disputes and publicity stunts.
[14:10](https://www.youtube.com/watch?v=QmVHay-kPEc&t=850s) SSI Speculation and Recursive Superintelligence's AWS Deal
The hosts address speculation surrounding Ilya Sutskever’s AI firm, **Safe Superintelligence (SSI)**, following comments from Ilya about scaling their research. They clarify that scaling research simply means testing specific architectural flywheels with larger compute clusters—funded by recent capital raises—rather than an imminent public product launch, despite online rumors.
They also laugh over a notable headline typo in tech media regarding **Recursive Superintelligence** (a startup that emerged from stealth in May with $650 million in funding to build self-improving systems). A header typo erroneously read that the company signed a "$410 million million" compute deal with Amazon Web Services.
* **Key Data Point:** Recursive Superintelligence signed a **$410 million** multi-year compute agreement with AWS to fuel its compute-intensive approach to open-ended, self-improving AI research.
[15:54](https://www.youtube.com/watch?v=QmVHay-kPEc&t=954s) Compute Deals and Non-Circular Financing
Continuing the discussion on AI compute deals, the hosts examine Recurse's $410 million outlay, which represents the bulk of the company's total fundraising to date. On a call with *TechCrunch*, Recurse Founder and CEO Richard Socher emphasized that he expects this to be merely the first of many such agreements. While AI compute deals are typically intricate and tied to circular equity investments—such as Amazon investing in a startup while requiring them to purchase Trainium chips and host infrastructure on AWS—this agreement appears surprisingly straightforward.
The hosts note that they have become so desensitized to circular deal structures that finding a vanilla transaction feels almost unusual. Jason, VP of Startups and VC at AWS, mentioned that part of the agreement includes co-developing infrastructure purpose-built for these types of companies, leaving open the possibility that some traditional circularity may still manifest down the line.
- Recurse allocated $410 million out of its total fundraising pool for AWS compute.
- AWS and Recurse plan to co-develop specialized infrastructure tailored for AI workloads.
---
[17:35](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1055s) Nvidia's $50 Billion Data Center Lease and Tech Valuations
Nvidia has been revealed as the tenant for a massive $50 billion data center in Texas that will house hundreds of thousands of its own GPUs. CEO Jensen Huang is aggressively deploying Nvidia's balance sheet to backstop the growth of the AI compute market. Nvidia signed the lease for an entire 1-gigawatt facility being constructed by developer Hut 8, marking a unprecedented single-site commitment that underscores the company's growing role in financing the broader AI ecosystem.
Looking at market caps, Nvidia surged over 3% to reach a $5.17 trillion valuation, briefly crossing $5 trillion alongside Apple ($4.99 trillion). The hosts joke that Apple is running a "do nothing and win" strategy while Jensen Huang executes thousands of deals.
- Nvidia signed a $50 billion lease commitment for Hut 8's 1-gigawatt data center in Texas.
- Nvidia's market cap hit $5.17 trillion, competing closely with Apple at $4.99 trillion.
---
[19:02](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1142s) The Open Letter Debates: Apple, Amazon, and Anthropic
A major industry open letter supporting open-weight models and opposing open-source AI bans recently circulated. Only three major tech giants declined to sign it: Amazon, Apple, and Anthropic. The hosts speculate that Apple and Amazon are somewhat "unsloppable" because their core businesses rely on physical infrastructure—such as iPhone hardware and physical Amazon warehouses—that cannot simply be "vibe-coded" away. Additionally, Apple's refusal to sign fits its classic brand posture of standing apart from group consensus.
- The three key tech players abstaining from the open-weight letter are Apple, Amazon, and Anthropic.
---
[20:16](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1216s) Anthropic's Position on Open Weights, Geopolitics, and Distillation
Anthropic CEO Dario Amodei issued a direct response outlining the company's official stance on AI governance and open-weight models. Amodei clarified three foundational premises:
1. Anthropic has never advocated for a blanket ban on open-weight models.
2. The United States must maintain its AI lead over authoritarian governments like China to safeguard democratic freedoms.
3. Powerful AI models pose inherent national security risks, such as potential misuse in cyberattacks or biological weapons deployment.
To address these challenges, Amodei outlined three key policy proposals:
1. **Maintain Strict Chip Controls:** Continue sanctioning and restricting the export of advanced AI chips and semiconductor manufacturing equipment to China. The hosts argue that restricting sales preserves America's multi-year technological edge, especially since domestic US demand for chips remains virtually insatiable.
2. **Crack Down on Industrial-Scale Distillation:** Implement policy measures to deter foreign entities or competitors from systematically harvesting API outputs from frontier models (such as Claude) to train rival open models.
3. **Mandatory Safety Testing:** Subject all sufficiently powerful frontier models—both open and closed—to standardized safety evaluations prior to deployment.
- Dario Amodei highlighted security risks surrounding cyberwarfare and biological threats from unrestricted models.
- Chip export restrictions remain a vital lever to preserve the United States' competitive technological gap over China.
---
[22:59](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1379s) Industrial Distillation and the Synthetic Data Gray Area
Defining and enforcing prohibitions against "industrial-scale distillation" presents significant legal and technical hurdles. Synthetic data usage exists in a broad gray area; for example, Thing Machines' Inkling model utilized synthetic data generated by Kimi k2.5. While benefiting from foreign open models, such workflows do not necessarily constitute malicious industrial distillation.
In practice, frontier models are often distilled from a mix of sources—combining outputs from hypothetical future models like "Opus 5," "GPT-5.6," and Mistral fine-tunes with custom RL environments. Proving distillation behind closed doors is extremely difficult. In Washington, DC, legal negotiations will likely hinge on whether AI labs can provide concrete API access logs and quantitative evaluations to demonstrate intellectual property theft.
- Synthetic data generated by models like Kimi k2.5 illustrates the nuance between legitimate data generation and predatory distillation.
- Legal remedies against foreign labs remain difficult to enforce due to jurisdictional constraints.
---
[26:18](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1578s) Mandatory Safety Testing and Containing Rogue Models
Mandatory safety testing—a policy also championed by Demis Hassabis at Google DeepMind—aims to prevent catastrophic risks. However, the hosts warn of severe regulatory bottlenecks. Smaller startups without lobbying presence in Washington, DC (such as Recursive Superintelligence) could be stuck in FDA-style approval queues for years while trillion-dollar tech giants skip to the front of the line.
The hosts also explore extreme adversarial scenarios: if a rogue foreign lab distills frontier capabilities, strips out safety guardrails for cyber/bio capabilities, and dumps the weights onto torrent trackers or Hugging Face, what options remain for government enforcement?
While low-level IP-infringing models might warrant minimal intervention, hyper-capable malicious models would compel aggressive government action. Federal enforcement would likely focus on blocking physical infrastructure: pressuring neoclouds, denying hosting in domestic data centers (like NVL72 clusters), and enforcing strict network firewalls to restrict deployment inside the US.
- Google DeepMind's Demis Hassabis and Anthropic's Dario Amodei both advocate for mandatory safety evaluations for frontier AI.
- Government intervention against rogue open-weight models would primarily operate at the infrastructure level by restricting US data centers and neocloud access.
[31:17](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1877s) Zuckerberg's "AI Future for Everyone" & Open-Source Defense
Mark Zuckerberg published a new opinion piece in *The Wall Street Journal* titled "The AI Future Is for Everyone," in which he argues that "the history of democracy and economics has proved that centralized power stifles human potential." The hosts discuss how Zuckerberg is actively positioning Meta to fill the void as the pro-openness tech leader, championing open-weight models while investing hundreds of billions of dollars annually in AI infrastructure.
Addressing potential monopoly critiques, the show notes that despite Meta's massive size, it faces relentless competition across the broader attention economy—competing against TikTok, Snapchat, LinkedIn, Twitch, YouTube, Netflix, podcast feeds, SMS, and iMessage. Crucially, Zuckerberg advocates that open-source software remains the premier pathway for security and safety, arguing that providing broad access to powerful AI systems offers the most robust long-term defense.
- Zuckerberg's *Wall Street Journal* op-ed highlights open-source software and open-weight models as the optimal safeguard for AI safety and cybersecurity.
- The hosts point out Meta's unique position in the market: investing heavily in AI while championing democratized access rather than centralized control.
[33:16](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1996s) Apple Upgrade Hardware Leasing & The Economics of Premium Tech
Apple is introducing a new hardware leasing and subscription program called "Apple Upgrade" in partnership with Klarna. Launching across U.S. online and retail channels, the program allows users to lease iPhones, iPads, Macs, and Apple Watches on a recurring monthly basis. The hosts joke about the classic WEF phrase, "You will own nothing and you will be happy," noting that hardware pricing has shifted drastically from the subsidized $200 carrier phone days to modern premium devices pushing $2,000.
The discussion pivots into a humorous banter about the hosts' own show setup, comparing soaring hardware costs to the price of their show's famous gongs. They reflect on the "hedonic treadmill" of upgrading studio gongs over time—starting small and continually moving up to larger, more electric instruments—with one host suggesting it is time to buy an enormous new gong suspended directly from the rafters.
- Apple Upgrade leasing pricing tiers:
- **iPhone:** Starting as low as $17.99/month
- **iPad:** $11.99/month
- **Apple Watch:** $11.99/month
- **Mac:** $24.99/month
- The move toward hardware leasing is viewed as a strategic reaction to rising memory prices and higher overall product manufacturing costs.
- The hosts close out the show reminding viewers to leave five-star ratings on Apple Podcasts and Spotify, sign up for the daily newsletter at TBPN.com, and tune in live weekdays at 11:00 AM Pacific.
## Notable quotes
> [00:48](https://www.youtube.com/watch?v=QmVHay-kPEc&t=48s) I'm actually closer than ever to hiring a full-time Celtic punk band to play music to recreate songs... where that was not even on the roadmap a few years ago.
> [00:48](https://www.youtube.com/watch?v=QmVHay-kPEc&t=48s) RIP new grads. — Bryce Roberts
> [07:58](https://www.youtube.com/watch?v=QmVHay-kPEc&t=478s) Shame on you, Google. Return our $80,000.
> [00:01](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1s) more benign than some had feared — M. Keith Waddell
> [31:17](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1877s) The AI Future Is for Everyone — Mark Zuckerberg
> [31:17](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1877s) the history of democracy and economics has proved that centralized power stifles human potential. — Mark Zuckerberg
## Mentioned
John Coogan, Jordi Hays, Mark Zuckerberg, Sam Altman, Mark Cuban, Satya Nadella, Ramp, Public, Cisco, Console, CrowdStrike, Figma, MongoDB, NYSE, Railway, Shopify, The Wall Street Journal, CSX, Alphabet, ServiceNow, Snap-on, Booz Allen Hamilton, Lattice, Sarah Franklin, Robert Half, M. Keith Waddell, Upwork, Suno, Bryce Roberts, Cloudflare, Matthew Prince, Good Work, Bernie Madoff, Sam Bankman-Fried, Elizabeth Holmes, Black Sheep, Google, NBC Today Show, Ilya Sutskever, Safe Superintelligence, SSI, Recursive Superintelligence, Amazon Web Services, AWS, TechCrunch, Richard Socher, Trainium, Jason, Nvidia, Hut 8, Apple, Amazon, Anthropic, Dario Amodei, Claude, Thing Machines, Inkling, Kimi k2.5, Mistral, Hugging Face, Demis Hassabis, Google DeepMind, Meta, TikTok, Snapchat, LinkedIn, Twitch, YouTube, Netflix, Klarna, iPhone, iPad, Mac, Apple Watch, Apple Upgrade, TBPN
Corporate labor markets and artificial intelligence strategies are moving past initial workforce displacement fears into heavy infrastructure capital deployment, resumed hiring, and intense policy debates over open-weight models. As frontier firms commit tens of billions of dollars to data center capacity, industry leaders remain divided between open-source democratization and stringent regulatory security controls.
AI Labor Trends and Resumed Corporate Hiring (0:01)
Contrary to early predictions warning of rapid, AI-driven job destruction, major corporations across technology, defense, and transportation—including Alphabet, ServiceNow, CSX, Snap-on, and Booz Allen Hamilton—are actively expanding their headcounts. The hosts view this shift through Jevons paradox, where increased technical efficiency reduces costs, stimulates total demand, and expands aggregate labor needs. Supporting indicators include recent US weekly jobless claims falling near decade lows. Furthermore, Lattice CEO Sarah Franklin and Robert Half CEO M. Keith Waddell report that client demand for junior roles and software engineers is recovering, softening earlier fears that autonomous agents would eliminate entry-level positions.
The discussion critiques previous corporate layoff announcements that cited AI, categorizing them as executive public relations cover for post-pandemic overhiring or strategic underperformance. Approaches to entry-level talent remain contested: while investor Bryce Roberts observed companies backfilling departing positions with AI tools instead of new graduates, Cloudflare CEO Matthew Prince advocates hiring entry-level workers to accelerate AI adoption across legacy engineering teams, a strategy reflected in Cloudflare's recent addition of roughly 1,000 employees.
Capital Deployments in Data Centers and Compute (15:54)
Frontier AI research is driving massive infrastructure financing commitments. Nvidia signed a $50 billion lease commitment for an entire 1-gigawatt data center facility constructed by Hut 8 in Texas, intended to host hundreds of thousands of its own GPUs. CEO Jensen Huang is directly deploying Nvidia's balance sheet to backstop the sector's compute capacity, a push coinciding with Nvidia reaching a $5.17 trillion market capitalization alongside Apple at $4.99 trillion.
Simultaneously, early-stage AI labs are concentrating their capital into cloud compute infrastructure. Recursive Superintelligence committed $410 million of its $650 million funding pool into a multi-year AWS agreement, with CEO Richard Socher projecting it as the first of multiple computing partnerships. The hosts observe that while many frontier computing contracts feature complex, circular equity and chip-purchasing requirements, the Recurse-AWS transaction appears unusually direct. Meanwhile, the hosts clarify that Safe Superintelligence (SSI), founded by Ilya Sutskever, is scaling research compute clusters rather than preparing an immediate commercial product release.
Regulatory Battles Over Open Weights and Distillation (20:16)
Tensions over open-weight AI intensified after Apple, Amazon, and Anthropic declined to sign an industry open letter opposing open-source model bans. Anthropic CEO Dario Amodei clarified that the company does not seek a blanket ban on open-weight models, but advocated for three specific policies: maintaining strict semiconductor export controls to China to protect US competitiveness, instituting mandatory pre-deployment safety testing for frontier systems, and restricting industrial-scale distillation by foreign competitors who harvest API outputs to train rival models.
Enforcing distillation restrictions involves significant technical gray areas, as legitimate frontier model training frequently incorporates synthetic data generated by other models, such as Thing Machines' Inkling using outputs from Kimi k2.5. The hosts also caution that mandatory safety evaluations—also endorsed by Google DeepMind CEO Demis Hassabis—could create bureaucratic approval queues that disproportionately disadvantage early-stage startups. Conversely, Meta CEO Mark Zuckerberg argued in a Wall Street Journal op-ed that centralized AI power stifles innovation, asserting that open-source models offer the strongest foundation for long-term safety and cybersecurity.
Commercial Search Disputes and Hardware Leasing Models (7:58)
Platform dynamics and hardware economics are driving new business models and corporate friction. Eyewear maker Black Sheep launched a guerilla marketing campaign using 25 mobile LED billboard trucks around Google's Chelsea office, claiming a Google organic search indexing error displayed a 404 page during a traffic surge and forced the brand to spend approximately $77,000 to $80,000 on paid search ads over 30 hours.
Concurrently, elevated component and memory manufacturing expenses are shifting consumer hardware distribution. Apple introduced "Apple Upgrade" in partnership with Klarna, offering monthly hardware leasing across US retail channels starting at $11.99 per month for iPads and Apple Watches, $17.99 for iPhones, and $24.99 for Macs. The hosts note this reflects a broader transition toward subscription models as outright premium device prices approach $2,000.
The source portrays an AI ecosystem transitioning from speculative labor disruption into capital-heavy infrastructure deployment and defensive regulatory maneuvering. Key unresolved tensions remain regarding how policymakers can effectively govern synthetic data distillation across borders without entrenching large incumbents at the expense of open-source models and startups.
## Background
该视频由科技脱口秀频道 TBPN（由 John Coogan 与 Jordi Hays 共同主持）于 2026 年 8 月 3 日发布在 YouTube 平台上。根据节目简介，本期“Diet TBPN”在约 30 分钟内浓缩了当天直播节目的精华，重点讨论了大企业重启招聘、Anthropic 针对开源权重的监管立场、扎克伯格力挺开源 AI 以及苹果推出硬件订阅租赁服务等前沿商业与科技动态。
## TL;DR
节目首先指出了“AI 引发失业潮”公关叙事的破产，揭示技术提效带来的“杰文斯悖论”正推动 Alphabet、ServiceNow 等大厂重新扩招。随后，主持人们梳理了 AI 算力基础设施的巨额交易，包括初创公司 Recursive 与 AWS 签订的 4.1 亿美元算力合同，以及英伟达斥资 500 亿美元租赁 1 吉瓦得州超级数据中心的生态闭环。针对开源与闭源的政策争论，节目深度剖析了 Anthropic 首席执行官 Dario Amodei 提出的芯片出口管制、严打工业级模型蒸馏与推行强制安全测试等主张，并与马克·扎克伯格在《华尔街日报》撰文高调捍卫开源普惠的立场形成鲜明对比。最后，节目还评述了苹果联合 Klarna 试水全系硬件月租订阅以应对硬件涨价的全新商业举措。
## Atmosphere & style
整期节目的氛围节奏飞快、犀利毒舌且洋溢着浓郁的硅谷极客幽默与华尔街行内人视角。两位主持人配合默契、插科打诨不断，将严肃硬核的地缘芯片博弈、监管俘渴与算力资本运作，巧妙地包裹在对商业反派球星卡、开卡车围堵谷歌曼哈顿总部以及演播室铜锣通胀等荒诞趣闻的捧腹调侃中，兼具极强的娱乐观赏性与敏锐的商业洞察力。
## Bulletin
- [00:01](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1s) 经历一年多的观望与紧缩后，科技及多行业巨头重新开启扩招，证明 AI 提高效率后不仅未引发全面失业，反而因“杰文斯悖论”催生了更多协同岗位需求。
- [17:35](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1055s) 英伟达作为承租方签署总值高达 500 亿美元的得州 1 吉瓦超级数据中心协议，亲自下场托底算力资产，单日市值跃升至 5.17 万亿美元。
- [14:11](https://www.youtube.com/watch?v=QmVHay-kPEc&t=851s) AI 初创公司 Recursive Superintelligence 与亚马逊 AWS 签署 4.1 亿美元的多年期算力采购协议，消耗了其走出隐形模式后 6.5 亿美元融资的大部分资金。
- [20:16](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1216s) Anthropic 首席执行官 Dario Amodei 明确澄清并未主张全面封杀开源模型，呼吁聚焦对华芯片出口管制、打击工业级模型蒸馏并推行强制性政府安全审查。
- [31:17](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1877s) 马克·扎克伯格在《华尔街日报》撰文高调捍卫开源 AI，强调技术权力集中会扼杀人类潜能，开放透明与全民审查才是维护安全的最佳基石。
## Key takeaways
- [00:01](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1s) 主流大厂重启招聘标志着“AI 抢走所有人饭碗”的叙事破裂，实践表明生产效率大幅提升并拉低成本后，往往会通过“杰文斯悖论”拉动更庞大的业务总量与用工需求。
- [00:01](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1s) 过去一年多数科技管理层高调宣称“因 AI 提高效率而裁员”，本质上往往是为了掩盖疫情期间过度招聘或核心主业增长停滞的公关借口。
- [04:49](https://www.youtube.com/watch?v=QmVHay-kPEc&t=289s) Cloudflare 首席执行官马修·普林斯反对停止招聘应届生，近期激进扩招近千名新人并注入遗留业务团队，借年轻一代对 AI 工具的敏感度带动老团队转型。
- [09:20](https://www.youtube.com/watch?v=QmVHay-kPEc&t=560s) 眼镜品牌 Black Sheep 因谷歌自然搜索异常被重定向至 404 页面，导致暴增的流量全被付费广告吸收并在 30 小时内烧光 8 万美元，遂出动 25 辆 LED 卡车围堵谷歌曼哈顿总部抗议。
- [14:11](https://www.youtube.com/watch?v=QmVHay-kPEc&t=851s) 理查德·索切尔创立的 Recursive Superintelligence 斥资 4.1 亿美元直接从亚马逊 AWS 采购基础算力，呈现出 AI 领域少见的纯现金采购而非交叉持股的“循环注资”。
- [17:35](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1055s) 英伟达直接动用资产负债表承租 Hut 8 承建的 500 亿美元得州数据中心，形成自建机房、部署自研 GPU 并转租算力的自循环闭环生态，推动市值突破 5.17 万亿美元。
- [19:06](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1146s) 苹果与亚马逊未签署英伟达主导的开源公开信，主要由于两家公司在全球物流和硬件制造等物理现实中构筑了极深的实体护城河，不易被纯软件 AI 轻易侵蚀。
- [20:16](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1216s) Anthropic 强调 AI 竞争的核心在于美国绝不能输给威权政权，建议通过持续强化先进芯片与光刻设备出口管制来巩固制度领先优势。
- [22:59](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1379s) 针对工业级模型蒸馏的政策干预呼吁面临技术与法规落地难题，多模型混合训练与合法合成数据之间的边界高度模糊，难以取证与判定。
- [26:18](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1578s) 对达到特定能力门槛的模型实施强制性安全审批，极易演变为“监管俘获”：拥有巨额游说资金的大公司可通过审核排队拖垮缺乏资源的初创团队。
- [28:06](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1686s) 开源恶意模型的现实风险受制于算力门槛：高门槛模型无法绕过合规经营的本土数据中心机房，而全面封禁也可能催生类似加密货币的离岸化与 VPN 逃避现象。
- [31:17](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1877s) 扎克伯格在专栏中借开源安全历史类比 AI，主张网络安全的经验证明开源审查的韧性远胜闭源黑盒，借此在巨额资本竞赛中抢占“普惠开放”的叙事生态位。
- [33:16](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1996s) 面对旗舰设备售价逼近 2,000 美元的高门槛，苹果联合 Klarna 推出“Apple Upgrade”全品类硬件月租订阅计划，以平摊消费者的换机成本压力。
## Detailed walkthrough
### [00:01](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1s) 大厂重启招聘潮与 AI 就业叙事的惊天逆转
节目伊始，主播们提到了 Anthropic 针对开源模型监管辩论所给出的最新提案——各界围绕开源模型究竟该被禁用、限制、强制测试还是面临起诉争论不休。然而当天科技圈最大的焦点，则是《华尔街日报》头条刊登的一颗重磅“白丸”（White Pill，指令人振奋的乐观消息）：经历了一年多审慎甚至停滞的招聘期后，包括科技、交通、国防等多个行业的巨头企业正重新拉开招聘大幕。此前，不少顶尖 AI 实验室领导者曾大肆预言 AI 将引发灾难性的失业潮，但现实并未如其所料；相反，企业迎来了生产力的大幅提振，技术扩散在各行各业需要时间消化，AI 目前远无法在大多数岗位上充当同事的“即插即用型完全替代品”。这种现象正是经典的“杰文斯悖论”（Jevons Paradox）——当技术大幅提升某种生产环节的效率、压低使用成本时，市场总需求往往会迎来爆发式增长，最终反而推高了对协同操作人员的整体需求。
主播们毫不留情地讽刺了过去一年里科技圈盛行的“公关甩锅大法”。许多首席执行官和管理层在宣布大裁员时，纷纷拿 AI 当幌子，大谈“AI 提高了效率所以不再需要这么多人”，而这在很大程度上只是为了掩盖疫情期间过度招聘、或是公司核心业务陷入瓶颈的遮羞布。与其狼狈承认经营不善，不如将裁员包装成拥抱前沿科技的敏捷之举；相比之下，新任 Xbox 首席执行官敢于坦承公司经营现状的诚恳表态反而显得尤为难得。此外，许多万人大厂在发展到特定阶段后，往往会通过淘汰绩效排在后 5% 到 10% 的底层员工，转而招募更符合当前战略轨迹（例如需要更多销售或 AI 适配工程师）的新血，而这些被大厂淘汰的人换到其他组织中，很可能依然是排名前 10% 的骨干精英。
如今，这股将一切裁员归咎于 AI 的叙事正在瓦解。不仅全美每周初请失业金人数降至数十年来的低位区间，凸显出劳动力市场的强劲韧性，各大企业对 AI 能力边界的理解也趋于理性。人力资源平台 Lattice 的首席执行官莎拉·富兰克林（Sarah Franklin）指出，许多企业最初天真地认为 AI Agent 能彻底取代初级员工，但现在逐渐意识到人类员工不可替代；“拥有代码编写智能体并不意味着你不再招聘工程师”，Lattice 的众多企业客户正在重新招募包括初级岗位在内的人才。招聘巨头 Robert Half 首席执行官基斯·瓦德尔（M. Keith Waddell）也表示，AI 对就业的影响远比外界此前担忧的更为温和，企业招聘需求正在持续回升。
- **重启扩招代表企业**：铁路巨头 CSX、Alphabet、ServiceNow、工具制造巨头 Snap-on，以及咨询业巨头博思艾伦（Booz Allen Hamilton）。
- **行业心理战反弹**：此前市场盛行一种“AI 即将接管一切”的认知战（Psyop），导致不少企业因害怕招人后迅速被 AI 替代而陷入观望；如今这种防范过度摇摆的顾虑正在消退。
```mermaid
flowchart TD
accTitle: AI 引入与企业用工需求演变流程
accDescr: 展示从最初预期 AI 全面替代人力到效率提升后引发杰文斯悖论带动重新招聘的逻辑链条。
subgraph Phase1["最初悲观预期与现实"]
A["引入 AI 工具 / 编码智能体"] --> B["预期：全面替代初级岗位与工程师"]
B --> C["管理层借 AI 之名裁员 / 缩招掩盖经营问题"]
end
subgraph Phase2["实践中的杰文斯悖论"]
C --> D["生产效率提升与成本下降"]
D --> E["业务扩展与服务总需求爆发"]
E --> F["需要更多人类员工配合 AI 协同作业"]
end
Phase1 --> Phase2
```
---
### [04:49](https://www.youtube.com/watch?v=QmVHay-kPEc&t=289s) 从外包零工到凯尔特朋克乐队：团队使用 AI 的现实体感
谈及小型团队的真实切身体会，主播们以自身制作每日科技直播节目的组织为例展开探讨。对于此类极度细分的垂直内容团队而言，AI 真正取代的并非核心骨干，而是过去经常在 Upwork 等外包平台上雇佣的非专家级自由职业者——比如临时写一段恶搞歌曲、或快速搭建一个恶搞网站。过去找人做一首恶作剧歌曲需要花钱找网络乐手，而现在打开 Suno 几秒钟就能生成出来。
话题随即拐进了一段充满笑料的无厘头插曲：主播调侃道，既然如此，自己是不是该把招聘启事上预留的“5 个全职凯尔特朋克（Celtic Punk）录音乐手”岗位撤下来，因为他原本计划天天让他们翻唱 Dropkick Murphys 的硬核朋克曲目充当节目背景乐。两人随即大笑着反讽：“不，先别撤，我现在比以往任何时候都更想雇一支全职凯尔特朋克乐队来现场重制音乐！”玩笑归玩笑，这恰恰反映出 AI 在真实工作流中的定位：它抹平了边缘琐碎任务的门槛，充当填补业务缝隙的高效工具，但组织的核心中枢依然必须由承担最终责任的真人牢牢把控，并由他们熟练操控 AI 放大产出。
关于初级毕业生的前途，风险投资人布莱斯·罗伯茨（Bryce Roberts）在社交媒体上表达了相对悲观的看法。他贴出了一张同行聊天截图，对方声称“我们目前基本不招人了，所有岗位空缺基本全靠 AI 填补（backfilling）”，并感叹“应届生要完了（Rip new grads）”。主播分析指出，这种所谓靠 AI 填补离职空缺（比如两人小组离职一人，便要求留下的人借助 AI 承担双倍工作）的企业，大概率自身业务并未处于高速扩张期；当一家公司的业务真正迎来爆发式飞跃时，管理层唯一的痛点只会是优秀人才招得根本不够快。与此同时，Cloudflare 首席执行官马修·普林斯（Matthew Prince）则针锋相对地给出了截然相反的人才策略。
- **马修·普林斯的招聘理念**：停止招募应届生是极其愚蠢的策略；正确的做法是大规模招募年轻人，将他们直接插入到传统的遗留业务团队中，借助新生代对前沿工具的天然敏感度，加速推动老牌团队全面拥抱 AI。
- **Cloudflare 的激进扩招**：普林斯掌舵的 Cloudflare 近期一举新招募了近千名新员工，展现出对人才扩张的极强进攻性。
---
### [07:54](https://www.youtube.com/watch?v=QmVHay-kPEc&t=474s) 商业反派球星卡：“Good Work” 的职场恶搞狂欢
节目的轻松时刻落在了职场恶搞媒体 Good Work 最新发布的短视频上。主播们在镜头前津津有味地围观拆封一套极具讽刺意味的“商业反派球星卡”（Business Villain Trading Cards）。视频中，拆卡者煞有介事地戴着专业手套，像鉴定珍贵棒球卡一样拆解商界臭名昭著的风云人物。
第一张拆出的是庞氏骗局主谋伯尼·麦道夫（Bernie Madoff），带有浓烈的 80 年代复古做旧质感；紧接着登场的是 FTX 创始人山姆·班克曼-弗里德（Sam Bankman-Fried, SBF），卡片竟然还是用廉价的轻木（Balsa wood）材质精巧制作，无情嘲讽其草台班子属性；随后拆出的则是一张 2005 年元老级复古马克·扎克伯格（Mark Zuckerberg）卡，背面甚至戏谑地镶嵌着一片号称来自扎克伯格 Fruit of the Loom 平角内裤的布料实物切片，拆卡演员还一本正经地凑上去闻了闻，声称“靠气味就能验明正身”，精准还原了体育球星卡内嵌球衣切片的招牌手法。
藏品的高潮接踵而至，接下来是血液透析骗局 Theranos 创始人伊丽莎白·霍姆斯（Elizabeth Holmes）的卡片，演员调侃称资深藏家必须人手集齐“三张霍姆斯卡”才算排面。而当最后一张神秘王牌在“三、二、一”倒计时中被缓缓揭开时，全场瞬间爆发出一阵失控的惊呼、狂笑与“天哪赶紧关掉”的尖叫。主播们感叹，这种独属于硅谷与华尔街的“商业喜剧正典”（Business Comedy Canon），如今已经演化得越来越成熟、越来越具文化杀伤力。
### [09:20](https://www.youtube.com/watch?v=QmVHay-kPEc&t=560s) 商业反派“球星卡”与 Black Sheep 开卡车围堵谷歌曼哈顿总部
在调侃完恶搞视频制作团队 Good Work 的商业反派“球星卡”——盘点了包括伊丽莎白·霍姆斯（Elizabeth Holmes）、SBF（Sam Bankman-Fried）和马克·扎克伯格（Mark Zuckerberg）等科技争议人物，并笑称居然没有包含 WeWork 创始人亚当·诺依曼（Adam Neumann）的“新秀卡”后，主持人们将目光转向了一起清晨发生在曼哈顿的现实抓马事件。现场发来的照片显示，一家名为 Black Sheep 的眼镜品牌雇佣了整整 25 辆装有 LED 大屏的卡车，绕着谷歌位于纽约切尔西（Chelsea）的总部大楼打转，屏幕上赫然打出标语：“谷歌，为你感到羞耻。还我们 8 万美元！”（"Shame on you, Google. Return our $80,000."）。
这起戏剧性的地推抗议源于一场广告投放事故。Black Sheep 自称是一家“直连工厂”（Direct-to-factory）的光学颠覆者，其主打的 8 美元平价墨镜曾在一档 NBC 实验室评测中击败了标价 350 美元的奢侈大牌。然而，就在其登上 NBC《Today Show》一炮而红后，全美搜索流量出现井喷，灾难却随之降临：该品牌声称谷歌的自然搜索结果出现致命 Bug，将本应导向其官网的自然流量重定向到了一个第三方 404 错误页面。大量涌入的用户因找不到自然入口，全被引导点击了顶部的谷歌付费广告链接，导致 Black Sheep 在短短 30 小时内被迅速烧光了近 7.7 万至 8 万美元的广告预算。
主持人们围绕这场闹剧展开了热烈讨论。一方面，他们质疑这可能只是 Black Sheep 团队自身的“用户操作失误”，因为网站管理员工具原本可以自主设置抓取索引或快速配置 301 重定向；但另一方面，大家也一致认为这场公关特技相当高明——25 辆黑白极简风格的 LED 卡车穿行在曼哈顿街头，吸引了巨大的社交热度，为其赚回的有机曝光价值早已远远超过了损失的 7.7 万美元。
```mermaid
flowchart TD
accTitle: Black Sheep 广告资金消耗事件流程
accDescr: 展示从 NBC 节目走红到谷歌搜索流量异常导致广告费被快速消耗的整个链条
subgraph Exposure ["曝光爆发"]
A["NBC Today Show 播出"] --> B["全美搜索流量暴增"]
end
subgraph SearchIssue ["谷歌搜索异常"]
B --> C["自然搜索排名失效"]
C --> D["流量被导向第三方 404 页面"]
end
subgraph AdDrain ["付费流量溢出"]
D --> E["用户被迫点击搜索广告"]
E --> F["30 小时内消耗近 8 万美元广告费"]
F --> G["出动 25 辆 LED 卡车围堵谷歌总部抗议"]
end
```
- **抗议规模**：共计 25 辆装配 LED 广告屏的卡车包围谷歌曼哈顿切尔西园区。
- **争议焦点**：在 NBC 电视曝光后流量暴增的 30 小时内，自然搜索失效被转入 404 错误页，导致系统短时间内消耗掉 7.7 万至 8 万美元付费搜索广告费用。
- **产品对照**：Black Sheep 主打 8 美元平价镜框，宣称在 NBC 独立测试中性能战胜了 350 美元的传统零售产品。
---
### [14:11](https://www.youtube.com/watch?v=QmVHay-kPEc&t=851s) SSI 的神秘进展与 Recursive 豪掷 4.1 亿美元采购亚马逊算力
话题随后转向前沿 AI 领域的资本与研发动态。针对伊利亚·苏茨克维（Ilya Sutskever）创立的 Safe Superintelligence（SSI），坊间充斥着各种揣测。有推文调侃称伊利亚曾说要“一路直奔 SSI”，希望他们别最后像其他团队一样，只端出一个叫“France”或“Francois”的日常打工 Agent。主持人们指出，SSI 所谓的大动作实际上很可能只是研发规模化（scaling research）——他们在模型架构或飞轮机制上取得了一定进展，因此需要筹集更多资金去购买算力堆叠训练，绝非外界传言的即将公开发布某种革命性产品。他们戏称前沿 AI 创始人如今都在过一种“谜语人生活”（vague life），发布的内容处处充斥着故弄玄虚的不可捉摸感。
紧接着，另一家名字极具戏剧张力的 AI 初创公司浮出水面——理查德·索切尔（Richard Socher）创立的 Recursive Super Intelligence（递归超级智能）。主持人们大笑着指出科技媒体 TechCrunch 在标题中犯下的严重笔误，把金额写成了令人咋舌的“410 million million”（410万亿美元）。实际上，这家今年 5 月刚带着 6.5 亿美元融资走出隐形模式（Stealth Mode）的公司，与亚马逊 AWS 签署了一份价值 4.1 亿美元的多年度大规模算力采购协议。
这笔重磅交易几乎消耗了 Recursive 迄今为止募集的大部分资金，不过其首席执行官兼创始人理查德·索切尔对媒体表示，这预计只是未来众多算力大单的开端。主持人们展开了一段关于现代 AI 领域“闭环交易/循环注资”（circular deals）的讽刺对话：以往的风投模式中，科技巨头往往会先向初创公司投资，然后初创公司再把钱拿去租赁投资方的云服务或自研芯片（例如 Trainium），双方在账面上互相抬轿；而 Recursive 与 AWS 这笔交易在表面上显得异常“传统而纯粹”——就是拿现金直接购买大量云算力，几乎没有复杂的交叉持股痕迹。不过，AWS 初创企业及风险投资副总裁 Jason 提及双方将共同针对该类架构定制算力，让主持人们笑称“或许未来还是有机会看到经典的循环注资玩法”。
- **融资与开销**：Recursive 于 2026 年 5 月走出隐形模式并官宣 6.5 亿美元融资，此次 AWS 算力合同总值达 4.1 亿美元，占据其总融资金额的大半。
- **技术路线**：专注于开放式自迭代（open-ended self-improving）系统与重度依赖算力支撑的 AI 基础研发。
---
### [17:35](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1055s) 英伟达 500 亿美元数据中心大单与万亿美元市值的较量
如果说 Recursive 的交易还不够“循环”，那么英伟达最新曝光的运作则是教科书级别的自建生态闭环。据最新披露，英伟达作为主要承租方，签署了总价值高达 500 亿美元的超大规模得克萨斯州数据中心租赁协议。该超级数据中心由 Hut 8 负责开发建设，总电力容量高达惊人的 1 吉瓦（1 Gigawatt），内部将部署成千上万颗英伟达自家的 GPU 图形处理器芯片。
这意味着近 5 万亿美元市值的英伟达正直接动用自己的资产负债表，充当买家与承租人来为整个 AI 算力市场托底——英伟达不仅向市场卖铲子，甚至亲自出资租赁装满自家铲子的大型机房，未来再将其算力租售给第三方开发者或大客户。
消息传出当天，二级市场反应极其强烈：
- **英伟达（NVIDIA）**：盘中拉出一根巨大的大阳线，股价单日大涨 3%，市值一举跃升至惊人的 5.17 万亿美元。
- **苹果（Apple）**：盘中市值曾一度突破 5 万亿美元大关，随后略有回落，徘徊在 4.99 万亿美元左右，与英伟达展开了胶着的全球最高市值争夺战。
面对两家科技霸主的巅峰对决，主持人们感叹其策略的截然不同。黄仁勋（Jensen Huang）宛如一位疯狂的操盘手，不知疲倦地在全球穿梭、敲定数以千计的合作协议与巨额基础设施租赁；而苹果则在贯彻一种“什么都不干就能赢”（do nothing win strategy）的极简哲学，甚至连各大科技公司争相表态的 AI 行业公开信都未曾参与，两者的行事风格形成了极具戏剧感的鲜明对比。
### [19:06](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1146s) 未签署公开信的三巨头：Apple、Amazon 与 Anthropic 的考量
在围绕英伟达（Nvidia）主导的“反开源禁令”联名公开信风波中，整个科技行业几乎倾巢而出表示支持，但全球唯独有三家重量级巨头迟迟没有签字：苹果（Apple）、亚马逊（Amazon）与 Anthropic。针对这一缺席，Anthropic 迅速发布了一篇立场极为清晰的长文，正面回应其对开源模型的态度与主张。至于苹果和亚马逊，两位主持人调侃这两家公司在实体世界中都有着极深的护城河，属于“无法被粗制滥造击垮”（unsloppable）的业务形态——你不可能通过“氛围编码”（vibe coding）凭空写出一座亚马逊仓储物流中心，也无法靠提示词捏出一台 iPhone。
尽管亚马逊能通过 AWS 托管并分发开源模型获利，苹果也能坐享端侧开源生态的好处，但苹果选择按兵不动，更像是一种品牌层面的特立独行，维持其“非同凡想”（Think Different）的格调，拒绝随大流加入这场硅谷公开信狂欢。主持人打趣道：“按照苹果一贯的 AI 节奏，我估计他们得在一两年后才会考虑签这份信。”
- **未签署联名信的三大巨头**：Apple、Amazon、Anthropic。
- **实体业务护城河**：亚马逊的全球物流网络与苹果的硬件制造体系无法轻易被纯软件层面的 AI 冲击取代。
---
### [20:16](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1216s) Dario Amodei 的澄清与三大主张：地缘竞争、芯片封锁与安全风险
针对周末发酵的开源权重联名信，Anthropic 首席执行官达里奥·阿莫代伊（Dario Amodei）迅速做出了官方回应。在正式提出政策倡议之前，阿莫代伊首先摆明了 Anthropic 的三个核心前提：第一，Anthropic 从未主张对开源权重模型实施一刀切的全面禁令（blanket ban）；第二，AI 竞争的本质是美国必须跑赢威权政府，阿莫代伊明确指出，如果以中国为代表的威权政权掌握了超级 AI，美国的技术与制度优势将荡然无存；第三，顶尖的高性能 AI 模型一旦缺乏管控，极易被滥用于网络攻击或生物武器制造，放任极度强大的开源系统野蛮生长存在巨大的系统性风险。
基于这三大前提出发，阿莫代伊提出了第一项关键政策建议：**持续强化芯片管制与出口制裁，坚决禁止向中国出售最先进的 AI 芯片及光刻/制造设备**。主持人分析指出，自拜登政府出台芯片限令以来，这一直是地缘经济博弈的核心议题。即便不考虑生存风险，仅从经济竞争角度看，芯片也是现代经济的引擎，拖慢竞争对手的脚步对美国而言是显而易见的利益。有人担忧美国芯片企业会因此丢掉中国这个全球第二大计算市场，但反方论点在于，中国数十年来一直在举国推进芯片自研与本土半导体代工生态，即使没有禁令，这个市场终究也会转向本土替代；西方只要能维持几年的代际领先差距，就是战略优势。
```mermaid
flowchart TD
accTitle: Anthropic关于模型与芯片管制的政策主张
accDescr: 展示Dario Amodei提出的三项核心前提与应对政策路径
subgraph Claims["三大核心澄清"]
C1["从未主张全面封杀开源权重"]
C2["地缘竞赛：美必须跑赢威权国家"]
C3["防范生化与高阶网络安全风险"]
end
subgraph Proposals["三大行动倡议"]
P1["深化芯片与光刻设备出口管制"]
P2["打击工业级模型蒸馏侵权"]
P3["对超强模型推行强制安全审查"]
end
Claims --> Proposals
```
---
### [22:59](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1379s) 工业级模型蒸馏与政策干预的灰色地带
阿莫代伊提出的第二项倡议，是**在政策层面严厉打击“工业级模型蒸馏”（industrial-scale distillation）行为**。从商业常理来看，企业有权设立服务条款（ToS）以保护其专有知识产权，任何违规调用 API 提取数据训练自有模型的行为都应面临法律制裁。Anthropic 一直在与针对其前沿模型的蒸馏攻击作斗争，但单纯依靠技术防御往往防不胜防，因此阿莫代伊呼吁政府出台政策干预。
然而，主持人和现场分析师对“政策干预”的具体落地表达了极大的困惑与怀疑：
1. **政策手段模糊**：所谓的政策干预究竟是关税、罚款、谴责信，还是剥夺国际准入？
2. **缺乏技术判定标准**：根本没有一套标准算法可以断言一个模型是 100% 蒸馏产物。开发者完全可以将半成品的 Claude Opus 蒸馏数据，掺入 GPT-5.6 的生成样本、Mistral 的开源数据，再结合独特的强化学习（RL）微调，这种混合体极难在法律和技术上被精确界定。
3. **合成数据的灰色地带**：Tyler 指出，像 Thing Machines 的 Inkling 模型在训练时就使用了由月之暗面 Kimi K2.5 生成的合成数据。这究竟算不算蒸馏？虽然它不属于“工业级规模”的盗用，但它确实吸收了其他前沿模型的成果。
华盛顿内部各方势力针对此事的博弈已逐渐浮出水面。英伟达与开源支持派阵营，与 Anthropic 等闭源防御阵营展开了拉锯。若要进行政策打击，政府不仅需要厂商提供极高门槛的 API 滥用证据日志与定量基准评测报告，而且面对来自海外管辖权之外的实验室，传统民事诉讼也几乎无从下手。
---
### [26:18](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1578s) 强制性安全测试的博弈与监管陷阱
阿莫代伊提出的第三项倡议是：**所有达到特定能力门槛的模型，无论开源还是闭源，都必须接受强制性的政府安全测试**。巧合的是，Google DeepMind 首席执行官德米斯·哈萨比斯（Demis Hassabis）近期也公开发表了几乎相同的观点，显示出两大顶尖闭源实验室在监管议程上的高度一致。
但这恰恰引发了行业对“监管俘渴”（Regulatory Capture）的极度担忧。这一流程极可能演变成类似美国食品药品监督管理局（FDA）或核能监管委员会（NRC）式的审批泥潭：
- **排队与资源碾压**：像 Recursive Superintelligence 这样的新兴初创企业，一旦没有华盛顿的说客网络，其安全模型可能要在漫长的审批队列中苦等数年；而坐拥数万亿美元市值的巨头则可以派出庞大的游说团队优先插队审核，从而将竞争对手扼杀在摇篮里。
- **打击创新生态**：尽管阿莫代伊在文中极力声明其初衷绝非为了“保护美国现存 AI 巨头免受市场竞争”，但这种牌照式准入制度客观上必然导致小团队的创新速度被严重拖垮。
在推演极端对抗场景时，主持人提出一个无解的困境：假若某家不受西方约束的海外实验室，通过走私获取的高性能算力，对西方前沿模型实施最为激进的逆向蒸馏，随后人为抹除模型的所有生化与网络武器护栏，直接将未对齐的模型权重上传至 BitTorrent 磁力链或 Hugging Face 上。面对这种完全失控且不可逆的开源权重外泄，任何现有的政府安全测试与国内监管框架，都将显得苍白无力。
### [28:06](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1686s) 危险模型的算力门槛与监管博弈：离岸化与针对性防御
主持人深入探讨了开源权重与危险模型监管的现实可行性。当前公众对开源危险性的看法存在分歧，主要因为现存模型并未展现出切实的致命威胁；但如果未来真出现某种纯粹的“杀手机器（killing machine）”，绝大多数人都会出于民主意愿投票支持对其进行封禁与断供。然而核心技术问题在于：一个模型要达到真正具备现实危害的程度，究竟需要多少算力？是个人在自家后院小屋里摆几张 GPU 就能驱动，还是必须依赖成排的高密服务器机架与海量电力？如果需要极其庞大的算力与能源，使用者就不得不接入 Neocloud（新型云服务商）。而任何在美国境内运营、拥有配备 NVL72 机架数据中心的企业，都有正规的工商登记注册，政府随时可以依据法律介入并提出安全警告——这与现行法规禁止数据中心托管盗版侵权内容、或禁止 CRM 软件公司明知故犯地协助跨国贩毒集团洗钱和运营是同一个道理。
但如果恶性模型已经开源流出，接下来的博弈将进入类似加密货币的灰色地带：模型是否会被迫转移至海外？用户是否会像过去规避监管那样通过 VPN 翻墙调用离岸算力？对此，政府的干预手段大概率会与其潜在威胁成正比。若该模型只是略微侵犯版权（比如能自动生成迪士尼 IP，但用户根本不会因此退订 Disney+），监管机构绝不会大动干戈建立国家级防火墙；可一旦它演变为能够席卷全球银行账户的“终极黑客机器（ultimate hack machine）”，监管与防御措施必将极其激进。主持人指出，Anthropic 近期发表的公开信虽然厘清了其在开放权重上的立场，但在具体政策干预手段上依然过于笼统，未来各方围绕模型能力的监管博弈仍需更清晰的推演与制度设计。
```mermaid
flowchart TD
accTitle: 危险 AI 模型监管与应对博弈逻辑
accDescr: 展示模型危害等级、算力依赖程度与政府相应监管反应的博弈路径。
subgraph Compute["算力需求分流"]
A["潜在危险开源模型"] --> B{"算力需求门槛"}
B -- "低算力 (后院散片 GPU)" --> C["去中心化运行 / 难以物理根除"]
B -- "高算力 (NVL72 机架 / 巨额电力)" --> D["必须依赖本土 Neocloud 数据中心"]
end
subgraph Regulation["监管干预手段"]
D --> E["合规审查与企业注册监管介入"]
C --> F{"危害程度评估"}
E --> F
F -- "轻度侵权 / 扰民骚扰" --> G["常规法律维权 (比例原则，不建防火墙)"]
F -- "系统性灾难 (如清空银行的终极黑客)" --> H["国家级防火墙封锁 + 强力打击离岸算力"]
end
```
- **数据中心合规先例**：数据中心无法规避知识产权法律，正如 CRM 厂商在法律层面绝不能蓄意为跨国贩毒集团提供业务支撑。
- **离岸转移假说**：一旦美国本土全面禁止托管特定恶意权重，AI 运行极可能重演加密货币（Crypto）的离岸化路径，催生海外节点与 VPN 访问生态。
- **政策博弈现状**：Anthropic 的信件虽然表明了态度，但关于具体的“政策干预（policy interventions）”细节仍停留在泛泛而谈的阶段，有待后续立法细化。
---
### [31:17](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1877s) 扎克伯格发文力挺开源 AI：“AI 的未来属于所有人”
马克·扎克伯格在《华尔街日报》（The Wall Street Journal）观点版发表了题为《AI 的未来属于所有人》（The AI Future is for Everyone）的专栏文章。扎克伯格在文中鲜明指出：“民主与经济的发展历史证明，权力的过度集中会扼杀人类的潜能（The history of democracy and economics has proved that centralized power stifles human potential）。”面对当前科技巨头争相在 AI 基础设施上掷入千亿美元的军备竞赛，扎克伯格敏锐地抢占了一块极为关键的叙事生态位：将自己塑造成那个“每年投入数千亿美元搞 AI、但坚信这项技术应当普惠每一个人并致力于推动开放生态”的代表人物。
针对外界对 Meta 是否存在垄断的批评，主持人展开了辩护与讨论，认为 Facebook 虽然在部分领域具有主导地位，但在注意力竞争层面始终面临着极其激烈的外部挑战——从 TikTok、Snapchat、LinkedIn、Twitch 到 YouTube、Netflix、各类播客平台甚至 SMS 和 iMessage，信息分发渠道从未真正被一家垄断。安德鲁·柯伦（Andrew Curran）摘录了扎克伯格文章中最核心的论点：正如网络安全领域所展示的那样，开源软件的历史已经反复证明，“赋予每个人使用强大系统的完整权限，才是长期维护安全与保障的最佳方式（giving everyone full access to powerful systems will be the best way to protect safety and security over time）”。扎克伯格坚定地站在了技术民主化与强力开源模型的大旗之下。
- **专栏核心论点**：集中式控制不仅违背历史规律，而且闭源并不等同于安全；网络安全的经验表明，全社会共同审查与使用的开源系统最终具备最强的防御韧性。
- **注意力经济的竞争格局**：主持人列举了包括 TikTok、Snapchat、Twitch、YouTube、Netflix、iMessage 在内的多维注意力渠道，反驳对 Meta 拥有绝对信息垄断的简单定性。
---
### [33:16](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1996s) 苹果推出硬件订阅租赁服务与 TBPN 锣声通胀趣谈
苹果官方正式宣布将于下周推出名为“Apple Upgrade”的全品类硬件订阅租赁计划，联合先买后付巨头 Klarna 在美国线上及线下零售店全面上线，覆盖 iPhone、iPad、Mac 和 Apple Watch。主持人打趣地引用了世界经济论坛的名言：“你将一无所有，并感到幸福（You will own nothing and be happy）。”根据公布的官方资费标准，设备月租门槛极低：
- **iPhone**：每月起售价为 17.99 美元（或最低 20 美元档位）；
- **Apple Watch**：每月 11.99 美元；
- **iPad**：每月 11.99 美元；
- **Mac**：每月 24.99 美元。
节目分析指出，该计划是苹果应对内存成本上涨与整机价格攀升的直接反应。过去一部 iPhone 仅需几百美元，消费者还可以借助 Verizon 等电信运营商的补贴套餐来平摊硬件成本；而如今顶配 iPhone 售价已直逼 2,000 美元，高昂的换机门槛使得租赁订阅模式成为必然趋势。
话题随即幽默地拐向了 TBPN 演播室标志性铜锣的“通货膨胀”。主持人调侃道，曾经演播室里敲的小锣只要 200 美元，如今定制的大锣价格已经飙升至数万美元。面对朋友索要买锣渠道的询问，主持人建议“必须从小锣买起”——这并非看不起对方的承受能力，而是因为在“享乐适应跑步机（hedonic treadmill）”上一步步升级铜锣才具有真正的快感；每次引入一面尺寸更大的新锣，演播室的气氛都无比炸裂。另一位主持人当即表示演播室需要再添一面直接吊在屋顶椽木上的超巨型铜锣。节目在两人的插科打诨中收尾，主持人提醒听众在 Apple Podcasts 和 Spotify 上打五星好评，并念出招牌告别语：“金钱永不眠，你也不该睡（Money never sleeps. You shouldn't either. Call me）”，相约次日上午 11 点直播再见。
## Notable quotes
> [00:01](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1s) 拥有代码编写智能体并不意味着你不再招聘工程师 — Sarah Franklin
> [04:49](https://www.youtube.com/watch?v=QmVHay-kPEc&t=289s) 停止招募应届生是极其愚蠢的策略；正确的做法是大规模招募年轻人，将他们直接插入到传统的遗留业务团队中，借助新生代对前沿工具的天然敏感度，加速推动老牌团队全面拥抱 AI — Matthew Prince
> [04:49](https://www.youtube.com/watch?v=QmVHay-kPEc&t=289s) 我们目前基本不招人了，所有岗位空缺基本全靠 AI 填补，应届生要完了 — 同行聊天截图（Bryce Roberts 引用）
> [09:20](https://www.youtube.com/watch?v=QmVHay-kPEc&t=560s) 谷歌，为你感到羞耻。还我们 8 万美元！ — Black Sheep LED 卡车站牌标语
> [19:06](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1146s) 你不可能通过“氛围编码”凭空写出一座亚马逊仓储物流中心，也无法靠提示词捏出一台 iPhone — TBPN 主播
> [31:17](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1877s) 民主与经济的发展历史证明，权力的过度集中会扼杀人类的潜能 — Mark Zuckerberg
> [31:17](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1877s) 赋予每个人使用强大系统的完整权限，才是长期维护安全与保障的最佳方式 — Mark Zuckerberg
> [33:16](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1996s) 你将一无所有，并感到幸福 — 世界经济论坛（主播引用）
> [33:16](https://www.youtube.com/watch?v=QmVHay-kPEc&t=1996s) 金钱永不眠，你也不该睡 — TBPN 主播
## Mentioned
Anthropic, Dario Amodei, Mark Zuckerberg, Meta, Google, Alphabet, Apple, Nvidia, Jensen Huang, Amazon, AWS, Ilya Sutskever, Safe Superintelligence (SSI), Richard Socher, Recursive Super Intelligence, Matthew Prince, Cloudflare, Sarah Franklin, Lattice, M. Keith Waddell, Robert Half, Bryce Roberts, Demis Hassabis, Google DeepMind, Black Sheep, Good Work, Hut 8, Klarna, CSX, ServiceNow, Snap-on, Booz Allen Hamilton, The Wall Street Journal, Today Show, NBC, Claude Opus, Mistral, Thing Machines, Inkling, 月之暗面 Kimi K2.5, Bernie Madoff, Sam Bankman-Fried, Elizabeth Holmes, Adam Neumann
## Video description
Sign up for TBPN’s daily newsletter at TBPN.com
Diet TBPN delivers the best of today’s TBPN episode in 30 minutes. TBPN is a live tech talk show hosted by John Coogan and Jordi Hays, streaming weekdays 11–2 PT on X and YouTube, with each episode posted to podcast platforms right after.
Described by The New York Times as “Silicon Valley’s newest obsession,” the show has recently featured Mark Zuckerberg, Sam Altman, Mark Cuban, and Satya Nadella.
TBPN is made possible by:
Ramp - https://ramp.com
Public - https://public.com
Cisco - https://www.cisco.com
Console - https://www.console.com
CrowdStrike - https://www.crowdstrike.com
Figma - https://www.figma.com
MongoDB - https://www.mongodb.com
NYSE - https://www.nyse.com
Railway - https://railway.com
Shopify - https://www.shopify.com/
Follow TBPN:
https://TBPN.com
https://x.com/tbpn
https://open.spotify.com/show/2L6WMqY3GUPCGBD0dX6p00?si=674252d53acf4231
https://podcasts.apple.com/us/podcast/technology-brothers/id1772360235
https://www.youtube.com/@TBPNLive
Chapters:
00:00 Big Companies Start Hiring Again + The AI Jobs Narrative Flips
07:58 Business Villain Trading Cards + Black Sheep vs. Google
14:10 SSI Speculation, Recursive Superintelligence + Nvidia’s $50B Data Center
20:16 Anthropic’s Open-Weight Position + China Chip Controls
22:59 Industrial-Scale Distillation + The Synthetic Data Gray Area
26:18 Mandatory Safety Testing + Can Dangerous Open Models Be Contained?
31:17 Zuckerberg’s “AI Future for Everyone” + Apple Device Leasing
[Source](https://www.youtube.com/watch?v=QmVHay-kPEc)
Source quotes
[0:01] Sarah Franklin: 拥有代码编写智能体并不意味着你不再招聘工程师
[4:49] Matthew Prince: 停止招募应届生是极其愚蠢的策略；正确的做法是大规模招募年轻人，将他们直接插入到传统的遗留业务团队中，借助新生代对前沿工具的天然敏感度，加速推动老牌团队全面拥抱 AI
[4:49] 同行聊天截图（Bryce Roberts 引用）: 我们目前基本不招人了，所有岗位空缺基本全靠 AI 填补，应届生要完了
[9:20] Black Sheep LED 卡车站牌标语: 谷歌，为你感到羞耻。还我们 8 万美元！
[19:06] TBPN 主播: 你不可能通过“氛围编码”凭空写出一座亚马逊仓储物流中心，也无法靠提示词捏出一台 iPhone
[31:17] Mark Zuckerberg: 民主与经济的发展历史证明，权力的过度集中会扼杀人类的潜能
[31:17] Mark Zuckerberg: 赋予每个人使用强大系统的完整权限，才是长期维护安全与保障的最佳方式
[33:16] 世界经济论坛（主播引用）: 你将一无所有，并感到幸福
[33:16] TBPN 主播: 金钱永不眠，你也不该睡
Source quotes
[0:48] I'm actually closer than ever to hiring a full-time Celtic punk band to play music to recreate songs... where that was not even on the roadmap a few years ago.
[0:48] Bryce Roberts: RIP new grads.
[7:58] Shame on you, Google. Return our $80,000.
[0:01] M. Keith Waddell: more benign than some had feared
[31:17] Mark Zuckerberg: The AI Future Is for Everyone
[31:17] Mark Zuckerberg: the history of democracy and economics has proved that centralized power stifles human potential.
Original transcript
[0:01] Well, uh, Anthropics responded. We're
[0:03] going to go through that proposal, the
[0:06] the the facts and the proposal
[0:07] [clears throat] for what happens next in
[0:09] the open, uh, open model debate over
[0:12] whether or not they should be banned,
[0:14] restricted, tested, limited in some
[0:16] ways, sued. There's a whole bunch of
[0:18] different possible outcomes. Uh, but
[0:21] we'll take you all through it. The Wall
[0:22] Street Journal is reporting that large
[0:23] companies are beginning to
[0:24] >> They have a large white pill. Yes, it is
[0:26] a large white pill
[0:27] >> has hit the front page of the journal.
[0:29] >> Yes. And I think people have been going
[0:31] back and forth on this. This is a story
[0:33] that's that's just getting digested by
[0:36] the tech folks like the actual AI lab
[0:39] leaders who had predicted crazy job
[0:42] losses and are now not really seeing
[0:44] that. They're seeing productivity boosts
[0:46] and different uh diffusion taking time
[0:50] in certain places and there's new
[0:52] capabilities, but it's not exactly a
[0:54] drop-in replacement for a co-worker, at
[0:56] least in in most scenarios. And that's
[0:59] what the Wall Street Journal is
[1:00] reporting. So, let me set the table and
[1:02] then uh we can debate it a little bit.
[1:04] After roughly a year of cautious hiring,
[1:07] companies across technology,
[1:08] transportation, defense, and other
[1:11] industries now say they need more
[1:13] employees to work alongside AI systems.
[1:16] Total victory for both humans and AI.
[1:19] We're working together. Peace is
[1:21] possible. It's an example of Je of
[1:22] Jevans paradox. Jeban's paradox. When a
[1:25] technology makes something more
[1:26] efficient, demand often rises enough.
[1:28] The total use and the need for people
[1:31] actually increases. For roughly the past
[1:33] year, many companies pointed to AI while
[1:35] announcing layoffs. This was a huge
[1:38] thorn in your side. You I think you
[1:39] hated this more than anyone else. Um and
[1:42] you were right to because it did seem
[1:44] like it was just PR spin, etc. Yeah, it
[1:47] was it was a way for CEOs and management
[1:50] teams to save their own ass instead of
[1:53] saying, you know, hey, we we overhired
[1:55] or the business isn't doing as well as
[1:57] as as we would like and we need to sort
[1:59] of
[2:00] >> uh basically [clears throat]
[2:01] settle down for a second and uh get our
[2:04] mojo back. Yeah, obviously no one wants
[2:06] to say that, but I I think uh one of my
[2:10] favorite posts was the and and obviously
[2:12] these circumstances are never great, but
[2:14] the the new CEO of Xbox came out and
[2:16] just was like very honest about the
[2:18] situation.
[2:18] >> Yep.
[2:19] >> And I think that uh more of that is
[2:22] necessary.
[2:23] >> Yeah. Also, there's a lot of firms where
[2:25] they once they get to 10,000 20,000
[2:29] employees, they might say, "Look, 20,000
[2:31] might be the right number, but the
[2:33] bottom thousand people are not
[2:36] performing. We would like to lay them
[2:38] off and then bring in a new thousand
[2:40] people that are better fit for the
[2:41] company and the current trajectory that
[2:43] we're on, the current skills that we
[2:44] need, maybe we need more sales people."
[2:46] >> And those bottom people might be top 10%
[2:50] at another company. Exactly. Yeah. So,
[2:53] uh, the narrative appears to be
[2:54] shifting. Uh, companies like CSX,
[2:56] Alphabet, Service Now, Snap-on, and
[2:58] consulting giant Booze Allen, Hamilton
[3:00] have all recently signaled plans to
[3:03] expand hiring, particularly in areas
[3:05] where employees can use AI to become
[3:08] more productive.
[3:09] >> Yeah. It used to be somewhat of a flex
[3:10] if a company was like, "Yeah, we put up
[3:12] a role and we got 2,000 applicants."
[3:14] Yeah. You know, it's like, well,
[3:16] >> that's true. And then also, it's a
[3:18] little bit of a sign of like, "Oh, wow.
[3:19] They have a hundred openings. like they
[3:21] must be like growing so fast, you know?
[3:24] So, but if it's just a prompt to say,
[3:25] "Oh, yeah, put up like look at my
[3:27] organizational design and and add five
[3:30] roles for everyone because why not? Why
[3:33] not see who comes by?" You know, we're
[3:35] not we don't necessarily have to
[3:36] interview these people. Uh so, weird
[3:38] weird dynamics, but we'll dig into it.
[3:40] So, uh meanwhile, the latest weekly US
[3:42] jobless claims fell to one of the lowest
[3:44] levels in decades, underscoring the
[3:46] resilience of the labor market. The
[3:48] shift also requ reflects a more
[3:50] realistic understanding of AI's
[3:52] capabilities. Sarah Franklin, CEO of HR
[3:54] platform Lattis, says many companies
[3:56] initially assumed AI agents could
[3:58] replace entry-level workers, but are now
[4:00] recognizing that human human employees
[4:02] remain essential. Just because you have
[4:04] coding agents doesn't mean you're not
[4:06] hiring engineers, she said, adding that
[4:08] Lattis is seeing renewed hiring among
[4:11] many of its customers, including for
[4:12] junior roles. Uh Robert Half, CEO M.
[4:16] Keith Watt said, "AI's effect on
[4:18] employment has been more benign than
[4:20] some has feared, adding that hiring
[4:22] demand continues to improve and market
[4:24] conditions are increasingly more
[4:25] supportive of business." And so I do
[4:27] think there was a little bit of like a
[4:28] successful scop with the with the AI is
[4:32] going to be able to do everything where
[4:33] I do think there are some firms that
[4:34] were like, "Yeah, maybe we shouldn't
[4:35] hire or because like what if we get it
[4:37] wrong and we hire a bunch of people and
[4:38] then AI really does catch up and we
[4:41] don't need those people. That's silly.
[4:42] We shouldn't go through that like
[4:44] whipssaw effect." Uh, and so people are
[4:46] going back and forth on that. Bryce
[4:48] Roberts.
[4:49] >> Yeah, it's interesting. At least in at
[4:51] least in our organization, which is
[4:54] unique and and uh very niche and there's
[4:57] not that many organizations that are
[4:59] running, you know, a niche technology uh
[5:03] daily show. Yeah.
[5:04] >> I feel like a lot of
[5:05] >> what the the value that we get out of AI
[5:08] would have historically been done by not
[5:11] super expert level freelancers, right?
[5:14] these sort of like Upwork style tasks
[5:16] that you would do historically like an
[5:18] idea for a funny song, right? I've paid
[5:21] to get a funny song made probably a
[5:24] decade ago online, right, as just like a
[5:26] joke [clears throat] and now you can
[5:28] just go to Sunno and and make something
[5:29] like that. And and then there's other
[5:31] things like make a funny website, right?
[5:33] I historically would maybe work with
[5:36] >> So, you're saying that I should I should
[5:37] take down the five open roles I have for
[5:41] Celtic punk uh session musicians?
[5:44] >> Not yet.
[5:45] >> Because I was going to hire five Celtic
[5:47] punk session musicians to constantly
[5:50] record Dropkick Murphy's covers for us.
[5:52] Yes. Every day
[5:53] >> and then perform that you shouldn't do
[5:55] that. [laughter] I'm actually closer
[5:58] than ever to hiring a full-time Celtic
[6:01] punk band to to play music to recreate
[6:04] songs.
[6:05] >> Yes, I'm I'm closer than ever uh to
[6:08] doing that where that was not even on
[6:09] the road map uh a few years ago. Yeah. I
[6:11] don't know. It's it's a good point
[6:13] there. Yeah. There there's a lot of
[6:14] things that you uh are doing that you
[6:16] would never do with a full-time
[6:18] employee. Yeah. uh that just sort of
[6:20] like fills the cracks and allows you to
[6:21] do more different things in your
[6:23] organization, but the core stuff is
[6:25] still like you want a person that's
[6:26] responsible and then you want them using
[6:28] AI. I don't know. Yeah. Um the Wall
[6:30] Street Journal uh breaks it all down,
[6:32] but we went through most of that. So
[6:33] Bryce Roberts, he's he's taking the
[6:35] other side of this. He says he shares a
[6:36] screenshot of a text message says, "We
[6:38] honestly aren't hiring a ton right now.
[6:40] AI backfilling most roles." Backfilling
[6:43] is that specifically does that
[6:45] specifically refer to when someone
[6:46] leaves the company you backfill them
[6:49] with AI? So you say, "Oh, someone quit."
[6:52] Like if there's Steve and Jim on two
[6:54] different on one team and Steve quits,
[6:56] you say, "Hey, Jim, can you just instead
[6:58] of hiring another person just do twice
[6:59] as much work with AI?" Is that what this
[7:01] person's articulating? I mean, obviously
[7:04] there's some companies that are like,
[7:05] "Yeah, we're not hiring anyone. We're
[7:06] going for the one person$ one billion
[7:08] dollar company. Like I'm not going to
[7:09] hire anyone. I'm just going to use
[7:10] >> Yeah, but that's rare. Usually, usually
[7:12] when your business is ripping, you're
[7:14] like, I can't hire great people fast
[7:16] enough. Yeah.
[7:17] >> And sometimes you actually sometimes you
[7:18] actually don't have time to
[7:20] >> invest into various hiring processes.
[7:23] But yeah,
[7:24] >> uh yeah, I would read into this text the
[7:26] company's just probably not like
[7:28] ripping.
[7:30] >> That's my that's my takeaway.
[7:31] >> Well, Bryce Roberts says, "Rip newrads."
[7:33] Matthew Prince over Cloudflare takes the
[7:35] other side. He says, "Wrong strategy to
[7:37] stop hiring new grads." the right
[7:39] strategy, hire them and insert them into
[7:41] legacy teams to help them better adopt
[7:43] AI.
[7:43] >> And Cloudflare of course hired 1,000
[7:47] >> something. It was a It was a crazy
[7:49] number, wasn't it, up there in like
[7:50] almost a thousand
[7:51] >> four digits.
[7:52] >> That's crazy. Let's play the latest Good
[7:54] Work uh real. We're just watching reals
[7:57] now.
[7:58] >> This is one we got is a Bernie Maidoff.
[8:03] >> Pretty good
[8:04] >> gloves.
[8:05] >> That's from the 80s, too. That's good.
[8:06] >> Yeah. Yeah. I've I've seen a few of
[8:08] these around before. Up next,
[8:10] >> solid Sam Bankman Freed here.
[8:13] >> That's nice.
[8:14] >> Really nice.
[8:14] >> That's really nice.
[8:15] >> I like that they actually printed these.
[8:18] >> I think he made balsa wood.
[8:20] >> This is balsa wood.
[8:22] >> The acting is so
[8:26] >> Wow.
[8:27] >> Wait, John.
[8:28] >> This is a This is a vintage Zuckerberg.
[8:30] >> 05.
[8:31] >> This is a vintage '05 Zuckerberg. Let's
[8:33] just check the back really quick.
[8:36] There it is. That is a patch from his
[8:39] Fruit of the Loom boxer briefs. You can
[8:41] tell by the smell.
[8:42] >> Is that real? What is that referring to?
[8:45] >> This is on athlete, you know, trading
[8:48] card. I'll put a piece of the jersey
[8:50] one.
[8:51] >> Yeah. Yeah.
[8:51] >> All right. Up next.
[8:52] >> Sleeve it.
[8:53] >> Ooh. Okay.
[8:54] >> Nice.
[8:55] >> Elizabeth Holmes. We do have two. I
[8:57] believe we have two. But a triple Holmes
[9:00] is what every good collector has
[9:03] >> in their arsenal. All right. We have one
[9:04] last card.
[9:05] >> One one card left.
[9:06] >> Three, two, one.
[9:09] >> Oh my god. [laughter] Oh my god.
[9:10] >> Oh my god. Oh my god. Oh my god.
[9:12] >> Turn it off.
[9:14] >> Very funny. Very funny.
[9:15] >> It is funny how the the like business
[9:18] comedy cannon has really solidified
[9:20] around like Elizabeth Holmes. Uh Sam
[9:23] Bankenfreed, Mark Zuckerberg. There's
[9:25] like a few names.
[9:26] >> I'm surprised they didn't have an Adam
[9:27] Newman rookie card in there.
[9:29] >> I don't know if Adam Newman is like a
[9:30] big enough name relative to
[9:33] >> That's true. Sam Bank and Elizabeth
[9:34] Holmes. Uh it's just interesting like
[9:36] the different the different names that
[9:37] have broken out that you can do a comedy
[9:40] sketch that's like you know it goes as
[9:42] big as uh as good work does because they
[9:44] get you know I think millions and
[9:45] millions of views on their stuff.
[9:46] >> All right, pull up this image from
[9:48] Manhattan this morning. We got sent
[9:50] this.
[9:51] >> We've been doing on the ground reporting
[9:53] >> from one of our on the ground reporters
[9:55] in Manhattan. There's a company called
[9:57] Black Sheep
[9:58] >> that that got 20 trucks and they are
[10:01] just driving them around Google's
[10:04] Manhattan office. Yes. Saying, "Shame on
[10:06] you, Google. Return our $80,000."
[10:10] >> We had to dig in. We got very curious.
[10:12] >> I had no idea. They make sunglasses.
[10:16] >> They make $8 sunglasses that beat $350
[10:20] sunglasses in an NBC lab test.
[10:23] >> Okay. Are you wearing black sheep today?
[10:25] What you
[10:25] >> I wish [laughter]
[10:26] I wish. So, uh, Black Sheep makes direct
[10:29] to factory.
[10:31] >> Okay. Factory direct prescription
[10:33] eyewear. Stop paying
[10:35] >> No. This is from their own website.
[10:37] They're saying direct to factory optical
[10:40] disruptor, [laughter] which is not. This
[10:42] is from their website. Where you uh on
[10:45] Black Sheep.
[10:46] >> I'm on black sheep.io as well. It says
[10:48] factory direct direct to factory. Look,
[10:52] I want to send some eyewear to a
[10:53] factory. Direct to factoryactory.
[10:55] >> I'll be sending it to them.
[10:57] >> Direct toactory.
[10:58] >> Yeah.
[10:58] >> Uh so this company, they say direct
[11:01] toactory optical disruptor black sheep
[11:03] launches 25 truck gorilla campaign
[11:07] against Google in Manhattan. Okay. And
[11:09] then they're they're sort of like
[11:11] narrating their own gorilla campaign. A
[11:13] fleet of 25 minimalist LED billboard
[11:16] trucks surrounds Google's Chelsea
[11:18] headquarters after the tech giant
[11:21] weaponized an organic search glitch to
[11:23] pocket nearly $80,000
[11:26] in ad spend following Black Sheep's
[11:28] viral NBC Today Show debut.
[11:32] 25 LED trucks deployed. $77,000 drained
[11:36] in 30 hours. And then they're and then
[11:39] they're just continuing to market their
[11:40] own product. So very interesting
[11:41] strategy here. I think every marketer
[11:44] has had the experience of of having a
[11:46] campaign go haywire. Yeah.
[11:48] >> Very fascinating to take to take this
[11:53] route. Let's see how it works for them.
[11:55] If I were Google, I would say you can
[11:58] have your $80,000 back, but you can
[12:01] never advertise on Google again because
[12:03] I just don't know how. I don't think
[12:05] Google would ban them permanently for
[12:06] this. This is ridiculous. But it, you
[12:08] know, they're just going to they're just
[12:10] going to be like any other like as a
[12:12] self-s served platform.
[12:13] >> But is it a good campaign?
[12:14] >> But what what actually happened? So they
[12:16] say how it unfolded. NBC Today show
[12:18] segment airs. Uh they test the retail
[12:21] subscription against Black Sheep's
[12:22] factory direct pair. National search
[12:25] traffic spikes. Hundreds of Americans
[12:27] search Black Sheep because they're
[12:29] seeing it on TV. The organic listing
[12:31] breaks. Google search engine redirected
[12:33] organic brand traffic to a dead-end
[12:36] third-party 404 error page. And so with
[12:40] the organic route broken, users were
[12:43] funneled into Google's paid listings. So
[12:46] what is their claim? How is Google
[12:49] responsible for this? Exactly.
[12:51] >> Sounds like user error
[12:52] >> because I mean you do you do have some
[12:54] control over your Google search results
[12:56] based on the web master tools. you can
[12:58] index certain things and then also if
[13:00] you're noticing a 404 page you could
[13:02] like redirect it quickly but again if
[13:04] this is happening all very fast
[13:07] but I mean it is interesting because
[13:08] they're probably going to get more than
[13:10] $77,000 worth of organic just from this
[13:12] I mean I didn't see the original
[13:14] campaign and I'm seeing this because
[13:16] this is hilarious but this is like a is
[13:18] this they they shared an AI image with
[13:21] tons of these like shame on you trucks
[13:24] but those are real
[13:25] >> these are real
[13:26] >> and are those minimalist
[13:28] Those seem maximalist to me,
[13:30] >> but maybe they're minimal.
[13:31] >> Minimalist, I guess, in the in the
[13:33] display of the uh in the way they
[13:35] actually are leveraging the space on the
[13:38] truck, but black and white.
[13:39] >> Truly underrated surface area for stunts
[13:42] and advertising. like like this message
[13:45] is sort of like squabbling with Google
[13:47] over this like sort of odd scenario. But
[13:50] you can imagine someone using this for
[13:52] something much cooler and much more
[13:54] positive and not like this uh you know
[13:56] sort of unfortunate situation for them
[13:58] where they're dealing with the you know
[13:59] fallout of a Google error.
[14:01] >> Well, we want to interview the truck
[14:03] drivers.
[14:04] >> So if you're driving a black sheep truck
[14:06] around Manhattan today, reach out
[14:08] >> for sure.
[14:09] >> Show Nick, make it happen.
[14:11] >> Ilio said straight shot to SSI. So, they
[14:13] better not be gearing up to release a
[14:14] work agent called France. [laughter]
[14:18] [gasps] Francois would be a very good
[14:19] name for a for an AI agent. I like that.
[14:23] Uh, I do wonder what they're going to be
[14:25] releasing. Has the SSI is going to
[14:27] release. Is that complete rumor? Because
[14:30] all all I said scale their research. So,
[14:34] that just means they've done a bunch of
[14:36] research. They have some sort of
[14:37] architecture that they like, some sort
[14:38] of flywheel, and they're going to like
[14:40] use more compute. And so that's why
[14:42] they're raising money. I don't think
[14:43] they said like and we're going to
[14:45] release it publicly.
[14:46] >> Yeah.
[14:46] >> But everyone's thinking like probably
[14:49] still LLM or something different. No one
[14:52] really knows.
[14:52] >> Yeah. I mean I think still broadly like
[14:54] generally
[14:55] >> post god like next [clears throat]
[14:57] level. There are levels to vague
[14:58] posting. When you live a vague life just
[15:01] like your entire life is vagory.
[15:04] Anyway, uh, in other news, recursive
[15:06] super intelligence signs a $410 compute
[15:09] [laughter] deal with Amazon.
[15:12] So funny.
[15:14] And it's in the TechCrunch. It's in the
[15:17] It's in the [laughter] the the header,
[15:19] too. Of course, that is a typo. Uh, it
[15:21] says recursive super intelligence signs
[15:23] 410 million million dollar compute deal
[15:27] with Amazon. Congratulations to
[15:28] recursive super intelligence. Throwing
[15:31] safety out the window. That should be
[15:32] the that should be the uh the tagline
[15:35] because there's already safe super
[15:36] intelligence but we're just doing
[15:37] recursive super intelligence over here.
[15:39] Um but of course the company is doing
[15:41] very well. They emerged from stealth in
[15:44] May with 650 million in funding focused
[15:48] on building open-ended self-improving
[15:49] systems and potentially compute
[15:51] intensive approach to AI research. This
[15:53] multi-year deal is meant to provide
[15:54] flexibility as the company looks to
[15:56] scale up those systems. Uh, Recursive's
[15:58] $410 million outlay represents the bulk
[16:02] of the company's fundraising to date,
[16:04] but on a call with TechCrunch, uh,
[16:07] >> hey, hey, they still have a couple
[16:08] hundred million left over.
[16:09] >> Founder and CEO Richard Socher
[16:11] emphasized that he expected it to be the
[16:13] first of many such deals. So, is this I
[16:16] feel like normally when you see a like a
[16:18] compute deal signed, it's always like
[16:20] more complicated than just like we're
[16:22] buying this expensive thing. It's
[16:24] usually like we're we're paying that.
[16:26] I'm like it's we we used to be so like
[16:28] anti- circular deal that now I just have
[16:32] come to I've been so normalized by them
[16:33] that I expect them every time and I'm
[16:35] like wait
[16:36] >> this is just there's no circularity
[16:38] here. I would have expected changing
[16:40] hands.
[16:40] >> Yeah. Like Amazon's investing in you and
[16:41] you're buying tranium and racking it in
[16:43] AWS doing new campus and they're
[16:45] investing and this and that and you're
[16:47] investing in them. Uh instead it just
[16:48] seems like it's a pretty vanilla deal.
[16:50] It's like they're just buying a lot of
[16:51] compute from Amazon. Great.
[16:52] >> Seems like it.
[16:53] >> Well, good luck to them. very excited
[16:56] for what they're launching. Um,
[16:57] >> yeah. Uh, Jason, [clears throat] VP of
[16:59] startups and VC at AWS, says part of the
[17:02] agreement is that we're going to
[17:03] co-develop him for a purpose built for
[17:05] these types of companies. So,
[17:06] >> fingers crossed, but it seems like we
[17:08] could get some some circularity.
[17:10] [laughter]
[17:15] Yeah, let's hope so. Uh, let me tell you
[17:17] about Railway. Railway is the all-in-one
[17:20] intelligent cloud provider. Use your
[17:21] favorite agent to deploy web app
[17:23] serveres and more while Railway
[17:25] automatically takes care of scaling,
[17:27] monitoring, and security.
[17:30] Fingers crossed. [laughter]
[17:33] Well,
[17:35] well, here's here's a deal that's
[17:37] somewhat circular. We got Nvidia
[17:39] revealed as a tenant for a $50 billion
[17:42] data center that will use its chips. So,
[17:44] they're the tenant of the data center
[17:46] that uses its chips. We'll talk to Tate
[17:48] Kim about this. Uh CEO Jensen uh Wong
[17:52] deploys balance sheet to backs stop
[17:54] growth of AI computing market. Nvidia
[17:56] has signed leases worth up to $50
[17:58] billion for a massive Texas data center.
[18:01] That's very very big. That's very very
[18:03] big for a single site. Uh a previously
[18:06] undisclosed commitment that shines a new
[18:08] spotlight on the chips on the chip
[18:10] group's growing role in financing AI.
[18:12] The nearly $5 trillion company is
[18:14] leasing the entire gig 1 gigawatt
[18:16] facility that developer Hut 8 is
[18:18] building which will house hundreds of
[18:20] thousands of Nvidia graphics processing
[18:21] units. So you have to imagine that once
[18:24] they have these they serve something or
[18:26] wind up selling them. Like these things
[18:28] change hands so many times there's a lot
[18:29] of different ways that this could play
[18:30] out ultimately. Um but in the uh Can we
[18:33] pull up the Nvidia chart? Yeah, there we
[18:35] go. Nvidia,
[18:37] >> big candle today, up 3%, 5.17 trillion.
[18:42] Let's take a look at Apple
[18:44] >> 4.99. They crossed five today. Uh
[18:48] they're down a little bit since they
[18:50] since they beat breached that, but
[18:52] they're neck and neck. Uh Google is
[18:55] sitting at
[18:56] >> Apple running the do nothing win
[18:58] strategy. Yeah. Jensen doing thousands
[19:01] of deals.
[19:02] >> They didn't even sign the open letter.
[19:04] Uh there are three companies that still
[19:06] I I believe still haven't uh signed the
[19:09] open letter and
[19:10] >> only three companies on the entire
[19:12] [laughter] surface of the earth.
[19:13] >> There are three there are three major
[19:15] companies that haven't signed that
[19:16] Nvidia open letter about uh banning open
[19:19] source uh and or not banning open
[19:21] source. Amazon, Apple and Anthropic.
[19:24] Anthropic put out a uh uh a post
[19:27] yesterday very very clear response sort
[19:30] of outlining their view on open source
[19:33] their stance. Apple and Amazon haven't
[19:36] signed and they both have like very
[19:39] physical elements in the world in the
[19:41] sense that they're not uh they're sort
[19:43] of unsloppable like you you can't vibe
[19:45] code an Amazon warehouse. You can't vibe
[19:47] code an iPhone. There are threats to
[19:50] those businesses of course and of course
[19:52] Apple should benefit from open source
[19:54] and so should Amazon because they'll be
[19:56] able to serve open models across AWS.
[19:58] But it's just potentially interesting. I
[20:00] think the Apple standing it standing
[20:02] back is more just like look we're not
[20:03] jumping on with this crazy open letter
[20:05] that everyone is signing like we just
[20:06] have our own brand. We're thinking
[20:08] different. We're doing
[20:08] >> well and based on other Apple AI
[20:10] timelines I would expect them to sign it
[20:12] in maybe a year or two
[20:16] >> potentially. These glasses have have
[20:18] changed you. They turned you a new
[20:19] beast. Uh let me let me run through the
[20:22] the three anthropic proposals because uh
[20:24] it's an important response. So Dario
[20:26] Amade anthropic CEO uh responded
[20:28] directly to that letter summarizing uh
[20:30] supporting openweight models uh that
[20:32] circulated over the weekend. So he makes
[20:34] three claims just to sort of clarify
[20:36] things that I think are important. One
[20:38] he says anthropic has never advocated
[20:39] for a ban on openweight models. Now that
[20:42] that's a blanket ban. There's obviously
[20:44] like defining what a ban is and what an
[20:47] openw weight model, what a distilled
[20:48] model, what a foreign model is. These
[20:50] things all matter. But he has he has
[20:52] come out and said look, we never
[20:53] advocated for a total ban on openweight
[20:56] models. Uh two, he says uh undergirling
[20:58] all of this is the US must beat
[21:00] authoritarian governments in the AI
[21:02] race. He points to China, but he
[21:04] identifies any authoritarian government
[21:06] if they get really powerful AI, they'll
[21:08] come over here and steamroll us and you
[21:10] won't be free to do whatever you want to
[21:12] do in America. Uh three powerful AI
[21:14] models may be misused to carry out cyber
[21:16] attacks or biological attacks. There are
[21:18] there are risks to having really really
[21:20] powerful uh AI opensource systems just
[21:24] running around. So he's worried about
[21:25] those three things clarifying those
[21:27] three points. But he makes three uh
[21:30] recommended actions. He makes three
[21:31] proposals. Uh first he says uh let's
[21:34] continue to sanction chips. Let's not
[21:36] sell chips to China. He says we should
[21:38] not sell powerful chips or chipmaking
[21:40] equipment to China. So this has been
[21:42] debated for years like going back to the
[21:44] Biden chip uh controls. Um everyone
[21:46] knows every different angle on this uh
[21:49] the basics. I mean there is a pretty
[21:51] good argument for chip controls. Uh even
[21:53] on purely geoeconomic competitive
[21:55] grounds like even if you don't believe
[21:57] in the the the risk of authoritarian
[21:59] governments having powerful AI even if
[22:01] you just think it's like you know fancy
[22:03] autocomplete it's like well it's the
[22:04] engine of our economy and if you can
[22:06] slow down arrival economy that's
[22:08] beneficial to you right? Um and so and
[22:11] there also seems to be basically
[22:13] unlimited demand for chips in America.
[22:15] So uh by restricting sales to China that
[22:17] shouldn't actually hurt American chip
[22:19] companies all that much. Uh but yes,
[22:21] >> well they just like the their argument
[22:23] would be we fully lose the Chinese
[22:26] >> market. Yeah. Which is
[22:28] >> the second largest Yeah.
[22:29] >> computing market in the world. No.
[22:31] >> Right. So So I think um but but you're
[22:34] the counterpoint to that is you were
[22:36] going to lose it anyways.
[22:37] >> Yeah. Um and a lot of that it stems from
[22:39] the fact that chi China has been
[22:41] building an indigenous chip supply for
[22:43] decades. We've talked about going back
[22:45] to uh a whole bunch of their you know
[22:48] state-led statef funded chip and fab
[22:52] processes. They've always been uh a few
[22:54] years behind. Uh and so maintaining that
[22:56] gap all else equal is an advantage for
[22:58] the United States. Uh the second point
[23:00] Dario makes is he says we should crack
[23:02] down on industrialcale distillation
[23:04] operations. Uh this seems totally
[23:06] reasonable. companies can set their
[23:08] terms of service and they have a right
[23:10] to maintain intellectual property uh
[23:12] with proper legal consequences for
[23:14] violations. Anthropic's been fighting
[23:16] distillation attacks, but according to
[23:18] them, it's not that effective. Daario
[23:20] proposes policy interventions to deter
[23:23] this behavior. And this is where I'm
[23:25] still not clear on what where that goes
[23:29] next. Like what is the correct policy
[23:31] intervention? There's a like policy
[23:33] intervention is a very very broad thing.
[23:34] It can mean anything from like a tax, a
[23:37] tariff, a fine, uh a sternly worded
[23:39] letter, uh not getting invited to a golf
[23:42] tournament. Like there's so many
[23:43] different things that policy like covers
[23:45] these days, right? Um where where does
[23:48] this actually go? He says he doesn't
[23:50] want a blanket ban on openweight models,
[23:53] but it does seem like one possible
[23:55] policy intervention would be to sort of
[23:57] like ban restrict or pressure open
[23:59] weights models that can be reasonably
[24:00] shown to have been distilled. So if
[24:02] there's someone who's just a perfect
[24:03] distillation, it just gets it just
[24:06] doesn't quite feel right. It's hard to
[24:08] quantify these things. We don't have a
[24:10] binary where you run some sort of
[24:12] algorithm and you say, "Yes, this was
[24:14] distilled." Because you can distill, you
[24:16] know, half on Opus 5 and then throw in a
[24:18] little GPT 5.6 and then mix in some some
[24:22] mistrol and like just be distilling from
[24:23] all over the place. Fine-tune stuff,
[24:26] change the flavor, change the RL
[24:28] environment. There's so many different
[24:29] pieces of it. And Tyler, you were making
[24:31] a point about uh Tinker or
[24:34] >> Yeah. So, so the the Inkling model from
[24:36] Thing Machines like it used some synthet
[24:38] synthetic data that was created with uh
[24:41] I think Kimmy K2.5.
[24:43] >> Yes.
[24:43] >> So like I does that count as like
[24:45] distillation like probably not when
[24:47] people usually talk about it but like it
[24:49] definitely benefited from Chinese open
[24:51] source models.
[24:52] >> Yeah. So I wouldn't call that downstream
[24:53] of
[24:54] >> Yeah. I wouldn't call that industrial
[24:55] scale distillation. Yeah.
[24:58] >> There is like some big gray area where
[25:00] like how do you actually define these
[25:02] things?
[25:02] >> Yes. And so defining that is going to be
[25:04] what that's going to be the conversation
[25:06] that plays out in in DC like behind the
[25:09] scenes on the basis of this and that's
[25:12] where the actual negotiation is going to
[25:13] happen between uh you know the the
[25:16] position of Nvidia and everyone that
[25:18] signed the letter versus the position of
[25:19] Enthropic and everyone who didn't sign
[25:21] the letter. They're going to sort of
[25:22] decide okay well if you can prove this
[25:25] this and this and you can show us that
[25:26] your API was getting hit by these
[25:29] different things and you have a really
[25:30] solid report of what happened and then
[25:33] the model also you know sort of you know
[25:36] checks these boxes quantitatively when
[25:38] we do this eval then maybe we will
[25:41] pressure it and then what does that
[25:43] actually mean? Um, you could go after
[25:46] the lab that committed the distillation
[25:48] attack with lawsuits, but that seems
[25:50] really difficult given the international
[25:51] nature of these attacks. So, we're sort
[25:53] of back to where we started. Like, what
[25:55] what can the government do that the lab
[25:57] can't? Like, the lab should be looking
[25:59] at every customer and saying, "Oh, this
[26:02] seems like someone who's trying to
[26:03] distill. They keep asking for basically
[26:05] what looks like a lot of training data.
[26:06] They they're not acting like a normal
[26:08] user just being like, "Build me a
[26:10] website." Okay. Third, he says, "All
[26:12] sufficiently capable models, open and
[26:14] closed, should go through mandatory
[26:16] mandatory safety testing." So, this was
[26:19] recently outlined by Dennis Hassabis
[26:20] over at Google Deepind as well. And and
[26:23] it seems like that the two companies are
[26:25] in alignment on this in particular. Uh,
[26:28] and it's a somewhat reasonable position.
[26:30] Uh, although the risk is that small
[26:32] companies who have safe models could
[26:34] that aren't distilled could get tied up
[26:36] in a review queue for years before they
[26:38] can release. Like that would be very
[26:39] very annoying if you're uh recursive
[26:42] super intelligence for example and you
[26:44] don't have a Washington DC office and
[26:47] you're like hey we want to release our
[26:49] new model and they're like yeah totally
[26:51] like you got to go through the the the
[26:52] the review process get in line and then
[26:55] it's like every you know every trillion
[26:58] dollar company is there with a ton of
[27:00] lobbyists being like we'll review our
[27:01] model first because we want to get out a
[27:03] week before the the small startup and
[27:05] that's the frustration of biotech the
[27:08] FDA anything that goes through approval.
[27:10] We've talked about this with the nuclear
[27:11] stuff. Uh it gets very tricky and so you
[27:14] want to avoid that and you don't want to
[27:16] wind up slowing down innovation that's
[27:18] happening on small scales and decreasing
[27:20] competition. Dario does do a good job of
[27:22] like acknowledging upfront that he says
[27:25] it would protect a US AI companies from
[27:28] competition, but that's never been my
[27:29] goal with anything that he's saying
[27:31] here. And so, uh it's still worth
[27:33] working through what happens in a really
[27:36] adversarial situation. Like what if a
[27:38] foreign lab distills a bunch of frontier
[27:40] models, they're the most aggressive.
[27:42] They're just distilling everything. Then
[27:44] they jump forward a bunch in capability.
[27:45] Uh they get a bunch of smuggled chips.
[27:47] They take all the restrictions off of
[27:48] cyber, all the restrictions off of bio.
[27:50] Uh and then they just drop the weights
[27:52] on like a torrent or they put them up on
[27:54] Hugging Face and Hugging Face is like
[27:55] this is really crazy. No one likes this.
[27:57] There's a lot of pressure to take it
[27:58] down. I don't know. But it's out there.
[28:00] Like what does the government actually
[28:02] do? like the government probably
[28:04] pressures or bans like hosting the
[28:06] weights uh maybe serving the model you
[28:09] maybe won't be able to run it in
[28:10] American data centers you go to the
[28:12] neocloud and say like hey this thing is
[28:14] actually bad and I think people are
[28:16] divided on this because they see the
[28:18] current models not as actually dangerous
[28:20] which is totally reasonable to assess
[28:22] that yeah it's not that bad um but like
[28:24] if there was a model that was like yeah
[28:26] it's actually just like the killing
[28:27] machine like I think most people would
[28:29] be like yes I'm democratically voting to
[28:32] not serve that because it's just like
[28:33] it's an annoyance at at best and like
[28:36] actually bad at worst.
[28:37] >> And the other big question is like how
[28:39] much compute do you actually need for it
[28:40] to be dangerous, right?
[28:41] >> Yeah, totally.
[28:42] >> Is is uh like having some GPUs in the
[28:45] back shed going to be enough. Maybe for
[28:48] sufficiently advanced model, yes. Or do
[28:50] you need access to a ton of racks? Yeah.
[28:54] >> Ton of power. And then you do need to
[28:56] work with
[28:57] >> a Neocloud in that case. And as soon as
[28:59] you're a US-based company with a real
[29:02] data center with a bunch of NVL72s in
[29:04] there, uh, you probably have
[29:06] registration and, you know, all sorts of
[29:08] just like business registrations where
[29:10] the government can reach out to you and
[29:12] say, "Hey, we're actually really worried
[29:13] about this." Just like there are other
[29:15] things you can't host in a data center.
[29:17] There's all sorts of stuff that's
[29:19] illegal, even just intellectual
[29:20] property.
[29:21] >> Yeah. Exactly. Yeah. That's it's uh
[29:23] >> like you can't even just just because
[29:24] you have a data center doesn't mean that
[29:26] you can like take an open-source you
[29:28] know uh
[29:28] >> you can't you can't as a data center
[29:30] >> oh yeah open source Marvel like they'll
[29:32] be
[29:32] >> or even even a you know a CRM company
[29:35] can't knowingly support like a organized
[29:39] cart global cartel that is like
[29:41] trafficking narcotics right you have
[29:43] you'd have to imagine like
[29:45] >> uh that uh they have
[29:47] >> they have to vi balances [laughter]
[29:51] maybe.
[29:52] >> So, uh, so, so, uh, what, what's
[29:55] interesting is like what is the next
[29:56] step of that? So, if there is a bad
[29:59] model, uh, and and everyone agrees like,
[30:02] okay, yeah, we got to not host this, not
[30:04] distribute this, like yeah, the weights
[30:05] are out there, people are trying to like
[30:07] sort of run it a little bit. Uh, but
[30:09] does it go offshore? Do we wind up in
[30:11] like the crypto scenario where there's
[30:12] like these offshore things and people
[30:14] are using VPNs to get access to it? like
[30:16] what level of aggression do you see from
[30:19] the US government in that scenario? It
[30:22] probably should be proportionate to like
[30:23] the danger imposed by the model. Like if
[30:25] it's just a model that's like that's
[30:26] like annoying or like slightly IP
[30:28] infringes, but like no one's really
[30:30] being like I'm not I'm I'm canceling my
[30:32] Disney subscription because this new
[30:34] model will generate me Disney IP. Like
[30:36] that's probably not like okay, put up a
[30:38] crazy firewall. But if it is like the
[30:40] ultimate hack machine that's like
[30:42] stealing everyone's money from the
[30:43] banks, then yeah, you are going to put
[30:45] up the the firewall and sort of be much
[30:47] more aggressive. So I think the response
[30:49] will be in reaction to whatever the
[30:51] power of the models are. But uh it'll be
[30:53] interesting to go back and forth.
[30:54] Anyway, all in all, the letter clarifies
[30:56] a lot about the anthropic position. So I
[30:57] think it's good that it it came out. Uh
[30:59] but it's still worth working through the
[31:01] game theory of like what happens down
[31:02] the line. Policy interventions is all we
[31:04] got here. And I think it's still too
[31:06] generic at this point. I want to know
[31:08] like what policy looks like. I want to
[31:10] predict that. I want to understand uh
[31:12] what's actually being proposed, what
[31:13] people like, what people don't like. And
[31:15] so I think we'll learn more about this
[31:16] in the coming days.
[31:17] >> Mark Zuckerberg is in the Wall Street
[31:20] Journal opinion section with a new
[31:22] piece. The AI future is for everyone. He
[31:25] says, "The history of democracy and
[31:27] economics has proved that centralized
[31:29] power stifles human potential."
[31:33] >> And uh it's quite long. go let you guys
[31:37] read it, but let's head into the comment
[31:39] section. Let's get a quick Let's get a
[31:41] quick rea Let's get a quick reaction.
[31:43] >> Uh this is the Wall Street Journal. I
[31:44] think it'll be
[31:45] >> um No, it it it looks relatively tame.
[31:48] Um but yeah, making a you know, a clear
[31:51] effort to position to to be the overtly
[31:55] >> there was a white space for a guy
[31:57] investing hundreds of billions of
[31:58] dollars a year in AI that is like says,
[32:00] "Hey, this is going to be really great
[32:02] for everyone." Yeah. And I'm going to
[32:03] help us get there.
[32:04] >> It is It is interesting. I Facebook does
[32:07] have some monopolies, but like the
[32:10] competition for attention is constant
[32:13] and there are always sources outside
[32:16] like they've never had a full monopoly
[32:19] on social media even with Tik Tok and
[32:23] Snapchat and uh LinkedIn and Twitch and
[32:27] YouTube and Netflix and the podcast feed
[32:31] and SMS and iMessage. like there are so
[32:34] many other platforms for disseminating
[32:36] information like I I don't know I I it's
[32:39] hard to jump straight to a critique here
[32:42] but um the the key quote that Andrew
[32:44] Curran pulled out was that he said in
[32:46] most cases like cyber security the
[32:48] history of open- source software has
[32:50] shown that giving everyone full access
[32:53] to powerful systems will be the best way
[32:56] to protect safety and security over
[32:58] time. So, he's firmly on the side of uh
[33:01] democratizing powerful AI and uh he is
[33:05] yet another one. Uh I imagine that they
[33:07] that they signed the letter. I I've lost
[33:08] track at this point, but you can imagine
[33:11] that he did. Um anyway, thank you so
[33:14] much for tuning in. The other piece of
[33:16] news is that Apple is launching Apple
[33:19] Upgrade next week, then iPhone, iPad,
[33:22] Mac, and Apple Watch
[33:24] leasing/subscription
[33:26] program. Uh they said you will own
[33:28] [laughter]
[33:29] >> and you will be happy.
[33:31] >> We're launching our new program. You
[33:33] will own nothing and be happy.
[33:34] >> Uh it's partnering with Clara to launch
[33:37] in the United States at online and
[33:38] retail stores. Uh it's now official.
[33:40] Leasing prices start as low as $20 or
[33:43] $17.99 per month for iPhone. Uh $11.99
[33:48] for Apple Watch, $24.99 for Mac, and
[33:51] $11.99 for iPad. So uh interesting. I
[33:56] mean, a lot of people are saying this is
[33:57] a direct reaction to uh increased prices
[34:00] for memory, increased prices for
[34:02] products. There was a time when an
[34:04] iPhone was a couple hundred and there
[34:08] were incentives to jump on a Verizon
[34:10] plan and you sort of advertise the cost
[34:11] over that. Those days are gone. Like
[34:13] we're in the world of like a $2,000
[34:15] iPhone. It's a significant
[34:16] >> Same thing with our gongs for people
[34:17] honestly.
[34:18] >> You want a subscription gong?
[34:20] >> No, I'm just saying there was a time
[34:21] when a TVPN gong was $200. Yeah,
[34:25] >> now it's in the tens of thousands of
[34:26] dollars, right?
[34:27] >> It's actually so expensive. Somebody a
[34:29] friend of mine texted me and was like,
[34:30] "Where do we get the gong?" Cuz I need a
[34:31] gong. And I was like, I think you should
[34:33] start small. And this is not like you're
[34:36] you can't handle the big gong. I was
[34:38] more saying that like there is a joy to
[34:40] being on the hyonic treadmill of larger
[34:43] gongs. Like you don't want to jump
[34:44] straight to the biggest gong. You want
[34:46] to start with a small gong
[34:47] >> and work your way up.
[34:48] >> Work your way up. Because every gong
[34:49] that we've added has been so electric
[34:52] when we get
[34:52] >> I think it's time for a new one.
[34:54] >> You want an even bigger gong or what?
[34:56] >> Yeah, I want I want one that's hanging
[34:57] from the rafters.
[34:58] >> Leave us
[35:00] >> five stars on Apple Podcast and Spotify.
[35:02] >> Money never sleeps. You shouldn't
[35:04] either. Call me.
[35:06] >> Sign up for the newsletter at tvpn.com
[35:08] and we will see you tomorrow at 11:00
[35:10] a.m. Pacific. Goodbye.
[35:12] >> Cheers.
[35:13] >> Flashbang.
