Trump Rejects Anthropic Slowdown Plan as Labs Seek Antitrust Shield

The AI Slowdown Debate | Diet TBPN — TBPN

TBPN

Frontier AI labs are pushing coordinated pacing pacts to establish a defensible duty of care against tort liabilities, even though voluntarily holding back model deployments risks destroying incumbent margins by letting rivals commoditize the market.

AI-assisted summary of The AI Slowdown Debate | Diet TBPN — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by TBPN on September 14, 2026, this episode of 'Diet TBPN' delivers a condensed 25-minute version of the daily live tech talk show hosted by John Coogan and Jordi Hays. The description positions TBPN as Silicon Valley's go-to tech broadcast, tackling the latest industry headlines and major political and technological debates.

## TL;DR The hosts dissect the explosive 'Pace the Frontier' AI safety debate triggered by Dario Amodei's essay calling for third-party evaluators, antitrust exemptions, and international coordination. They explore fierce pushback from David Sacks and Donald Trump, who rejected the slowdown under the doctrine of 'whoever wins AI wins' against China, while Gavin Baker contextualized the push around tort liability and a lack of Section 230 immunity. The discussion also examines whether intentional pacing actually erodes frontier margins for OpenAI and Anthropic by enabling fast followers to commoditize models. The episode concludes with breaking news of Donald Trump phoning Nvidia CEO Jensen Huang live on stage at the All-In Summit to endorse continued AI acceleration.

## Atmosphere & style The broadcast carries an energetic, irreverent tech-insider vibe, balancing serious legal and economic analysis with punchy Silicon Valley humor. Despite host John Coogan dialing in fresh out of the hospital from an intense bout of workout-induced rhabdomyolysis, the banter remains rapid-fire and sharp. The hosts move effortlessly between dissecting antitrust statutes and giggling at the absurdity of Trump placing diamonds alongside the transistor in his hierarchy of human inventions, giving the episode an authentic, live-on-the-fly conversational momentum.

## Key takeaways - [03:08](https://www.youtube.com/watch?v=wKcf6RorKvM&t=188s) The "Pace the Frontier" movement rapidly gained traction between July and September 2026, culminating in Dario Amodei's essay calling for frontier pacing, which received high-profile endorsements from Sam Altman and Elon Musk. - [05:51](https://www.youtube.com/watch?v=wKcf6RorKvM&t=351s) Dario Amodei's slowdown framework centers on three pillars: embedding third-party safety evaluators like METR, coordinating government safety thresholds, and establishing international treaties with rival nations like China. - [07:00](https://www.youtube.com/watch?v=wKcf6RorKvM&t=420s) Coordinating model pauses and safety thresholds across competing frontier labs presents severe antitrust hurdles under the Sherman Act, requiring statutory waivers to avoid being classified as an illegal cartel. - [09:17](https://www.youtube.com/watch?v=wKcf6RorKvM&t=557s) David Sacks pushed back forcefully against Amodei and Altman, arguing that frontier incumbents should pause unilaterally rather than seeking regulatory capture that imposes heavy compliance burdens on smaller startups. - [11:39](https://www.youtube.com/watch?v=wKcf6RorKvM&t=699s) Donald Trump explicitly rejected the AI slowdown narrative on Truth Social, maintaining that "whoever wins AI wins" and warning that domestic compute restrictions would cede leadership to China. - [12:00](https://www.youtube.com/watch?v=wKcf6RorKvM&t=720s) Trump positioned AI data centers and energy infrastructure as an unprecedented economic engine larger than oil or the internet, while dismissing AI existential risk warnings as hoaxes. - [15:56](https://www.youtube.com/watch?v=wKcf6RorKvM&t=956s) Investor Gavin Baker highlighted that because AI outputs lack Section 230 immunity, embedded third-party evaluations are primarily designed to establish a documented "duty of care" against catastrophic tort liability. - [16:40](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1000s) Real-world physical constraints—including electrical power grids, semiconductor supply chains, and rail transport—are already acting as natural pacing mechanisms for frontier AI deployment. - [18:45](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1125s) Unlike targeted microchip chokepoints, electricity infrastructure is broad and diffuse, meaning expanding energy production remains universally beneficial to society even if frontier AI is paused. - [19:30](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1170s) Frontier AI policy proposals have shifted away from a heavy-handed federal agency model (an "FAA for AI") toward an industry-led, self-regulatory framework modeled on financial regulators ("FINRA for AI"). - [20:54](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1254s) Challenging the regulatory capture thesis, commentators noted that slowing down frontier deployment could compress margins and destroy pricing power by enabling fast followers and open-source models to catch up. - [22:10](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1330s) Because OpenAI and Anthropic are structured as Public Benefit Corporations, leadership possesses the legal authority to sacrifice valuation and prioritize safety without violating traditional shareholder duties. - [24:34](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1474s) Donald Trump phoned Nvidia CEO Jensen Huang live during the All-In Summit, where Huang publicly thanked Trump for speaking out against catastrophic AI doom narratives.

## Detailed walkthrough ### [00:01](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1s) Weekend Chaos: Rhabdo in the Hospital and Never Missing a Show

The show kicks off with a check-in on the weekend, which was anything but mellow. Aside from the explosive AI pacing debates rocking Silicon Valley, John Coogan shares that he spent time in the hospital after being diagnosed with rhabdomyolysis ("rhabdo"). Calling it "the dumbest illness you can possibly get," he explains that it was brought on by overexertion in a workout combined with dehydration. After an IV, hospital medication, and breaking a fever, he is on the mend, joking that even if there is only a 1% chance that artificial intelligence could prevent or cure rhabdo, "it's worth any amount of risk" to push frontier AI capabilities forward.

Despite the hospitalization, the hosts boast that TBPN has never missed a broadcast due to illness. When the hospital doctor asked John if he wanted a doctor's note for work, he laughed at the thought. The hosts note that the show goes on regardless because hosting simply requires sitting in a chair and having a conversation—though John warns listeners he might be slightly slower than usual.

[01:27](https://www.youtube.com/watch?v=wKcf6RorKvM&t=87s) The US Open Flow State and 007 Production Mastery

Before diving into tech policy, the hosts unpack the finish of the US Open, which made the front page of *The Wall Street Journal*. They highlight a viral moment involving American tennis star Ben Shelton: upon hitting the winning shot in a grueling back-and-forth duel, Shelton immediately prepared for the subsequent serve, completely unaware he had just won the match. Only after noticing the standing ovation in the crowd did he finally look up at the scoreboard and realize the contest was over. The hosts marvel at this pure "flow state" and compare it to venture capitalists who should theoretically stay focused entirely on their next deal rather than taking public "VC victory laps" when a portfolio seed company exits or files an IPO.

The conversation turns to another broadcast highlight from the tournament: during a countdown clock with exactly seven seconds remaining on screen, the production director seamlessly cut the live camera feed straight to Pierce Brosnan—James Bond himself—sitting in the stands. The hosts praise this unscripted bit of creative switcher work as proof that "there's levels to the game of live production."

[03:08](https://www.youtube.com/watch?v=wKcf6RorKvM&t=188s) How "Pace the Frontier" Took Over AI: The July–September Timeline

The discussion transitions into the historical timeline of the "Pace the Frontier" campaign, tracing the rapid escalation of the AI safety and slowdown discourse over the preceding weeks. The phrase entered the mainstream on July 28 with a public open statement signed by prominent researchers and executives across OpenAI, Anthropic, Google DeepMind, and Meta. That same day, Sam Altman publicly suggested that the industry might need to pace the rate of frontier development. Two days later, OpenAI disclosed to *Axios* that it had helped shape the petition's wording and that Altman had directly briefed White House officials on pacing initiatives.

```mermaid flowchart TD accTitle: AI Pacing Discourse Timeline accDescr: Chronology of the AI slowdown and pacing the frontier debate from July to September. subgraph Phase1 ["Summer Buildup"] A["July 28: Open Letter on Pacing"] --> B["July 30: OpenAI confirms WH briefings"] B --> C["Aug 18: OpenAI paper & RL training pause"] C --> D["Aug 31: Anthropic post on pacing mechanisms"] end subgraph Phase2 ["September Climax"] D --> E["Sept 12, 7:00 AM: Dario Amodei publishes essay"] E --> F["Morning: Endorsements from Musk & Altman"] F --> G["Sept 13: Trump rejects slowdown (FT)"] end ```

The momentum accelerated through August. On August 18, OpenAI published a paper titled *Pacing Model Development in an Era of Cyber Critical Capabilities*, revealing that it had voluntarily paused scaling and held up a major reinforcement learning (RL) training run for two weeks. On August 31, Anthropic published a piece explicitly endorsing mechanisms to coordinate frontier pacing. The debate reached a crescendo early on Saturday, September 12, at 7:00 AM Pacific, when Anthropic CEO Dario Amodei dropped his long-form essay, *We Must Pace the Frontier*. Published during the quiet Saturday morning news window, the post captured tech's full attention:

- **8:00 AM PT:** Elon Musk publicly endorses Amodei’s essay. - **9:30 AM PT:** Sam Altman offers his endorsement. - **4:00 PM PT:** Demis Hassabis (who previously raised similar oversight concerns before shifting roles at DeepMind) indicates broad agreement, noting details still need refinement. - **8:15 PM PT:** David Sacks engages with critical questions about the proposal. - **Sunday Morning:** *The Financial Times* reports Donald Trump has formally rejected tech leaders' calls for an AI pause, followed by investor Gavin Baker publishing an extensive weekend recap.

[05:51](https://www.youtube.com/watch?v=wKcf6RorKvM&t=351s) Dario Amodei's Three-Step Slowdown Plan: METR, Antitrust Cartels, and US–China Talks

The hosts deconstruct Dario Amodei's three-step proposal for pacing frontier models, noting how each pillar raises critical legal, structural, and geopolitical complications:

1. **Independent Third-Party Safety Evaluators:** Amodei argues external auditors must be granted access inside frontier AI labs to inspect models and verify safety safeguards. He explicitly cites METR (Model Evaluation and Threat Research). The hosts point out that METR's independence has faced scrutiny given the perpetual "revolving door" of personnel circulating among OpenAI, Anthropic, and alignment organizations. Unlike traditional accounting firms that remain indifferent to industry outcomes, METR consists of deeply invested ideological insiders—a dynamic that provides deep technical context but also raises conflict-of-interest concerns.

2. **Coordinated Safety Standards via Government Integration:** The second pillar proposes that leading frontier labs in democratic nations work in lockstep under government supervision to mandate shared thresholds. The hosts note legal analyst Dean Ball's warnings regarding the Sherman Antitrust Act. Private market leaders cannot lawfully collude to restrict output or coordinate operations—just as major airlines like Delta, United, and American cannot collectively decide to reduce flights under the guise of safety to protect profit margins and increase ticket prices. Because such coordination constitutes an illegal cartel under existing antitrust law, AI labs cannot formalize these pacts without explicit statutory waivers or federal regulatory exceptions.

3. **International Coordination with Authoritarian Regimes:** The third and most controversial recommendation urges democratic nations to coordinate pacing thresholds directly with rival authoritarian powers, particularly China. The hosts compare this to historical Cold War nuclear non-proliferation treaties. However, achieving parity or structured pacing with Beijing poses game-theoretic dilemmas: if the United States insists on maintaining a decisive capability lead, negotiations collapse into an unstable dynamic where one side demands the other accept permanent technological inferiority.

[09:17](https://www.youtube.com/watch?v=wKcf6RorKvM&t=557s) David Sacks Pushes Back: Cartels, Antitrust, and Regulatory Capture

Following Dario Amodei’s call to "pace the frontier"—which was quickly endorsed by Sam Altman and Demis Hassabis—the broader tech timeline pushed back with immediate skepticism. Critics questioned who these unelected third-party evaluators would actually be, why frontier labs needed Washington's permission to slow down rather than simply doing it themselves, and whether this was an attempt at early regulatory capture. David Sacks weighed in with a sharply critical post on X directed squarely at Amodei and Altman. Sacks noted that OpenAI and Anthropic essentially form a duopoly at the frontier across market share, revenue growth, and model capabilities. If their unreleased internal models are genuinely terrifying enough to warrant tapping the brakes, Sacks argued, they should act responsibly and pause on their own.

```mermaid flowchart TD accTitle: Frontier AI Regulation Debate accDescr: Diagram contrasting Dario Amodei's proposal with David Sacks' pushback and Gavin Baker's liability thesis.

subgraph FrontierProposal["Amodei & Altman Proposal"] Pace["'Pace the Frontier'"] Eval["Embedded Third-Party Evaluators (e.g., METR)"] Gov["State / Federal Regulatory Approval"] Pace --> Eval --> Gov end

subgraph SacksCritique["David Sacks Pushback"] Cartel["Cartel Behavior & Antitrust Suspensions"] Capture["Regulatory Capture Crushing Startups"] Conflicts["METR Conflicts of Interest"] end

subgraph BakerThesis["Gavin Baker Tort Analysis"] No230["No Section 230 Shield for AI Outputs"] DutyOfCare["Evaluators Establish 'Duty of Care' for Litigation"] end

Gov -.->|"Risks"| Cartel Gov -.->|"Risks"| Capture Eval -.->|"Alleged"| Conflicts Eval -->|"Provides"| DutyOfCare DutyOfCare -->|"Mitigates"| No230 ```

Sacks pulled no punches in dismantling what he saw as the pretense behind the proposal: "Stop pretending you need anyone else's permission. Stop pretending antitrust law has to be suspended so you can form a cartel. Stop pretending you need a regulatory approval process that supersedes product liability." He also took direct aim at METR (Model Evaluation and Threat Research), warning that it is far from an independent referee given its deep entanglements with Anthropic’s own staff and investors, and rejected the idea that these evaluators should police smaller rivals who are nowhere near the frontier. The hosts echoed this danger of regulatory capture, pointing out that if heavy compliance requirements—like year-long negotiations with embedded government-approved auditors—are imposed across the board, breakout consumer startups (like Instinct) would be smothered before they can launch. They highlighted Greg Brockman's recent comments on *Odd Lots* arguing against imposing limits on application-layer software or individuals training models at home.

- Sacks stressed that OpenAI and Anthropic already claim their lead is widening through recursive self-improvement, giving them unilateral power to slow down without an antitrust waiver. - The hosts emphasized the distinction between frontier foundational models and application-layer tooling, warning that sweeping safety compliance regimes disproportionately crush smaller engineering teams.

---

[11:39](https://www.youtube.com/watch?v=wKcf6RorKvM&t=699s) Trump Rejects the AI Slowdown: "Whoever Wins AI Wins"

Donald Trump jumped into the discourse over the weekend with characteristically bombastic posts on Truth Social, forcefully dismissing calls to slow down AI development. Trump declared that the only "guard rails" artificial intelligence actually needs is a "strong and smart high IQ president," boasting that the United States already has that "in spades." Taking direct aim at Anthropic’s CEO, Trump wrote that his administration had already acted to prevent AI figures from doing bad things, singling out "Dario who is now pretending to be a perfect little angel." He labeled the push to restrict compute and data centers a "sick conspiracy" that directly serves Beijing's interests, flatly stating: "Whoever wins AI wins. We are leading China and all others and will continue to do so."

Trump followed up shortly after by questioning the fundamental business logic of the slowdown, asking when in the "history of business" industry leaders had ever begged for regulatory burdens that would drive their own companies into bankruptcy. He dismissed AI existential risk as an outright hoax, putting it in the same bucket as "Russia Russia Russia," Ukraine, and both impeachment inquiries. Highlighting global competition, Trump claimed President Xi Jinping had just signaled China would do nothing to impede AI development, and voiced frustration that domestic permitting friction is already driving US tech giants abroad—citing Google's decision to construct a massive data center plant in Finland due to American regulatory hurdles. Trump promised that AI and data center infrastructure will serve as the single greatest economic engine in human history: "Bigger than oil, gold, diamonds, or even the internet."

The hosts broke down into laughter over Trump's idiosyncratic hierarchy of historical inventions, marveling at diamonds being ranked alongside the internet, the transistor, and the wheel as a foundational technology. They joked about the hedge funds paying premium subscriptions for sub-second early feeds of Truth Social trying to figure out how to execute trades off the post, with one host deadpanning that algorithms must have immediately gone "long diamonds." They noted that the natural diamond industry is actually in a deep slump due to the rapid rise of lab-grown alternatives—making diamonds one industry that genuinely *is* begging for protective government regulation.

- Trump warned all "conspiracy theorists, treasonous traitors, and leakers" that his administration already holds immense criminal and regulatory oversight over domestic AI labs. - He framed permitting reform and domestic energy buildouts as existential national security priorities to keep infrastructure investment from fleeing to Europe or Asia.

---

[14:42](https://www.youtube.com/watch?v=wKcf6RorKvM&t=882s) Trump on AI Usage and the AI Slop Reality

The show transitioned to a video clip of Trump being questioned by reporters about whether he had spoken directly with tech leaders and why he seemed to downplay their safety warnings. Trump reiterated his core conviction: AI will ultimately bring "more good than bad, but by a lot," repeating his mantra that "whoever wins AI wins." When pressed directly by a reporter on whether he personally uses AI and what he uses it for, Trump gave a characteristically vague response, insisting, "You can use AI for a lot of things," without naming a single specific tool or application.

The hosts found the exchange hilarious because Trump’s Truth Social feed is perpetually flooded with AI-generated political slop. They pointed out that his feed constantly features hyper-stylized AI imagery, satirical video simulations, and "vibe-coded" clips generated by campaign staffers and digital communications teams. While Trump himself likely doesn't open ChatGPT or Claude to generate content, his political operation has embraced generative media as a primary distribution tool, creating an amusing contrast with his evasive press answers.

---

[15:56](https://www.youtube.com/watch?v=wKcf6RorKvM&t=956s) Third-Party Evaluators, Section 230 Liability, and the Energy Question

The conversation turned to tech investor Gavin Baker’s comprehensive breakdown of the weekend's chaotic developments. Observers had anticipated Elon Musk would oppose Dario Amodei's pacing proposal, but Musk surprised many by voicing support for safety checks, recalling his early funding and collaboration with Nick Bostrom on existential risk back in 2014. While Musk has oscillated between aggressive frontier competition (founding xAI and building massive clusters) and existential alarm, his fundamental concern over AI safety has remained consistent.

Baker argued that the only tangible new development across the entire debate is that OpenAI and Anthropic are moving toward embedded third-party evaluators, with METR floated as the primary candidate. Crucially, Baker contextualized this through the lens of legal liability: unlike social media and web platforms, frontier AI models enjoy no Section 230 statutory shield for their generated outputs. Establishing robust third-party evaluation allows frontier labs to document a legally defensible "duty of care" to protect against ruinous tort litigation once frontier models cause real-world harms. For investors looking for a "smoother for longer" cycle, Baker argued that physical constraints—semiconductor wafers, electrical watts, rail transport, and credit spreads—are naturally pacing the industry regardless of paper regulation.

This prompted the hosts to question the logic of AI safety advocates who focus on restricting compute and chip manufacturing. If frontier AI development is intentionally paced while existing models gradually diffuse into the economy, the hosts asked whether anyone could object to continuing a full-scale buildout of energy infrastructure. Redirecting capital, engineering talent, and permitting momentum toward nuclear reactors, solar, wind, battery storage, and grid modernization offers a universal win. Even under a worst-case safety scenario where artificial general intelligence (AGI) or artificial superintelligence (ASI) is permanently paused as too hazardous, society would still walk away with an abundant, resilient, low-cost domestic energy supply.

- Baker highlighted that several iconic internet giants would have faced bankruptcy early on without Section 230 immunity, making proactive legal derisking an existential priority for frontier AI labs. - Capital expenditure diverted away from runaway frontier model training could de-bottleneck regional power grids and bring overdue industrial energy projects online.

[18:45](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1125s) Regulating Energy, Microchips, and Deep Supply Chain Absurdities

The discussion shifts to whether regulators would ever attempt to throttle AI progress by curbing the physical build-out of the American electrical grid. The hosts find that notion implausible: while electricity is an essential input for artificial intelligence, power as an infrastructure layer is far too diffuse. While specialized microchips have a relatively narrow set of end uses—primarily powering data centers for AI or services like YouTube—energy can be directed anywhere across the economy.

```mermaid flowchart TD accTitle: AI Supply Chain Chokepoints accDescr: Comparison of different layers where regulators or safety advocates might intervene to control AI progress.

subgraph Inputs["Diffuse vs. Targeted Inputs"] Power["Electrical Grid & Power<br/>(Diffuse infrastructure, hard to justify capping)"] Chips["Semiconductor Chips<br/>(Narrow end use, high leverage point)"] end

subgraph DeepChain["Extreme Supply Chain Proposals"] Specialty["Upstream Manufacturers<br/>(e.g., Korean firms producing films and consumer goods)"] end

subgraph Frontier["Frontier Models"] Labs["Frontier AI Labs<br/>(Direct safety and compute regulations)"] end

Power -.-> Labs Chips --> Labs Specialty --> Chips ```

This leads the hosts to joke about how absurdly deep down the supply chain hypothetical "AI safety valves" could reach. They imagine safety advocates deciding to regulate an industrial conglomerate in South Korea that manufactures consumer toilets simply because an obscure subsidiary produces specialized chemical films used in semiconductor lithography. Such hypothetical extremes illustrate the bizarre territory governments enter once they attempt to slow down technology by targeting upstream ingredients rather than software deployments.

[19:30](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1170s) Gavin Baker’s Breakdown of Dario Amodei’s Proposal and the Sherman Act Waiver

The hosts turn to investor Gavin Baker’s comprehensive breakdown of the proposal Dario Amodei floated over the weekend, which Baker characterized as the most maximalist plan put forward. The proposal includes embedded third-party evaluators inside labs, a domestic regulatory regime triggered when models hit specific capability and ingredient thresholds, a binding international regulatory pact among democratic nations, and tighter restrictions on compute distillation to prevent Chinese firms from copying American frontier models. Crucially, Dario also suggested an international framework encompassing China before certain domestic regimes take effect, as well as an explicit Sherman Act waiver so Anthropic can freely coordinate safety standards and deployment schedules with OpenAI and other frontier rivals without fear of antitrust prosecution.

The hosts note that while critics view this as extreme, it is actually toned down compared to Dario’s earlier calls for an "FAA for AI." Current policy discourse has shifted toward proposing a "FINRA for AI"—a non-governmental, industry-funded self-regulatory organization. Baker believes Dario is sincere in his safety convictions, even if critics inevitably argue that regulatory capture would entrench Anthropic's commercial position:

* **The Sincerity vs. Cynicism Matrix**: When an action falls into the quadrant that is both good for safety and advantageous for an incumbent's balance sheet, observers automatically assume cynical motivations. * **Sherman Antitrust Relief**: Frontier labs argue they cannot coordinate on safety benchmarks or pause deployments collaboratively without federal protection against illegal cartel accusations. * **FINRA Model over FAA**: A move away from heavy-handed federal agency oversight toward an industry-led, self-regulatory organization modeled after financial industry watchdogs.

[20:54](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1254s) Margin Compression, the Duopoly, and the "Mid-Curve" 4D Chess Theory

Challenging the assumption that slowing down AI helps incumbents, the hosts explore an alternate economic thesis raised by commentators: pacing the frontier might actually destroy lab margins. If Anthropic and OpenAI deliberately hold back their top-tier capabilities, fast followers like Google DeepMind, xAI's Grok, and open-source alternatives will inevitably close the gap. Instead of an entrenched duopoly maintaining pricing power, the market risks collapsing into an oligopoly of five to seven comparable models, turning frontier intelligence into a commoditized, low-margin race to the bottom.

Data cited from Ara Kharazian indicates that an observed dip in overall AI spending in August was not caused by falling user demand, but by intense price cuts and aggressive discounting. Enterprise customers are reassessing their budgets, hesitant to deploy expensive million-dollar frontier models for mundane tasks like retrieving the weather when cheaper alternatives suffice. Furthermore, enterprise inertia—rather than open-source Chinese models like DeepSeek—is shaping the market. Enterprise customers stick with OpenAI, Anthropic, or cloud giants on AWS and Azure because they demand enterprise-grade SLAs, reliability, and guaranteed uptime rather than unmanaged open-source model weights.

The conversation touches on how corporate structures play into this dynamic. Both Sam Altman and Dario Amodei lead organizations structured as Public Benefit Corporations (PBCs) or nonprofit-governed entities. Altman has reiterated in interviews that OpenAI has warned investors for a decade that profit maximization is not their primary mandate. Because of their PBC status, leadership can legally prioritize safety over short-term revenue, shield themselves from shareholder lawsuits if margins drop, and sacrifice market valuation. The hosts mock the "mid-curve" internet commentary that treats every slowdown effort as a sinister "4D chess" maneuver designed to achieve a ten-trillion-dollar monopoly, pointing out that voluntarily surrendering market share is rarely a deliberate commercial masterstroke.

[24:34](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1474s) Donald Trump Phones Jensen Huang Live at the All-In Summit

The show is interrupted by breaking news reported by Mike Isaac: Donald Trump phoned Nvidia CEO Jensen Huang live on stage during his interview at the All-In Summit. Huang took the call directly in front of the audience, placing the former president on speakerphone. Jensen used the opportunity to thank Trump for his social media posts earlier that morning, which had publicly poured cold water on AI alarmism and catastrophic doom narratives.

The hosts note that Jensen Huang has remained steadfastly "anti-doom" throughout the current technological cycle, consistently championing accelerated computing and physical build-outs over regulatory pauses. They share a laugh at the recurring tech-industry irony: a brilliant technologist capable of architecting the most sophisticated computing chips on the planet fumbling with his smartphone when trying to get a live caller onto speakerphone. With that dramatic summit moment capping the day's safety-versus-acceleration debate, the hosts wrap up the episode, reminding listeners to rate the show, subscribe to their newsletter at TBPN.com, and tune in for their next live broadcast.

## Notable quotes > [00:01](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1s) the dumbest illness you can possibly get — John Coogan

> [00:01](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1s) it's worth any amount of risk — John Coogan

> [01:27](https://www.youtube.com/watch?v=wKcf6RorKvM&t=87s) there's levels to the game of live production.

> [09:17](https://www.youtube.com/watch?v=wKcf6RorKvM&t=557s) Stop pretending you need anyone else's permission. Stop pretending antitrust law has to be suspended so you can form a cartel. Stop pretending you need a regulatory approval process that supersedes product liability. — David Sacks

> [11:39](https://www.youtube.com/watch?v=wKcf6RorKvM&t=699s) strong and smart high IQ president — Donald Trump

> [11:39](https://www.youtube.com/watch?v=wKcf6RorKvM&t=699s) Dario who is now pretending to be a perfect little angel. — Donald Trump

> [11:39](https://www.youtube.com/watch?v=wKcf6RorKvM&t=699s) Whoever wins AI wins. We are leading China and all others and will continue to do so. — Donald Trump

> [11:39](https://www.youtube.com/watch?v=wKcf6RorKvM&t=699s) Bigger than oil, gold, diamonds, or even the internet. — Donald Trump

> [14:42](https://www.youtube.com/watch?v=wKcf6RorKvM&t=882s) more good than bad, but by a lot — Donald Trump

> [14:42](https://www.youtube.com/watch?v=wKcf6RorKvM&t=882s) You can use AI for a lot of things — Donald Trump

## Mentioned John Coogan, Jordi Hays, Ben Shelton, Pierce Brosnan, Sam Altman, Dario Amodei, Elon Musk, Demis Hassabis, David Sacks, Donald Trump, Gavin Baker, Dean Ball, Greg Brockman, Xi Jinping, Nick Bostrom, Ara Kharazian, Jensen Huang, Mike Isaac, TBPN, The Wall Street Journal, The Financial Times, Axios, OpenAI, Anthropic, Google DeepMind, Meta, xAI, Nvidia, Truth Social, METR, Odd Lots, All-In Summit, Sherman Antitrust Act, Section 230, DeepSeek, Grok, ChatGPT, Claude

## 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 Weekend Chaos + The US Open’s Wild Finish 03:08 How “Pace the Frontier” Took Over the AI Industry 05:51 Dario’s 3-Step AI Slowdown Plan + US–China Coordination 09:29 David Sacks Pushes Back: Cartels, Antitrust + Regulatory Capture 11:31 Trump Rejects the AI Slowdown + “Whoever Wins AI Wins” 16:27 Third-Party AI Evaluators, Liability + Should We Keep Building Energy? 19:30 Is Slowing AI Actually Good for OpenAI & Anthropic? 24:34 Trump Calls Jensen Huang Live at the All-In Summit

[Source](https://www.youtube.com/watch?v=wKcf6RorKvM)

中文简报

特朗普拒绝 Anthropic 放缓计划，前沿实验室寻求反垄断豁免 前沿 AI 实验室正在推动协同控速协议，以确立针对侵权责任的合规注意义务，尽管自愿推迟模型部署可能让竞争对手将市场大宗商品化，从而面临破坏现有领头羊利润率的风险。 主持人深入剖析了由 Dario Amodei 发文呼吁引入第三方评估机构、反垄断豁免和国际协调所引发的“控制前沿节奏”（Pace the Frontier）AI 安全激烈辩论。节目探讨了来自 David Sacks 和 Donald Trump 的强烈反对，后者基于对抗中国时“赢得 AI 者赢得一切”的信条拒绝了放缓主张；与此同时，Gavin Baker 结合侵权责任以及缺乏第 230 条（Section 230）豁免权的背景对这一诉求进行了分析。讨论还探讨了故意放慢节奏是否会因快速跟进者将模型商品化，进而侵蚀 OpenAI 和 Anthropic 的前沿利润空间。本期节目最后报道了一条突发新闻：Donald Trump 在 All-In 峰会现场连线 Nvidia 首席执行官 Jensen Huang，表态支持继续加速推进 AI 发展。 “控制前沿节奏”（Pace the Frontier）运动在 2026 年 7 月至 9 月间迅速升温，并在 Dario Amodei 发表呼吁前沿控速的文章时达到顶峰，该倡议得到了 Sam Altman 和 Elon Musk 的高调支持。 (3:08) Dario Amodei 的放缓框架围绕三大支柱：引入诸如 METR 等第三方安全评估机构、协调政府安全阈值，以及与中国等竞争对手国家达成国际条约。 (5:51) 在相互竞争的前沿实验室之间协调模型暂停和安全阈值，会面临《谢尔曼法案》（Sherman Act）下的严重反垄断障碍，需要获得法定豁免以避免被定性为非法卡特尔。 (7:00) David Sacks 强烈反对 Amodei 和 Altman 的观点，认为前沿现有企业应当单方面暂停，而不是寻求会给较小初创公司带来沉重合规负担的监管俘获。 (9:17) Donald Trump 在 Truth Social 上明确拒绝了 AI 放缓的论调，坚持认为“赢得 AI 者赢得一切”，并警告称国内算力限制会将领导地位拱手让给中国。 (11:39)

Reading summary

Anthropic and OpenAI have spearheaded calls to coordinate a slowdown in frontier AI development, igniting an intense debate over whether institutional safety oversight constitutes responsible risk management or an anticompetitive bid for regulatory capture. Frontier Pacing Proposals and Legal Cartel Obstacles (5:51) The campaign to slow artificial intelligence accelerated following a July 28 open statement and an August 18 OpenAI paper disclosing a two-week reinforcement learning pause. The movement peaked on September 12 with an essay by Anthropic CEO Dario Amodei, titled 'We Must Pace the Frontier,' which garnered swift support from Sam Altman, Elon Musk, and Demis Hassabis. Amodei outlined a three-part framework: embedding independent third-party evaluators like Model Evaluation and Threat Research (METR) inside labs, coordinating government-supervised safety standards across democratic nations, and pursuing non-proliferation pacts with rival powers, particularly China. Implementing this proposal faces severe legal hurdles. Under the Sherman Antitrust Act, private market leaders are legally barred from colluding to restrict operational output or coordinate deployment schedules without statutory waivers. Observers note that the policy discourse has shifted away from an overarching federal agency toward an industry-funded self-regulatory model, akin to a 'FINRA for AI.' However, critics question the independence of evaluators like METR, pointing to a revolving door of personnel between safety auditing organizations and the frontier labs themselves. Regulatory Capture Allegations Versus Tort Liability Shielding (9:17) Venture capitalist David Sacks forcefully rejected Amodei and Altman's proposals, arguing that OpenAI and Anthropic already operate as a frontier duopoly. Sacks contended that if unreleased models pose genuine perils, the companies should pause internally rather than petitioning Washington for antitrust exemptions to organize a cartel. Sacks and other critics warned that imposing onerous compliance regimens—such as protracted reviews with state-sanctioned auditors—would disproportionately smother early-stage startups and application-layer developers who lack incumbent resources. Offering an alternative interpretation, investor Gavin Baker framed third-party audits as a structural response to legal risk rather than commercial malice. Because frontier AI models lack statutory immunity under Section 230 of the Communications Decency Act, labs face existential exposure to catastrophic tort lawsuits. Establishing rigorous external evaluation protocols allows frontier developers to document a legally defensible 'duty of care' against future claims of real-world harms. Geopolitical Pushback and Physical Infrastructure Chokepoints (11:39) Political and industrial leaders have rejected calls for deceleration, framing the technological race against China as an absolute imperative. Donald Trump dismissed existential risk warnings as an unfounded conspiracy that compromises national competitiveness, arguing that domestic permitting hurdles are already driving major data center projects overseas, including a Google facility in Finland. Nvidia CEO Jensen Huang similarly reinforced this growth imperative, publicly praising the rejection of doom-centric narratives during an onstage appearance at the All-In Summit. Analysts also highlighted the impracticality of attempting to regulate diffuse inputs like the electrical grid to constrain AI progress. While microchips represent a narrow, high-leverage chokepoint, energy infrastructure serves the broader economy. Diverting capital away from continuous model training into nuclear, solar, battery storage, and transmission networks delivers persistent economic dividends regardless of future safety mandates. Margin Compression and Public Benefit Corporate Governance (20:54) Market analysts challenged the premise that artificial slowdowns inevitably benefit market leaders, arguing that pausing frontier training could instead erode incumbent margins. Holding back top-tier models gives fast followers—such as Google DeepMind, xAI, and open-source models—time to bridge capability gaps. This dynamic risks turning a differentiated frontier duopoly into a commoditized, low-margin oligopoly. Data cited from Ara Kharazian noted that an August spending dip was primarily driven by aggressive price cuts rather than declining enterprise demand. Corporate governance structures explain why leadership might accept these commercial risks. Both OpenAI and Anthropic operate under Public Benefit Corporation or nonprofit charters. As Sam Altman has emphasized, these legal frameworks explicitly decouple corporate fiduciary duties from pure profit maximization, shielding executives from shareholder litigation if safety-related pauses compress operational margins or depress company valuations. The debate leaves unresolved whether frontier AI development will ultimately be bounded by legally exempted self-regulatory pacts, or whether geopolitical rivalry and physical constraints—such as power availability, chip supplies, and enterprise liability—will supersede paper regulations.

精炼阅读摘要

Anthropic 与 OpenAI 近期呼吁放缓前沿人工智能（AI）开发并引入第三方评估，引发了硅谷围绕反垄断法案豁免、监管俘获与算力地缘竞争的激烈博弈；而前总统 Donald Trump 与科技投资者 David Sacks 等人则坚决反对降速，主张在国际技术竞争中全面推进加速开发。 前沿降速倡议与三步监管架构 (3:08) 从 7 月下旬多位行业高管签署公开信开始，关于前沿 AI 降速的讨论迅速升温。OpenAI 在 8 月发表论文披露曾自愿暂停重要强化学习训练两周，随后 Anthropic 首席执行官 Dario Amodei 提出了包含三大支柱的全面框架：首先，引入嵌入实验室内部的第三方独立安全评估机构（如 METR）以核验安全防线；其次，在民主国家政府监督下协同制定前沿能力红线，并通过豁免《谢尔曼反垄断法》（Sherman Antitrust Act）允许实验室之间合法协调安全标准与发布节点；最后，寻求与威权体制国家（尤其是中国）开展类似冷战核不扩散条约式的国际步调协调。 反垄断质疑、监管俘获与侵权责任抗辩 (9:17) 这一降速提议引发了科技界的强烈反弹。David Sacks 公开批评 OpenAI 与 Anthropic 实质上构成了前沿双头垄断，指责其呼吁政府干预是在以安全之名行卡特尔垄断之实，并试图通过严苛的合规流程构建行业壁垒，从而扼杀诸如 Instinct 这类新兴初创企业。此外，METR 等评估机构被指与前沿实验室之间存在紧密的“旋转门”人事与投资关联，独立性饱受质疑。投资人 Gavin Baker 则提出了不同的法律抗辩视角：由于 AI 生成内容不像社交平台那样享有《通信规范法》第 230 条（Section 230）的免责护盾，前沿实验室推动第三方嵌入评估，核心动机之一是在未来面临毁灭性侵权诉讼前，预先确立并举证法律认可的“注意义务”（duty of care）。 商业利润压缩与公益公司法律结构 (20:54) 针对“降速属于蓄意垄断”的阴谋论观点，节目探讨了相反的商业逻辑：主动放慢顶层模型研发可能削弱两家巨头的定价权。如果 OpenAI 与 Anthropic 延后部署最强模型，Google DeepMind、xAI 的 Grok 及开源替代品将迅速缩小代差，使前沿模型沦为微利的同质化商品。Ara Kharazian 披露的数据显示，8 月份 AI 支出下滑主因在于激烈的降价折扣而非需求萎缩。由于 Sam Altman 和 Dario Amodei 领导的企业均采用公益公司（PBC）或非营利架构运作，其法律章程允许管理层优先考虑安全性而承受利润率暴跌或估值受损，并非纯粹出于利润最大化的商业算计。 加速派的反击、地缘政治与能源瓶颈 (11:39) 政界与硬件阵营对降速论调做出了强硬反击。Donald Trump 在社交平台驳斥了 AI 存亡风险警告，定调“谁赢得 AI 谁就赢得一切”，并强调美国必须全面维持对中国的绝对领先，同时他还在 All-In 峰会现场与持坚定“反毁灭论”立场的 Nvidia 首席执行官 Jensen Huang 进行了连线通话。围绕产业物理约束的讨论表明，相比于通过限制微芯片这类专用硬件来卡死 AI，全面扩建涵盖核电、太阳能与电网升级的基础设施是无风险的共识选择；即使未来前沿通用人工智能研发被叫停，由此沉淀的充沛低成本电力也将直接造福整体工业经济。 当前 AI 降速辩论的核心矛盾在于：前沿实验室究竟是出于规避无免责保护的侵权责任与真实安全风险而寻求自律协调，还是试图借政府规制构筑阻挡追赶者的卡特尔壁垒；而在大国地缘博弈和物理能源基建迅速推进的现实下，人为在法理与行政层面放慢前沿算力步伐仍面临巨大阻力与实操争议。

Detailed notes — zh

## Background 本期内容由科技播客频道 TBPN 于 2026 年 9 月 14 日发布在 YouTube 平台上，属于该频道的精选快讯栏目《Diet TBPN》。根据页面简介，TBPN 是由 John Coogan 和 Jordi Hays 主持的科技直播谈话节目，被《纽约时报》誉为“硅谷的新宠”，本期视频围绕过去 48 小时内科技界关于“放慢 AI 前沿步伐”的激烈政策辩论展开讨论。 ## TL;DR 本期节目重点梳理了 Anthropic 首席执行官 Dario Amodei 提出的“放慢前沿步伐”三步走倡议，以及 OpenAI、Elon Musk 等行业大佬的附和与跟进。主持人深入剖析了白宫官员 David Sacks 对巨头借安全之名组建卡特尔与搞“监管俘获”的猛烈抨击，并报道了特朗普全盘否定 AI 减速论、坚称“谁赢了 AI 谁就赢了一切”的强硬态度。随后，节目探讨了前沿实验室借助第三方评估寻求责任豁免的深层诉求、算力与电力基础设施的战略分流，以及突发快讯——特朗普在 All-In 峰会现场直接连线正在台上的英伟达 CEO 黄仁勋。 ## Atmosphere & style 整期节目节奏明快、充满硅谷圈内特有的机智幽默与辛辣讽刺，两位主持人展现出极佳的默契与松弛感。开场从主持人高强度健身得横纹肌溶解症的狼狈自嘲与美网抓马瞬间切入，随后无缝切换至高密度的前沿政策交锋，游刃有余地在严肃的法案博弈、地缘政治与爆笑的政商八卦之间横跳。聆听过程如同置身硅谷硬核创业者与投资人的茶余饭后对谈，信息量拉满却毫无枯燥感，尾声突发连线黄仁勋的插曲更让现场氛围达到戏剧性高潮。 ## Bulletin - [03:08](https://www.youtube.com/watch?v=wKcf6RorKvM&t=188s) Anthropic 首席执行官 Dario Amodei 联合 OpenAI、DeepMind 等前沿实验室高管发布“放慢前沿步伐”（Pace the Frontier）倡议，提出引入第三方安全审计、前沿实验室协同制定标准以及与威权国家协调监管的三步走计划。 - [09:29](https://www.youtube.com/watch?v=wKcf6RorKvM&t=569s) 白宫 AI 与加密事务负责人 David Sacks 猛烈抨击该倡议，指责 OpenAI 与 Anthropic 企图通过监管俘获暂停《谢尔曼反垄断法》来组建行业卡特尔，并利用合规流程逃避产品侵权责任。 - [11:31](https://www.youtube.com/watch?v=wKcf6RorKvM&t=691s) 唐纳德·特朗普发文全盘否定 AI 减速论，直言“谁赢了 AI 谁就赢得一切”，认为 AI 是比石油、互联网更大的经济增长引擎，并称所谓毁灭论是只会让中国受益的骗局。 - [15:57](https://www.youtube.com/watch?v=wKcf6RorKvM&t=957s) 投资人 Gavin Baker 剖析巨头引入第三方评估人员（如 METR）的核心动机在于建立法律防御盾牌，在缺乏类似互联网第 230 条豁免权的情况下，以此向陪审团证明履行了“应尽注意义务”。 - [24:34](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1474s) 在 All-In 峰会现场，特朗普致电正在台上演讲的英伟达 CEO 黄仁勋，黄仁勋公开放免提通话并当场致谢特朗普对反驳“AI 毁灭论”的表态。 ## Key takeaways - [03:08](https://www.youtube.com/watch?v=wKcf6RorKvM&t=188s) “放慢前沿步伐”风波自 7 月底多位顶尖 AI 实验室员工联名信起逐步发酵，OpenAI 期间曾实际暂停重大强化学习（RL）训练两周以评估网络安全威胁。 - [05:51](https://www.youtube.com/watch?v=wKcf6RorKvM&t=351s) Dario Amodei 在万字长文中提出放缓发展的三步走方案：第三方独立评估核验、民主国家实验室协同调控发布节奏、推动美欧与中国达成全球协调。 - [05:51](https://www.youtube.com/watch?v=wKcf6RorKvM&t=351s) 第三方评估机构 METR 的独立性饱受质疑，其与 OpenAI、Anthropic 之间频繁的“旋转门”人员流动引发了巨大的利益冲突担忧。 - [05:51](https://www.youtube.com/watch?v=wKcf6RorKvM&t=351s) 实验室协同制定标准的设想直接面临《谢尔曼反垄断法》法律壁垒，巨头联手限制节奏形同组建垄断卡特尔，本质上是在向联邦政府索要反垄断豁免权。 - [05:51](https://www.youtube.com/watch?v=wKcf6RorKvM&t=351s) 呼吁美中在 AI 上同步减速被指陷入类似冷战“核不扩散”的死局，在缺乏可信核查机制及各方争夺技术制高点的现实政治下几乎不可行。 - [09:29](https://www.youtube.com/watch?v=wKcf6RorKvM&t=569s) David Sacks 指责 Dario 与 Sam 的提议是双寡头垄断对应用层小团队的“监管俘获”，繁琐冗长的政府合规与入驻审查将扼杀底层创新。 - [11:31](https://www.youtube.com/watch?v=wKcf6RorKvM&t=691s) 特朗普强硬反击 AI 监管呼吁，主张美国必须在 AI 竞争中绝对领先中国，指责监管阻力正迫使谷歌等企业将大型数据中心外迁芬兰。 - [15:57](https://www.youtube.com/watch?v=wKcf6RorKvM&t=957s) 前沿模型当前缺乏类似互联网《通信规范法》第 230 条的法定免责护盾，拉拢外部第三方评估入驻是为未来潜在的侵权诉讼提前构建法律合规抗辩。 - [15:57](https://www.youtube.com/watch?v=wKcf6RorKvM&t=957s) 针对 AI 风险的应对思路可以从单纯的芯片算力限制转向基础设施建设，将资本分流至核能、地热与电网升级，以形成进可攻退可守的绿色正和博弈。 - [18:45](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1125s) 行业监管架构讨论正从过去类似“AI 领域的 FAA（联邦航空管理局）”的政府集权模式，转向类似 FINRA（金融业监管局）式的非政府行业自律模式。 - [20:25](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1225s) 从经济学角度看，人为压慢前沿模型迭代节奏将大幅压缩领先者代差优势，让 Mistral、xAI、DeepMind 等追赶者迅速抹平差距并加剧价格战。 - [20:25](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1225s) 企业端 AI 支出小幅下滑的主因是激烈的价格战与更理性的成本控制，企业不再愿意花巨资调用前沿大模型去处理低难度琐碎任务。 - [20:25](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1225s) OpenAI 凭借公共利益公司（PBC）的法律构架，使其在即使放缓研发导致营收与估值受损时，也能有效抵御股东违反受托责任的法律指控。 - [24:34](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1474s) 特朗普在 All-In 峰会现场连线黄仁勋，凸显出硬件算力领军企业英伟达坚定反对“末日论”，并与政界反监管声音形成合流。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=wKcf6RorKvM&t=0s) 周末近况：横纹肌溶解症与美网公开赛的“心流”插曲 科技界的周末向来波澜壮阔，两位主持人刚一开场就分享了各自的魔幻经历。其中一位主持人周末竟然直接住进了医院——起因是严重脱水加上高强度过度健身，不幸患上了横纹肌溶解症（Rhabdomyolysis）。在经历输液、服药退烧之后，他总算脱离了险境回到演播椅上。他调侃称这是“人类能得的最愚蠢的疾病”，完全是因为不喝水硬练造成的自找苦吃，并开玩笑说：“如果 AI 哪怕只有 1% 的机会能找到预防横纹肌溶解的方法，冒任何风险推行 AI 都是值得的！”当医生问他要不要开一张请假条时，他直接回绝，笑言他们做播客以来“从未因生病缺席过任何一期节目”，反正只要坐在椅子上聊天就能开播，只不过大病初愈可能反应会稍显迟钝。 随后话题转向了登上《华尔街日报》封面的美网公开赛跌宕起伏的决赛。比赛中出现了一个极具戏剧性的名场面：美国选手本·谢尔顿（Ben Shelton）遗憾落败，而获胜选手在打出制胜分赢得整场比赛后，居然毫无察觉，直接转身回到发球线摆好姿势准备打下一个球；直到环顾四周发现全场观众起立欢呼狂叫，他才抬头看了一眼记分牌，恍然大悟自己已经赢了。主持人感叹这才是真正的“心流状态”（Flow State），并顺便揶揄了一把硅谷的风险投资家（VC）——风投就该学学这种专注精神，脑子里永远只想着下一笔投资，而不是天天在社交媒体上大搞毫无意义的“胜利巡游”（Victory Laps），连自己的种子轮项目何时敲钟上市都用不着去关心。此外，转播团队在倒计时剩下 07 秒时，神来之笔地把特写镜头切给了看台上的“初代 007”皮尔斯·布鲁斯南（Pierce Brosnan），展现了顶级现场导播的临场创造力。 --- ### [03:08](https://www.youtube.com/watch?v=wKcf6RorKvM&t=188s) “放慢前沿步伐”发酵史：从公开信到达里奥万字长文 闲聊过后，讨论迅速切入过去 48 小时席卷整个人工智能圈的核心风暴：“放慢前沿步伐”（Pacing the Frontier）的倡议。主持人系统梳理了这场政策游说风暴的时间线：这一切的序幕其实在当年的 7 月 28 日就已经拉开，当时一份呼吁放慢 AI 前沿推进步伐的公开信发布，签署人包括来自 OpenAI、Anthropic、Google DeepMind 和 Meta 等顶尖实验室的员工与高管；同一天，OpenAI 首席执行官山姆·奥特曼（Sam Altman）公开表态称“我们可能不得不调整前沿 AI 的发展节奏”；两天后，OpenAI 向 Axios 证实，公司不仅参与撰写了请愿书的措辞，奥特曼还就发展节奏放缓一事与白宫高级官员展开了深入探讨。 进入 8 月，这场行动进一步从口头呼吁走向实践：8 月 18 日，OpenAI 发表了题为《网络关键能力时代的模型开发步调控制》的论文，并透露公司实际上已放慢了 Scaling 进程，并实施了为期两周的强化学习（RL）大规模训练中断；8 月 31 日，Anthropic 官方正式跟进发文，明确探讨“放慢前沿步伐”并呼吁建立协调机制。到了 9 月 12 日周六清晨 7 点（太平洋时间），Anthropic 首席执行官达里奥·阿莫代伊（Dario Amodei）在 X 上投下重磅炸弹，发布了长篇论文《我们必须放慢前沿步伐》（We Must Pace the Frontier）。随后，埃隆·马斯克（Elon Musk）在早晨 8 点率先点赞支持，山姆·奥特曼于 9 点 30 分紧随其后力挺，前 DeepMind 掌舵人德米斯·哈萨比斯（Demis Hassabis）也在下午 4 点表示大体认同其核心原则，瞬间演变成一场科技大佬集体背书的行业运动。 - **7 月 28 日**：多家前沿实验室员工联合签署“放慢前沿步伐”公开声明，奥特曼公开附和。 - **7 月 30 日**：OpenAI 证实参与起草请愿书，且奥特曼已与白宫官员展开闭门沟通。 - **8 月 18 日**：OpenAI 官宣曾暂停重大强化学习（RL）训练达两周，放缓扩展速度以评估高危网络能力。 - **8 月 31 日**：Anthropic 正式呼吁构建跨机构的步调协调监管机制。 - **9 月 12 日**：达里奥发表万字长文，马斯克、奥特曼在数小时内接连背书。 - **9 月 13 日**：大卫·萨克斯（David Sacks）公开质疑；《金融时报》爆出唐纳德·特朗普明确拒绝科技大佬放慢 AI 的呼吁。 --- ### [05:51](https://www.youtube.com/watch?v=wKcf6RorKvM&t=351s) 达里奥的三步慢速计划：第三方审计、反垄断困境与中美协调 达里奥在文章中提出了具体的“三步走倡议”，但每一个环节都在科技界和法律界引发了尖锐的交锋与质疑： ```mermaid flowchart TD accTitle: Dario's Three-Step AI Slowdown Proposal accDescr: 达里奥提出的AI减速三步走计划及其引发的行业争议 subgraph Step1 ["第一步：第三方独立评估"] A["开放实验室底层权限"] --> B["评估机构（如 METR）核验安全"] B -. 争议点.-> C["人员高度重叠，存在'旋转门'利益冲突"] end subgraph Step2 ["第二步：民主国家实验室协同"] D["顶尖实验室与政府制定安全标准"] --> E["统一调控训练与发布节奏"] E -. 法律壁垒.-> F["违反《谢尔曼反垄断法》，涉嫌垄断卡特尔"] end subgraph Step3 ["第三步：与威权政府协调"] G["美欧等国推动与中国达成协议"] --> H["全球同步限制模型能力"] H -. 地缘博弈.-> I["核不扩散困境：均势难题与领先者诉求"] end ``` 第一步是允许独立的第三方评估机构全面进入 AI 实验室，审查并验证其安全实践。达里奥在文中专门点名了评估机构 METR（模型评估与威胁研究机构）。然而，这一提议立刻被质疑“缺乏真正独立性”：METR 与 OpenAI、Anthropic 之间存在着极度频繁的“旋转门”现象，大量研究员在两套班子间来回跳槽。这绝非传统审计行业中普华永道或安永那样与被审计方毫无瓜葛的中立角色，而是一群在 AI 意识形态圈内深陷其中的局内人，这虽然保证了专业度，却也引发了巨大的利益冲突质疑。 第二步是民主国家的前沿实验室在政府协助下联合制定安全标准与节奏。智库学者迪恩·鲍尔（Dean Ball）等人立刻指出这一举措面临极其严峻的法律死胡同——《谢尔曼反垄断法》（Sherman Antitrust Act）。在市场经济中，顶尖巨头联合宣布限制产能或放慢节奏，历来被视作非法卡特尔行为。这好比美国航空、达美航空与美联航串通一气，借口“飞行安全”共同削减航班班次，表面上冠冕堂皇，实质上却大幅抬高航司利润率并重创消费者权益。达里奥等人深知这一点，因此其实是在变相向联邦政府索要“反垄断豁免权”。 最令人瞠目结舌的是第三步：达里奥呼吁美国及盟友必须尝试与威权政府（尤其是中国）协调一致，共同为 AI 踩刹车。但这立即陷入了类似冷战时期“核不扩散协议”的死局：中国正在拼命追赶，若此时要求双方冻结在同一发展水准，谁来充当可信的核查机制？对于矢志保持技术绝对领先地位的美国决策者而言，要求对手只保留“一半的核武库”并在博弈中甘心落后，在现实政治中几乎是一纸空谈。 ### [09:17](https://www.youtube.com/watch?v=wKcf6RorKvM&t=557s) 大卫·萨克斯猛烈反击：卡特尔、反垄断与监管俘获 在达里奥·阿莫代伊（Dario Amodei）发表“协同放慢前沿模型研发步伐”（Pace the Frontier）的呼吁后，山姆·阿尔特曼（Sam Altman）和戴密斯·哈萨比斯（Demis Hassabis）相继表示赞同。然而，网络舆论场上随即涌现出大量质疑：这些所谓的第三方评估机构究竟是谁？为什么非要等华盛顿立法监管？如果 Anthropic 真的认为模型发展过快，为何不率先主动减速？白宫 AI 与加密货币事务负责人大卫·萨克斯（David Sacks）在 X 上发布长文，对这种行业巨头之间的“默契减速”进行了毫不留情的抨击。 萨克斯直言不讳地指出，无论从市场份额、营收增长还是模型能力来看，OpenAI 和 Anthropic 已经实质上构成了前沿智能的双寡头（Duopoly）。双方甚至曾宣称，由于模型的递归自我改进（recursive self-improvement），其领先优势还在持续扩大。萨克斯反讽道：“如果实验室里未发布的模型真的吓人到让你们觉得必须减速，那我完全支持你们做出负责任的商业决策。但别再假装你们需要别人的许可了！”他严词斥责这种行为是以安全为幌子，要求暂停反垄断法来组建行业“卡特尔”，企图利用监管审批流程凌驾于产品责任法之上。此外，萨克斯特别点名了评估机构 METR（Model Evaluation and Threat Research），质疑其与 Anthropic 的投资者和内部员工盘根错节，根本算不上真正独立，更不应该让这些机构反过来去监管那些根本不在前沿层级的小型竞争对手。 主持人指出，如果针对前沿安全设定的合规流程过于繁重，很可能引发严重的“监管俘获”（Regulatory Capture）。例如像 Instinct 这样只有小团队的应用层初创公司，哪怕自己不训练前沿大模型，一旦被强加冗长的第三方入驻审查和政府审批流程，任何现象级的消费级创新产品都将被扼杀在摇篮里。不过也有人补充提到，目前监管的核心矛头基本聚焦于前沿大模型，格雷格（Greg Brockman）当天早上在彭博社《Odd Lots》播客上也明确表态，不应该对应用层以及开发者在个人电脑上私下微调或训练模型设置限制。 - **双寡头垄断指控**：萨克斯批评 Dario 与 Sam 在前沿智能领域握有绝对主导权，减速倡议本质上是巨头试图通过政策巩固壁垒。 - **利益冲突质疑**：第三方评测机构 METR 被指与 Anthropic 关系密切，其独立性难以服众。 - **免责与卡特尔风险**：批评者认为前沿实验室企图借由政府背书逃避产品侵权责任，并在反垄断法外建立事实上的联盟。 ```mermaid flowchart TD accTitle: 监管俘获与卡特尔争议逻辑 accDescr: 展示前沿巨头呼吁减速到大卫·萨克斯指出潜在监管俘获风险的演变逻辑 subgraph FrontierLabs ["前沿实验室倡议"] A["Dario & Sam 呼吁协同放缓<br/>(Pace the Frontier)"] B["引入第三方评估机构<br/>(如 METR 进驻审查)"] A --> B end subgraph SacksCriticism ["大卫·萨克斯的抨击"] C["事实上的前沿双寡头垄断<br/>(Duopoly)"] D["暂停反垄断法以组建卡特尔"] E["借合规流程规避产品责任法"] B -.-> C C --> D C --> E end subgraph DownstreamImpact ["下游与初创公司风险"] F["繁琐合规审查与政府拖延"] G["扼杀应用层创新<br/>(如小团队产品被拖垮)"] E --> F F --> G end ``` --- ### [11:31](https://www.youtube.com/watch?v=wKcf6RorKvM&t=691s) 特朗普炮轰 AI 减速论：谁赢了 AI，谁就赢得了一切 唐纳德·特朗普在周末及播客录制当天早晨通过社交媒体接连发布极具个人风格的重磅帖文，对 AI“减速论”进行了全盘否定。特朗普宣称，AI 唯一需要的控制器或所谓“护栏”，就是一个“强大、聪明、高智商的总统”，而美国在这方面“绰绰有余”。他嘲讽达里奥·阿莫代伊如今是在“装成一个完美的小天使”，并坚称自己的政府拥有强大的刑事与监管权力来遏制任何潜在的恶意行为。 在特朗普看来，当下针对 AI 和数据中心的种种恐慌背后存在着一场“恶心的阴谋”，而唯一能从中坐收渔利的就是中国。他写道：“谁赢了 AI，谁就赢了一切（Whoever wins AI wins）。我们在领先中国及其他所有人，而且必须保持领先。”他痛斥所谓“AI 毁灭人类、接管世界”的灭世论调纯属一场彻头彻尾的骗局，将其与当年的“通俄门”调查以及两次弹劾案相提并论。特朗普还声称，中国最高领导人刚刚表态绝不会阻碍 AI 的发展，反观美国自身，谷歌却因为在本土遭遇繁琐严苛的审批许可难题，转而计划在芬兰建立大型数据中心设施。 谈及 AI 的宏大前景时，特朗普将其定性为人类历史上最庞大的经济增长引擎，“规模将超越石油、黄金、钻石，甚至是互联网”。演播室的主持人看到这一串排比忍不住捧腹大笑，调侃在蒸汽机、晶体管和互联网这种改变历史的发明旁边，居然单独拎出了“钻石”，戏称如果华尔街对冲基金花钱提前一分钟截获这条推文，恐怕会一脸茫然地去疯狂做多钻石股票。随后，节目切入了一段记者在现场追问特朗普是否亲自使用过 AI 的采访切片；特朗普在被连问两次后含糊其辞地打起太极，主持人笑称特朗普可能压根不直接用，但其竞选团队和支持者显然在大批量制造并发布充斥着“AI 垃圾（AI slop）”的模因视频。 --- ### [15:57](https://www.youtube.com/watch?v=wKcf6RorKvM&t=957s) 责任豁免护盾、加文·贝克点评与算力 vs 电力之争 针对过去 24 小时围绕 AI 减速引发的激烈风暴，著名科技投资人加文·贝克（Gavin Baker）发表了提纲挈领的深度复盘。与外界预期的不同，埃隆·马斯克（Elon Musk）此次并没有站在 Dario 的对立面，而是明确支持对 AI 施加审慎约束。马斯克回顾了自己早在 2014 年就曾与牛津哲学家尼克·博斯特罗姆（Nick Bostrom）探讨 AI 安全问题的历程，尽管他过去在“末日论者”与“全力推进算力竞争”之间反复摇摆，但其内心深处始终对不受控的通用智能怀有戒心。 加文·贝克指出，过去这狂暴的一昼夜里，真正具有实质性意义的唯一增量事实，是 OpenAI 和 Anthropic 确认将接纳来自外部机构的嵌入式第三方评估人员（Dario 提出了 METR 作为候选）。贝克分析称，这一举动背后的真正逻辑在于法律风控：当前法律并没有为模型输出提供类似于互联网时代的《通信规范法》第 230 条（Section 230）责任豁免盾牌。如果缺乏法律豁免，未来模型一旦引发灾难性后果或侵权诉讼，AI 公司必须在法庭上向陪审团证明自己已经履行了“应尽注意义务”（Duty of Care）。互联网行业如果当年没有第 230 条的保护早已大批破产，因此前沿巨头拉拢第三方机构做背书，本质上是在为将来的诉讼建立护城河。 贝克的分析随即引发了主持人对 AI 发展要素的深层思考。目前安全阵营的主要手段是通过半导体供应链施加物理约束——限制芯片的生产与流转。但主持人提出反思：既然人们担忧前沿超级智能突飞猛进，为何不去换一个思路，把过剩的资本和顶尖人才分流到基础能源设施的建设中？ - **能源竞赛而非单纯算力竞赛**：与其冒着安全风险在数据中心里盲目“极限扩张（Blitzscaling）”训练不可控的模型，不如将资本投入到核能并网、地热、风电与先进太阳能电站的建设。 - **重振美国能源供应链**：通过重构电池储能与电网承载能力，不仅可以化解民众对高昂电价的怨气，还能重建制造业根基。 - **退可守进可攻的战略**：即便未来证实 AGI 或 ASI 因为安全风险过高而无法全面落地，大规模建设至少让整个社会换来了充沛且廉价的清洁能源，这无论对技术派还是安全派而言，都将是一笔毫无争议的正和博弈。 ### [18:45](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1125s) 供应链深处的监管之争与达里奥的“放缓方案” 讨论首先切入到针对 AI 基础设施的极端监管设想。两位主持人探讨了是否可能因为 AI 安全理由而对美国电网建设或半导体上游极其细分的供应链进行干预。虽然芯片本身的使用场景较为狭窄（主要用于数据中心、流媒体或 AI 计算），而电网电力则具有高度泛用性，但有人调侃道，如果为了给 AI 踩刹车，监管者甚至可能把安全阀门伸向极深的供应链隐秘角落，比如去规管韩国一家既做智能马桶又生产半导体薄膜的公司。 随后，节目引述了投资人 Gavin Baker 对 Anthropic 首席执行官达里奥·阿莫代伊（Dario Amodei）周末提出的方案所做的系统总结。Baker 认为这是达里奥迄今为止最激进的监管主张之一，涵盖了多层次的管控构想： ```mermaid flowchart TD accTitle: Dario Amodei AI Regulatory Proposal Hierarchy accDescr: Diagram showing the layers of Dario Amodei's proposed AI regulation structure from national oversight to international treaties. subgraph National["国内与前沿实验室层级"] A["第三方评估员派驻机构"] --> B["能力与算力阈值之上的国家监管体制"] B --> C["谢尔曼反托拉斯法豁免<br/>(允许 Anthropic 与 OpenAI 安全协同)"] end subgraph International["国际多边框架"] D["民主国家间广泛监管协议"] --> E["对华算力蒸馏与硬件更严格限制"] E --> F["涵盖中美的前沿大模型国际协定"] end National --> International ``` - **派驻第三方评估员**：在前沿实验室内部嵌入第三方评估机制，实时审核前沿模型的安全性。 - **算力与能力阈值监管**：针对超越特定算力与能力门槛的模型，建立专门的国家级监管体制。 - **《谢尔曼反托拉斯法》（Sherman Act）豁免**：允许 Anthropic、OpenAI 等前沿实验室在免受反垄断诉讼恐惧的前提下，进行合法的安全协调与信息互通。 - **国际民主同盟与对华限制**：建立民主国家间的监管同盟，对中国实施更严厉的算力蒸馏限制，并在建立广泛的全球监管体制之前，先行构筑包含中国的国际监管框架。 - **监管机构的演变思路**：达里奥此前曾倡议设立“AI 领域的联邦航空管理局（FAA）”，而如今行业讨论的风向则转向了非政府性质但侧重行业的“AI 领域的金融业监管局（FINRA）”。 ### [20:25](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1225s) 放缓前沿步伐究竟是利好还是削弱前沿巨头？ 针对达里奥与萨姆·奥尔特曼（Sam Altman）力推监管的动机，舆论场存在截然相反的两种叙事。一部分批评者认为，这是精心包装的监管俘获（Regulatory Capture），如果某项行动恰好落在“既对业务有利，又对安全性有利”的象限，外界就会指责其虚伪，认为一切说辞都是为了保住自身垄断地位。然而，另一种截然相反的经济学视角指出，人为压慢前沿模型的迭代速度（Pacing the Frontier）反而会严重压缩领先者的利润率。因为一旦前沿实验室放慢脚步，Mistral、Google DeepMind、xAI（Grok）等后发追赶者就会迅速缩小代差，原本 OpenAI 与 Anthropic 之间的双头垄断将被打破，演变成四到七家巨头并存的寡头红海。 近期分析师数据也印证了价格竞争的残酷性。8 月份 AI 相关支出的微跌并非因为实际使用量减少，而是源于激烈的价格战与折扣。企业在成本控制上变得更加审慎，不再愿意花高昂的调用费让“百万美元级大模型”去执行查天气之类的琐碎任务。与此同时，目前给前沿实验室带来定价压力的并非 DeepSeek 等开源模型，绝大多数注重稳定性的企业出于 SLA 服务保障、系统正常运行时间和运维复杂度的考虑，仍会优先选择部署在 AWS 或 Azure 上的 OpenAI、Anthropic 或 Google 模型。Sam Altman 此前在接受《财富》（Fortune）杂志采访时曾强调，作为公共利益公司（Public Benefit Corporation, PBC），他们在过去十年里一直向投资人重申公司非纯粹追逐股东利益的属性；即使放缓模型研发导致收入增长停滞或估值缩水，投资人也无法因此起诉他们违背受托责任。 ### [24:34](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1474s) 突发快讯：特朗普在 All-In 峰会现场连线黄仁勋 节目尾声迎来突发新闻：在正在举行的 All-In 峰会（All-In Summit）现场，美国前总统唐纳德·特朗普（Donald Trump）直接致电正在台上接受采访的英伟达 CEO 黄仁勋（Jensen Huang）。《纽约时报》记者 Mike Isaac 及多名现场参会者证实了这一插曲。 黄仁勋在台上将手机切换至免提扬声器，并幽默地对电话那头说道：“总统先生，您现在正在向全世界发话。”黄仁勋当场对特朗普当天上午在社交媒体上泼冷水、反驳“AI 毁灭论”的表态表示感谢。两位主持人打趣道，黄仁勋在技术上无所不能，能架构出全球最复杂的芯片与计算集群，但在台上却常常被各种免提通话等现场小插曲逗乐；而英伟达在整轮 AI 发展周期中始终坚定站在“反末日论（Anti-doom）”立场。随着这则现场花絮的爆发，主持人宣布结束当天节目，并提示观众关注次日早间 11 点的直播及 Daily TBPN 邮件通讯。 ## Notable quotes > [00:00](https://www.youtube.com/watch?v=wKcf6RorKvM&t=0s) 如果 AI 哪怕只有 1% 的机会能找到预防横纹肌溶解的方法，冒任何风险推行 AI 都是值得的！ > [03:08](https://www.youtube.com/watch?v=wKcf6RorKvM&t=188s) 我们可能不得不调整前沿 AI 的发展节奏 — Sam Altman > [09:17](https://www.youtube.com/watch?v=wKcf6RorKvM&t=557s) 如果实验室里未发布的模型真的吓人到让你们觉得必须减速，那我完全支持你们做出负责任的商业决策。但别再假装你们需要别人的许可了！ — David Sacks > [11:31](https://www.youtube.com/watch?v=wKcf6RorKvM&t=691s) 谁赢了 AI，谁就赢了一切（Whoever wins AI wins）。我们在领先中国及其他所有人，而且必须保持领先。 — Donald Trump > [11:31](https://www.youtube.com/watch?v=wKcf6RorKvM&t=691s) 装成一个完美的小天使 — Donald Trump > [11:31](https://www.youtube.com/watch?v=wKcf6RorKvM&t=691s) 规模将超越石油、黄金、钻石，甚至是互联网 — Donald Trump > [24:34](https://www.youtube.com/watch?v=wKcf6RorKvM&t=1474s) 总统先生，您现在正在向全世界发话。 — 黄仁勋 ## Mentioned TBPN, John Coogan, Jordi Hays, Ben Shelton, Pierce Brosnan, OpenAI, Anthropic, Google DeepMind, Meta, Sam Altman, Dario Amodei, Demis Hassabis, Elon Musk, David Sacks, Donald Trump, METR, Dean Ball, Instinct, Greg Brockman, Nick Bostrom, Gavin Baker, Mistral, xAI, DeepSeek, Jensen Huang, Mike Isaac, All-In Summit, WSJ, Axios, Odd Lots, Fortune, Section 230, Sherman Antitrust Act ## 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 Weekend Chaos + The US Open’s Wild Finish 03:08 How “Pace the Frontier” Took Over the AI Industry 05:51 Dario’s 3-Step AI Slowdown Plan + US–China Coordination 09:29 David Sacks Pushes Back: Cartels, Antitrust + Regulatory Capture 11:31 Trump Rejects the AI Slowdown + “Whoever Wins AI Wins” 16:27 Third-Party AI Evaluators, Liability + Should We Keep Building Energy? 19:30 Is Slowing AI Actually Good for OpenAI & Anthropic? 24:34 Trump Calls Jensen Huang Live at the All-In Summit [Source](https://www.youtube.com/watch?v=wKcf6RorKvM)

Source quotes

[0:00] 如果 AI 哪怕只有 1% 的机会能找到预防横纹肌溶解的方法，冒任何风险推行 AI 都是值得的！ [3:08] Sam Altman: 我们可能不得不调整前沿 AI 的发展节奏 [9:17] David Sacks: 如果实验室里未发布的模型真的吓人到让你们觉得必须减速，那我完全支持你们做出负责任的商业决策。但别再假装你们需要别人的许可了！ [11:31] Donald Trump: 谁赢了 AI，谁就赢了一切（Whoever wins AI wins）。我们在领先中国及其他所有人，而且必须保持领先。 [11:31] Donald Trump: 装成一个完美的小天使 [11:31] Donald Trump: 规模将超越石油、黄金、钻石，甚至是互联网 [24:34] 黄仁勋: 总统先生，您现在正在向全世界发话。

Source quotes

[0:01] John Coogan: the dumbest illness you can possibly get [0:01] John Coogan: it's worth any amount of risk [1:27] there's levels to the game of live production. [9:17] David Sacks: Stop pretending you need anyone else's permission. Stop pretending antitrust law has to be suspended so you can form a cartel. Stop pretending you need a regulatory approval process that supersedes product liability. [11:39] Donald Trump: strong and smart high IQ president [11:39] Donald Trump: Dario who is now pretending to be a perfect little angel. [11:39] Donald Trump: Whoever wins AI wins. We are leading China and all others and will continue to do so. [11:39] Donald Trump: Bigger than oil, gold, diamonds, or even the internet. [14:42] Donald Trump: more good than bad, but by a lot [14:42] Donald Trump: You can use AI for a lot of things

Original transcript

[0:01] We're back. Did anything happen over the [0:03] weekend in tech or was it kind of [0:05] mellow? [0:06] >> Uh, it was a crazy weekend. There was a [0:07] lot of news, a lot of back and forth [0:08] about pacing the AI debate. I was in the [0:11] hospital because I got rabbo. It's [0:13] brutal, but I am recovering fortunately. [0:16] >> Explain because it sounds really scary [0:19] and it is bad. [0:20] >> It can be very bad. Fortunately, I I I [0:21] did okay. I got an IV, got some [0:23] medicine, broke a fever, got through it, [0:26] and I'm I'm healing up. Uh it it happens [0:29] and it's the dumbest dumbest uh illness [0:31] you can possibly get. It comes from uh [0:33] working out too much and not being [0:35] hydrated. Very avoidable. Hopefully AI [0:38] can find a cure. If there's even a 1% [0:40] chance that AI can find a prevention for [0:42] Rabdo, I think it's worth any amount of [0:44] risk. [0:45] >> Pushing forward. [0:45] >> Yeah, we we must push forward. Uh no, [0:47] lots of debate about how fast we should [0:49] push forward. [0:50] >> We have never missed an episode for [0:53] being sick. [0:53] >> Yeah. And I realized that because [0:55] someone's asking me the other day, "What [0:57] do you guys do when you're sick?" We [0:59] just we keep going. [1:00] >> I mean, I did talk to the doctor. They [1:02] said, "Do you want a do you want a [1:03] doctor's note for work?" And I was like, [1:05] "No." [laughter] I kind of laughed. Uh [1:08] but uh I do think that uh like this job [1:12] can be done while sick because you're [1:14] just sitting in a chair having a [1:15] conversation. But I might be a little [1:16] bit slow today, so bear with me if I'm [1:19] not the sharpest on any of the [1:20] questions. Before we get into the pacing [1:23] AI debate, uh, what happened with the US [1:25] Open? Give me a little tennis update. [1:27] What happened? [1:28] >> Um, [1:28] >> it's on the cover of the Wall Street [1:29] Journal, so you know, it's important [1:31] news. Tennis's US Open climaxes and back [1:33] and forth duel, but there was something [1:35] interesting twist. [1:36] >> Ben Shelton, the American, lost. [1:39] >> Okay. Uh and uh what was notable was [1:43] that when he actually [1:46] hit the winning shot and uh and won [1:50] >> uh he went back and was getting ready [1:53] for the next uh for the next um point. [1:57] >> Yeah. And uh he starts looking around [1:59] and everyone is st you can see people in [2:01] the crowd are standing up uh and and and [2:03] cheering and he slowly realizes that he [2:06] won [laughter] [2:07] by checking finally look at the [2:08] scoreboard and yeah it just became very [2:10] obvious that he was so focused on the [2:12] next point he was not actually not [2:14] looking at the scoreboard at all. It was [2:16] just like flow state just just playing. [2:18] >> That's awesome. [2:19] >> Um and I'm sure a lot of venture [2:20] capitalists out there were seeing that [2:22] and they're like yeah all that matters [2:24] is your next investment. You shouldn't [2:25] even know when one of your companies, [2:26] your seed companies, yeah, [2:28] >> are gonna IPO. You should be so focused [2:30] on [2:31] >> I think a lot of people that don't like [2:32] VC Victory Labs would appreciate that, [2:34] too. [laughter] [2:34] >> It's like, we get it. There was an [2:36] acquisition. Uh, let's focus on what [2:38] you're doing right now. Um, I also like [2:41] the cameraman work. I saw a clip of the [2:43] scoreboard was counting down. There was [2:45] 07 [2:47] seconds left. What who do they cut to? [2:50] Pierce Browman 007 himself. I thought it [2:52] was nice camera work. shows you that [2:53] there's levels to the game of live [2:55] production. [2:55] >> Yeah. [2:56] >> You know, you got to those little [2:58] touches clearly that was just spur of [2:59] the moment. That's not that's not [3:01] something that's scripted. That's just [3:03] someone being creative in the moment [3:05] with the tools that they have. The [3:06] switcher basically. Uh anyway, the the [3:09] pacing AI debate. I wanted to start with [3:11] some history here and then we can go [3:12] through the timeline of what actually [3:13] happened uh since we got off stream [3:17] Friday. Then all the news breaks of [3:19] course. But the pacing the frontier goes [3:22] back to July 28th of this year. So there [3:24] was a there was a uh pacing the frontier [3:27] statement an open letter. It was signed [3:30] by employees and leaders from OpenAI, [3:32] Anthropic, DeepMind, Meta and a bunch of [3:34] other places talking about this idea of [3:37] pacing the frontier. Uh then uh the same [3:40] day Sam Alman publicly said we may have [3:43] to pace the rate of AI development. Two [3:45] days later, OpenAI uh tells Axios that [3:49] OpenAI shaped the position, the [3:51] petition's language and Altman had [3:53] discussed pacing with White House [3:54] officials. So stuff was happening in DC. [3:57] This was back in basically the very [3:59] beginning of August, very end of July. [4:01] Then August 18th, OpenAI publishes [4:04] something called pacing model [4:05] development in an era of cyber critical [4:08] capabilities. OpenAI said it actually [4:10] slowed scaling and paused a major RL [4:13] run. there was that two-week training [4:15] pause that happened uh on August 31st. [4:18] Anthropic uh official anthropic post [4:20] explicitly discusses pacing the [4:22] frontier. So the idea is spreading and [4:24] calls for coordinating pacing [4:25] mechanisms. And then on September 12th [4:28] was a se it was actually Saturday [4:29] morning that that Daria dropped the [4:31] essay. We must pace the frontier turn [4:33] the concept into a much bigger public [4:36] anthropic campaign uh that everyone's [4:38] talking about today. So, the roll out of [4:40] this was pretty pretty quick. And [4:41] there's there's a few key things from [4:43] Daario's essay. It hit the timeline at 7 [4:46] a.m. Pacific. He was up early posting. [4:49] Perfect time to get a lot of likes on a [4:51] banger. You got to feel something. No. [4:53] Everyone else, there's no fundraising [4:55] news happening. Drop your essay early [4:57] Saturday morning. Let people digest it [4:58] with their morning coffee. Elon was [5:00] clearly engaging. By 8 a.m., Elon [5:02] endorses it. 9:30 a.m. Sam endorses it. [5:06] uh Saturday 4 pm. Deis generally agrees [5:08] with Daario's point uh and argues that [5:11] it needs some work. Remember Daario put [5:13] out something pretty similar talking [5:15] about uh pacing and AI development and [5:18] the need for oversight generally uh [5:21] months ago before transitioning out of [5:23] his role as CEO of DeepMind and stepping [5:26] up to I think chairman. Um then later [5:29] David Saxs on Saturday but at 8:15 has [5:32] had time to process everything. Uh, he [5:34] asks some questions about Dario's post, [5:36] goes back and forth. Sunday morning, the [5:39] Financial Times reports that Donald [5:40] Trump has rejected calls from tech [5:42] bosses for an AI slowdown. Later that [5:45] day, Sunday morning, Gavin Baker rang [5:47] round rounds up the weekend. So, we can [5:48] kind of go through some of this and go [5:49] through this post. Uh, but what actually [5:52] happened in uh Dario's essay, we must [5:54] pace the frontier. He has the three-step [5:56] plan on how to how he proposes to [5:59] >> number three will shock you. basically [6:02] uh number one give independent [6:04] third-party evaluators access to AI [6:07] companies so they can verify safety [6:09] practices. Uh he calls out meter which [6:12] drew a lot of attention because people [6:14] were debating how independent meter is. [6:16] There's some meter employees who have [6:17] worked at open AI have worked at [6:19] anthropic. There's people that work at [6:21] Anthropic who worked at meter like it's [6:23] it's all a uh very much a a revolving [6:26] door with these AI safety groups and [6:29] then the AI labs. you go back and forth [6:31] depending on what you're interested in. [6:32] So, not as independent as like, oh yeah, [6:35] we're we're we're going with a an [6:37] accounting firm who, you know, doesn't [6:39] really have an opinion about uh the like [6:42] the AI future. These are all people who [6:44] are deeply embedded in the discourse, [6:46] which could be a huge benefit because [6:48] they take the stuff really seriously. Uh [6:50] could also lead to some what people [6:51] flagged as like conflicts potentially, [6:53] but uh still uh some some optimism from [6:57] the community around what that would [6:59] look like. Two is Frontier Labs in [7:01] democratic companies should work [7:02] together with the with the government's [7:04] help to establish AI safety standards. [7:07] And this was the one that people were [7:09] going back and forth on. Do the labs [7:11] actually need government approval from [7:12] from this? Uh Dean Ball was talking a [7:15] little bit about the Sherman Antitrust [7:18] Act. There are very clear rules about [7:21] companies agreeing to do really anything [7:23] in concert with each other. Um, this is [7:26] the number one reason that you don't [7:28] want monopolies and cartels forming, uh, [7:30] is that if all of a sudden, you know, [7:33] oh, uh, you know, American Airlines and [7:36] Delta and United all say, yeah, we, you [7:38] know, for safety reasons, we should have [7:40] fewer flights. It's like, okay, well, [7:42] what would that do to the market? [7:43] Probably drive up their margins because [7:45] there's less flights, charge more for [7:47] them. They have fixed costs. So you [7:49] could very clearly see that collusion in [7:51] that market would lead to more profits [7:54] for them and more harm for the consumer. [7:57] That's where the Sherman Antitrust Act [7:59] came from. We flagged this a week or two [8:01] ago that like even even though it seems [8:03] like oh well if Demis and Daario and Sam [8:06] and Elon all agree like this should just [8:08] happen, that's not a problem, there are [8:10] actual legal and federal regulatory [8:13] rules that might need to be adjusted or [8:15] waved for something like this to happen. [8:17] Um, so that's I think that's why he's [8:19] he's calling for that specifically. He's [8:21] saying we need an exception to the rules [8:23] so that we can actually do this. Um, [8:25] because it's great to all write blog [8:26] posts that we loosely agree on, but to [8:28] have anything formal, we need government [8:30] approval. And then third, this is the [8:32] one that will shock you apparently. [8:34] [laughter] Uh, the US and other [8:35] democratic countries should try and [8:37] coordinate with authoritarian [8:38] governments. So he wants the US and [8:40] China to coordinate on AI pacing. Of [8:44] course, China still wants to catch up [8:46] and there's some tricks there is like [8:49] what will happen in equilibrium if both [8:51] countries have the exact same [8:53] capabilities. Is that the best outcome [8:54] or do you always want to be in the lead? [8:56] If you always want to be in the lead, [8:57] it's very hard to negotiate with someone [8:59] and say, "Oh, well, you know, we'll uh [9:02] why don't you just have half as many [9:03] nuclear weapons as we have and like [9:05] that'll be great. Like we'll pace you." [9:06] And so you you go back to nuclear [9:08] non-prololiferation and all the [9:10] difficult debates there. We obviously [9:11] wound up with a multi-polar world with [9:14] everyone wanting at least some [9:15] capabilities there. So Elon chimed in [9:17] agreeing with Daario. Then Sam and Demis [9:19] also agreed. Uh the timeline on the [9:21] other hand had some questions and [9:22] concerns about Daria's post. Who will [9:24] who will these third party evaluators [9:26] be? Why wait for Washington to regulate [9:28] the frontier? If Anthropic believes AI [9:30] is advancing too quickly, why not simply [9:32] slow down on its own? David Sax raised [9:35] some questions on X. And we can go to [9:38] David Sax's post which I believe here is [9:41] in the timeline. He said Daario has [9:43] written that we need to quote pace the [9:45] frontier and Sam has agreed. People may [9:47] be surprised by my response. Go ahead. [9:49] Uh you guys are the frontier by any [9:51] reasonable metric. Market share, revenue [9:53] growth, model capability. The two of you [9:55] have a duopoly on frontier intelligence. [9:57] You've also claimed the lead is widening [9:58] because of recursive self-improvement. I [10:00] don't think what you see I don't see [10:02] what you see in the lab. If the [10:04] unreleased models are scary enough that [10:06] you think you should slow down, I [10:07] support your decision to be responsible. [10:09] But stop pretending you need anyone [10:11] else's permission. Stop pretending [10:12] antitrust law has to be suspended so you [10:15] can form a cartel. Stop pretending you [10:17] need a regulatory approval process that [10:19] supersedes product liability. Stop [10:21] pretending meter is independent when it [10:23] is intertwined with anthropics investors [10:24] and staff. Stop pretending that you need [10:27] those same evaluators to police [10:29] competitors who aren't even at the [10:31] frontier. And so yeah, maybe that's a [10:33] little bit of an an issue although yeah [10:35] I mean the the regulatory capture angle [10:36] is is always like if the safety rules [10:41] become super cumbersome such that you [10:44] can't start a instinct like product [10:47] right it's like instinct is a small team [10:51] they're not building a frontier model [10:54] they're not training their own models at [10:55] least not yet but they're like doing [10:56] stuff with AI if all of a sudden they [10:58] have to spend a year negotiating ating [11:01] with a bunch of third party evaluators, [11:02] have a bunch of people embedded and then [11:04] go through some sort of review process [11:05] and the government's not getting back to [11:07] them. Like you can't have a breakout [11:08] [clears throat] consumer product. Uh and [11:11] that [11:11] >> doesn't feel like there's been any type [11:14] of effort from anywhere to try to [11:16] control uh or regulate like application [11:19] layer companies and and Greg was on [11:22] OddLots. this uh episode came out this [11:25] morning talking about not trying to uh [11:27] doesn't doesn't believe there should be [11:29] any sort of limits to um [11:30] >> application layer stuff. [11:31] >> Yeah. Basically like if you just want to [11:33] train a model at home for your own use [11:36] cases, you know, there shouldn't be [11:37] limits on that. Yeah. Uh Trump Trump uh [11:40] came out yesterday. He was pretty active [11:41] over the weekend. He said the only [11:43] controller quote unquote guard rails [11:45] that AI needs is a strong and smart high [11:48] IQ president. And the USA has that in [11:51] spades. Trump [11:53] [laughter] [11:54] >> wild post. [11:57] >> The Trump administration has stopped AI [11:59] people from doing bad or potentially bad [12:02] things like Daario [12:04] who is now pretending to be a perfect [12:06] little angel and we will continue to do [12:09] so. We already have tremendous criminal [12:11] and regulatory power over these [12:12] companies. There is a sick conspiracy [12:15] going on against AI and data centers and [12:17] the only one that is happy about it is [12:19] China. Whoever wins AI wins. We are [12:21] leading China and all others and will [12:23] continue to do so. Conspiracy theorists, [12:25] treasonous traders, and leakers beware. [12:28] Thank you for your attention to this [12:29] matter. It is funny to accuse uh to [12:31] accuse like the labs of being in a [12:34] conspiracy against AI. [12:36] >> Yeah, I that really really doesn't [12:38] follow. I don't Yeah, that does not make [12:41] >> but uh he also followed up just 40 [12:44] minutes ago said concerning AI when in [12:47] the history of business capital H [12:49] capital B history history of business [12:51] did anyone see the leaders of an [12:52] industry call for regulation that if [12:54] strongly implemented would drive them [12:56] into oblivion and bankruptcy. AI is [12:58] taking over the world, destroying [13:00] humanity and all other things bad is a [13:02] hoax. No different from Russia Russia [13:04] Russia Ukraine Ukraine Ukraine. [13:06] Impeachment hoax one, impeachment hoax [13:08] two, and all the other hoaxes and scams [13:10] that America was forced to endure. [13:12] President Xi of China just announced [13:14] that China will be doing absolutely [13:15] nothing to stand in the way of AI or its [13:17] future. Google has recently stated that [13:18] they want to build a massive plant in [13:20] Finland, all because they are finding [13:22] permitting too difficult in the United [13:24] States. I'm not happy about this and [13:26] want them to change their thinking. AI [13:28] and data centers will be the greatest [13:29] economic development engine in history. [13:31] Bigger than oil, gold, diamonds, or even [13:33] the internet. Bigger [laughter] [13:36] Diamond [13:40] [laughter] bigger than diamonds. [13:44] Like of all the technology, the [13:46] internet, the the internet, the [13:49] transistor, [laughter] [13:53] >> the wheel, diamonds, these are [13:56] equivalent things. [13:57] >> It will not be stopped by brilliantly [14:00] run destructive forces during the term [14:03] of President Donald Trump. Okay. [14:05] >> Um, [14:06] >> well, [14:06] >> anyways, I I'm I'm wondering what the [14:10] So, a bunch of hedge fun hedge funds are [14:12] paying, you know, they paid to get [14:13] access to this. Let's assume like a [14:15] minute early, something like that. How [14:17] did they trade this? [laughter] Because [14:19] it's very [14:20] >> long [snorts] to be. [14:22] Yeah. [14:23] >> Long long diamonds. [14:25] >> Yeah. Maybe diamonds are are are [14:27] underappreciated. [14:28] >> I think diamonds are actually in a [14:29] really rough spot. I think people are [14:30] moving to lab grown stuff really [14:32] aggressively. So, it's a it's actually a [14:34] tough tough period of time for the [14:36] diamonds and I'm sure the diamond [14:37] industry is calling for regulation for [14:40] for sure. Um, [14:42] >> we got to keep going because the [14:43] president was on a roll this weekend. [14:45] Uh, let's pull up this video. [14:48] >> We taking a lot of their cars. Car [14:50] tariffs. [14:50] >> President Trump, this week again there's [14:52] been so much concern about AI. Have you [14:54] heard directly from any of these AI tech [14:57] leaders and you seem to downplay some of [14:59] the concerns that we've heard? But it's, [15:01] you know, it's going to be more good [15:02] than bad, but by a lot. [15:05] >> And I've said in the beginning, whoever [15:07] wins AI, and we're leading by a lot. [15:11] Whoever wins AI wins. [15:12] >> Have you ever used it yourself or have [15:14] you been [15:14] >> use AI? Yeah, I do. [15:15] >> How do you use AI? [15:16] >> But, uh, I would say, [15:18] >> you know, for the most part, you can use [15:20] AI for a lot of things. [15:21] >> What do you use it for? [15:22] >> Right away. Speaking of AI, [15:24] >> the funny thing is that is that like [15:26] like there's AI generated slop on his [15:29] social true social account like [15:30] constantly like the slop videos are [15:31] obviously AI. Like both sides of that [15:35] debate or that discussion should have [15:37] known that I guess but I guess it's like [15:39] people on his team are creating like the [15:41] AI slot videos because like there will [15:43] be like a new vibecoded like like uh you [15:46] know some sort of simulator that they [15:48] put up on White House stuff all the [15:50] time. So, but I guess that's not him. [15:51] It's probably people on his team that [15:53] are like, I made this AI video. [15:54] >> Yeah. [15:54] >> But I don't know. Gavin Baker had a [15:57] roundup of what happened. Let's go [15:59] through what's going on. And lots of [16:00] people just sort of like processing [16:02] this, going back and forth. People were [16:03] expecting Elon to be sort of against [16:05] this, maybe anti-Dario. Uh he of course [16:08] came out in favor and has been talking [16:10] about how like he was working with Nick [16:11] Bostonramm in 2014 on AI safety. Has [16:14] been very concerned about this the whole [16:16] time. Um although at various points in [16:18] time Elon has been less of a doomer and [16:21] more like we got to build we got to go [16:23] we gota we we we gota you know actually [16:26] compete here uh working both alongside [16:29] labs co-founding OpenAI founding uh XAI [16:33] growing the the model capabilities there [16:35] but uh he's like firmly in favor of this [16:38] so let's see what Gavin Baker had to say [16:40] he said wild 24 hours for AI and lots of [16:42] different proposals have been made TLDDR [16:44] the only tangible new fact is that [16:46] OpenAI and anthropic are going to have [16:48] embedded third-party evaluators from [16:51] unknown organizations with DARIA [16:53] floating meter as a possibility. Having [16:56] third-party evaluators is smart as [16:58] there's no section 230 style liability [17:01] shield for model outputs and showing a [17:04] duty of care will be important in future [17:06] litigation. Several internet companies [17:09] might have gone bankrupt without section [17:10] 230. So, limiting liability really [17:13] matters. I think there are minimal [17:14] investment implications from this single [17:16] new fact. But I do think that anyone who [17:18] wants quote a smoother for longer cycle [17:22] than most constraints uh are good [17:24] wafers, watts, rail rates and spreads. I [17:26] was thinking about how the uh a lot of [17:29] the safety folks are worried about just [17:30] the capability build out from the [17:33] semiconductor supply chain like they [17:34] they would like to restrict the amount [17:36] of chips that are created. And I'm [17:37] wondering if u should we not be pacing [17:41] electricity generation? Does everyone [17:43] agree that we should be building more [17:45] solar, thermal, wind power right now? Is [17:49] that uncontroversial? Is there anyone [17:51] who says like no, we shouldn't be doing [17:53] that? Because if we wind up in a world [17:54] where you're pacing frontier AI [17:57] development, current models are sort of [18:00] diffusing and there's some economic [18:02] impact there, but you're not, you know, [18:04] blitzcaling the data centers and the [18:06] frontier intelligence. Can you apply [18:08] some of that capital, some of that human [18:11] talent to just getting nuclear power [18:13] online, getting solar power online, re [18:16] re-industrializing the American energy [18:18] supply chain, battery, relieving grid [18:21] tension because everyone who complains [18:23] about higher uh energy prices would be [18:26] happy with that. I don't think the AI [18:29] safety folks would be upset about that [18:31] because you could use it for a lot of [18:32] things. Worst case, it's like, okay, you [18:35] wound up with like, okay, we can never [18:36] build, you know, AGI or ASI because it's [18:39] too dangerous. At least we have a lot of [18:41] cheap energy. Is there any reason why [18:42] you wouldn't why you would want to pace [18:45] the build out of the American electrical [18:46] grid? I can't see why. What do you [18:48] think? I mean at some level like yeah it [18:51] makes sense because it it is like an [18:52] input but it's it's like seems like way [18:54] too diffuse and also because like yeah [18:56] it's like if you build a lot of chips [18:58] like sure you can like you know make [19:01] more data centers that run YouTube or [19:03] whatever but like there's kind of a a [19:05] narrow use case for chips. [19:06] >> Yeah. [19:07] >> Where power you can kind of do whatever. [19:09] >> I wonder if you uh Yeah. Yeah. At what [19:11] point is there is there some sort of [19:13] like AI safety um valve that's like we [19:17] need to regulate that company in South [19:19] Korea that makes toilets that also makes [19:21] films for semiconductors because if we [19:22] can restrict them then we can slow down [19:25] the pace of AI progress uh going really [19:27] really deep in the supply chain for for [19:29] safety reasons potentially to summarize [19:31] the events Gavin Baker says Daario made [19:33] the most maximalist proposal of the [19:35] weekend embedded third party evaluators [19:37] national regulatory regime for models [19:39] beyond a certain capability ingredient [19:41] threshold, a broad international [19:43] regulatory packed between democracies, [19:45] stricter limits on compute distillation [19:47] for China, and then a different [19:49] international regulatory regime that [19:51] encompasses China before there is ever a [19:53] national regulatory regime. He wants a [19:55] Sherman Act waiver so that Anthropic can [19:58] safely coordinate with OpenAI and other [20:00] frontier labs without antitrust fears. [20:01] To be frank, this latest proposal is [20:03] much less maximalist than some of his [20:05] prior proposals like policy on the AI [20:08] exponential [clears throat] [20:09] where he advocated for an FAA of for AI. [20:12] Now people are talking about FINRA for [20:14] AI which is more non-governmental [20:16] regulation but industry focused. Uh I [20:18] believe he is sincere in his beliefs and [20:20] despite all the protestations all of [20:22] this would also probably be good for his [20:24] business over the long term. Yeah, [20:26] there's a there's there's a lot of [20:28] tricks here where uh people where [20:31] something there if if there are if there [20:32] are multiple quadrants of like actions [20:34] you can take and it's like it's either [20:36] good or bad for your business or it's [20:37] good or bad for safety, you can land in [20:40] the good for your business and good for [20:43] safety quadrant and everyone will accuse [20:46] you of being cynical and doing it only [20:47] because it's good for your business. And [20:49] so yeah, [20:50] >> it gets really difficult to to say okay [20:52] yes like the thing that I'm advocating [20:53] for is good for my business. Now run had [20:55] another take which is like no if you're [20:57] pacing the frontier that's going to [20:58] compress margins because everyone else [21:00] is just going to catch up and it's going [21:01] to become more oligopolistic. So if you [21:03] hold you know like yeah there's this [21:04] duopoly right now for the true frontier [21:07] amongst open AI and anthropic but if you [21:09] hold that constant like watermelon from [21:11] uh MSI uh MSL is going to catch up deep [21:14] mind's going to catch up like you're [21:16] going to have three or four Grock is [21:18] going to train to that level and so [21:19] you're going to have four five six seven [21:22] really capable competitors. Uh what was [21:25] interesting is like Ara Carzian was [21:27] saying that AI spend declined a little [21:29] bit in August and that was mostly [21:31] because of price cutting. It wasn't [21:33] actually people using less AI. So it's [21:35] not like people are pulling back on [21:37] their use of AI but the there's more [21:40] price competition. So open AI is [21:41] discounting and people are being we saw [21:44] like the fable adoption was a little [21:46] weak and a lot of it's like hey are we [21:48] sure we want to use the million-dollar [21:50] model for checking the weather? like [21:52] let's be a little bit more judicious [21:54] about this. And so if you wind up with a [21:57] with a world where there's much more [22:00] price competition that could compress [22:02] the frontiers margins, which is not good [22:04] for business. So doesn't really put you [22:06] in that top quadrant that you have to be [22:08] so cynical about. The interesting thing [22:10] is that he said that he doesn't think [22:11] that it's open- source Chinese models [22:13] that are actually putting pricing [22:15] pressure on the frontier. It's more of [22:17] the duopoly dynamic. And I think that's [22:19] because I think he sort of alluded to [22:20] this that there's just a lot of [22:22] companies that are like I am comfortable [22:24] doing business with an open AAI [22:25] anthropic uh a Google of course [22:28] something that's on AWS something that's [22:30] on Azure. Uh yes I could just download [22:33] the weights for DeepSeek but uh I'm not [22:36] really comfortable with that. I don't [22:37] really want to deal with that. I don't [22:38] want to deal with the scaling and uptime [22:40] of that. And so yes, there are APIs out [22:42] there that are reasonable, but none of [22:43] them are as mature as the as the [22:47] frontier labs, not in terms of the [22:49] intelligence, but in terms of the actual [22:51] if I implement this in [clears throat] [22:52] my business, will it reliably give me [22:53] the right results? Is there high quality [22:55] SLA? Will the service stay up? Will [22:57] uptime be good, etc. So, um there's [23:00] there is a world where you could get to [23:03] a place where this is bad for the [23:05] frontier companies business-wise. And [23:08] Sam has talked about that in an [23:10] interview with Fortune saying like if [23:12] it's bad for our business like we've we [23:14] have said to our investors at every step [23:16] for like a decade that like this is kind [23:18] of a nonprofit. This is a this is a this [23:21] is a there are a whole bunch of [23:22] different ways in which this is not [23:24] purely aligned to uh shareholder [23:26] interests like other companies PDC's. [23:29] Yeah. Public benefit corporations. So [23:32] they so there so so so I think both Sam [23:35] and Daario feel very confident about [23:37] being able to go to their investors and [23:39] say hey we're doing something that's [23:41] going to slow down revenue growth or [23:43] tank margins or you know just reduce the [23:47] market cap of the company but we think [23:49] it's the right thing to do so deal with [23:51] it and you can't really sue us because [23:53] you invested in a PBC but no one's [23:55] really giving them credit for that [23:57] everyone thinks like no this is [23:58] cynically what gets them to 10 trillion [24:00] this is what gets them 10 trillion and I [24:02] don't know may maybe maybe it is but uh [24:04] but it definitely seems like uh like the [24:07] the if we're if we're on the midcurve [24:09] you know bell curve meme like the dumb [24:11] guy thing is like slowing down is bad [24:14] for business right like like the [24:16] mid-curve is like slowing down is [24:18] actually good for business and the 40 [24:20] chess and like do you don't want to be [24:22] the best and have 100% market share you [24:25] want 20% market share because you let [24:27] everyone catch up to you like I I'm not [24:29] I'm not super convinced to that. But, [24:31] you know, there's plenty of people that [24:33] are making that case online. Uh, [24:34] >> got some breaking news. [24:35] >> What's the breaking news? [24:36] >> The president just phoned Jensen live on [24:39] stage at the All-In Summit. Whoa. [24:41] >> Mike Isaac says, "Okay, wow." Per folks [24:43] there. Trump called Jensen in the middle [24:45] of his interview. He was on stage and [24:47] Jensen put the phone on speaker phone. [24:49] Jensen thanked Trump for his social [24:50] media post this morning pouring cold [24:52] water on alarmism. [24:54] >> Yeah, Jensen's been anti-doom for this [24:57] entire cycle. We'll see. We'll see where [25:00] it all goes. I'm sure there'll be more. [25:02] >> Mr. President, [25:04] >> you're now talking to the planet. [25:06] >> See, the great thing about life is that [25:09] Jensen can develop the most complex. [25:11] >> It's the joke again for 10 years, but he [25:15] can't figure it out. Put me on speaker. [25:17] [laughter] [25:19] >> An odd joke. Okay. Well, anything [25:22] substantive come out of it? I'm sure [25:23] we'll know more tomorrow. [25:25] >> We'll be back. [25:26] >> We'll be back at 11:00 a.m. Pacific. [25:29] Thank you for tuning in. Leave us five [25:30] stars on Apple Podcast and Spotify. I'm [25:32] off to go get some rest and we will see [25:35] you tomorrow. [25:37] >> Our newsletter, tbp.com. [25:39] [music]