The AI Slowdown Debate: Dario's Three-Step Plan, and Both Sam and Elon Nodded
Dario posted Saturday morning demanding we "pace the frontier," and Sam and Elon both endorsed it the same day; but the only genuinely new fact is that OpenAI and Anthropic will embed third-party evaluators — and that step may require an antitrust exemption first.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Pace the frontier is not a new phrase — the line was laid in late July
The starting point of this discussion is not Dario's Saturday essay. On July 28, employees and leaders at OpenAI, Anthropic, DeepMind, Meta and other institutions signed an open letter on "pacing the frontier"; the same day Sam Altman said publicly that "we may have to slow down AI development"; two days later OpenAI told Axios that the letter's wording was shaped with OpenAI's involvement, and that Altman had discussed pacing with White House officials. On August 18 OpenAI published a document on the pace of model development in the era of cyber-critical capabilities, saying it had indeed slowed scaling and paused one large RL training run — that was the two-week training pause. On August 31 an official Anthropic post explicitly discussed pacing the frontier and called for a coordination mechanism. On the morning of Saturday, September 12, Dario's essay turned this into a public movement.
— Speaker 2Dario's third step is the one that will actually blow up
Dario's three-step plan: one, let independent third-party evaluators into AI companies to verify safety practices — he named METER, which sparked the argument over whether METER is really independent: METER has employees who worked at OpenAI and Anthropic, and people at Anthropic have been at METER, so it's a revolving door. Two, frontier labs and democratic governments establish AI safety standards with government help — the controversy here is whether labs need government approval. Three, the US and democracies should try to coordinate with authoritarian governments, meaning US-China coordination on AI pacing. The show points out the difficulty of step three: China still wants to catch up, and if both sides have exactly equal capabilities, is that the optimal equilibrium? If you always want to be ahead, it's hard to negotiate with the other side that "you keep only half our nuclear weapons."
— Speaker 2Step two may require exempting the Sherman Antitrust Act first
Dario's step two asks frontier labs to coordinate safety standards with government help, and the show points out a hard legal problem here: the Sherman Antitrust Act has very clear rules about firms acting in concert. The analogy is that if American, Delta and United all said "for safety we should cut flights," the likely result is less capacity, higher fares, higher profits and consumers harmed — which is exactly where the Sherman Act came from. So even if Demis, Dario, Sam and Elon all agree, actual federal regulatory rules would need to be adjusted or exempted. The show thinks this is precisely why Dario specifically calls for this step: it's easy to loosely agree in a blog post, but forming anything formal requires government approval.
— Speaker 2David Sachs's six straight "stop pretending"s
David Sachs posted on X, first conceding "by any reasonable metric — market share, revenue growth, model capability — you are the frontier," noting that OpenAI and Anthropic have a duopoly on frontier intelligence, and claiming the lead is widening because of recursive self-improvement. Then he fired off six "stop pretending"s: stop pretending you need anyone's permission; stop pretending antitrust law must be suspended so you can form a cartel; stop pretending you need a regulatory approval process that sits above product liability; stop pretending METER is independent when it's intertwined with Anthropic's investors and employees; stop pretending you need the same evaluators to govern competitors who aren't at the frontier at all. The show also points out the regulatory capture angle: if safety rules become extremely onerous, a small team like Instinct that doesn't train frontier models would have to spend a year negotiating with third-party evaluators, embedding a bunch of people, and going through government review — and couldn't ship a hit consumer product.
— Speaker 2Trump says the only guardrail AI needs is a high-IQ president
Trump was active over the weekend. He said the only control and guardrail AI needs is a strong, smart, high-IQ president, and that America has "plenty" of that. He also said this administration has already stopped AI people from doing bad things or things that could be bad, naming Dario, who now pretends to be a perfect little angel. He called it a sick conspiracy against AI and data centers, said the only one happy about it is China, and that whoever wins AI wins. Forty minutes later he posted again: when in business history has a leader of an industry called for regulation that, if strongly enforced, would push them into bankruptcy? He called AI taking over the world and destroying humanity a hoax, no different from Russiagate, Ukrainegate and the two impeachment hoaxes. He also mentioned that Xi Jinping had just announced China will put no obstacles on AI or its future, and that Google is building a big plant in Finland because US approvals are too hard.
— Speaker 1Gavin Baker: the only new fact is third-party evaluators
Gavin Baker's weekend summary said that in 24 hours proposals were flying everywhere, and the TLDR is the only substantive new fact: OpenAI and Anthropic will embed third-party evaluators from unknown organizations, with Dario offering METER as one possibility. He thinks third-party evaluation is smart because model outputs have no Section 230-style liability shield, and showing a duty of care in future litigation matters — without Section 230, several internet companies might already be bankrupt. He judges the investment impact of this single new fact to be minimal. He also summarized Dario's proposal as: embed third-party evaluators, establish a national regulatory regime for models above some capability threshold, a broad international regulatory agreement among democracies, stricter compute distillation restrictions on China, and a different international regulatory regime that includes China.
— Speaker 2Pacing could compress the frontier labs' own margins
The show offers a counterintuitive argument: if pacing actually happens, it could compress margins, because others catch up and the market becomes more oligopolistic. Right now OpenAI and Anthropic have a duopoly at the frontier, but if the current pace holds, MSL, DeepMind and Grok will all catch up to the same level, and in the end there will be four, five, six or seven genuinely capable competitors. Meanwhile Karazian said August AI spending dipped slightly, mainly because of price cuts rather than people using it less — OpenAI is discounting, Fable's adoption is weak, and people are asking "are we sure we want to use a million-dollar model to check the weather." The show argues the pricing pressure on frontier labs comes not from Chinese open-source models but from the duopoly dynamic itself: enterprises are willing to do business with OpenAI, Anthropic, Google or something running on AWS or Azure because it's mature, has SLAs and guaranteed uptime, whereas downloading DeepSeek weights and running it yourself is something many people don't want to deal with.
— Speaker 2Nobody gives them credit for their PBC structure
Sam said in a Fortune interview that if this is bad for our business, we've told investors at every step for ten years that this is more like a nonprofit, and there are many ways it isn't purely aligned with shareholder interests like other companies — a PBC, a public benefit corporation. The show thinks both Sam and Dario are confident they can tell investors: we're doing something that will slow revenue growth, compress margins, maybe even lower the company's market value, but we think it's right, and you can't do anything about it because you invested in a PBC. But nobody really gives them credit for it; everyone sees it as a cynical move to get them to a ten-trillion-dollar market cap. The show sums it up with the bell curve meme: the low end is "slowing down is bad for business," the middle is "slowing down is actually 4D chess and good for business," but the show itself says "I don't really buy it, though a lot of people online are saying it."
— Speaker 2In their own words · checked verbatim
The US and other democratic countries should try and coordinate with authoritarian governments. So he wants The US and China to coordinate on AI pacing.
Speaker 26:30
There are actual legal and federal regulatory rules that might need to be adjusted or waived something like this to happen.
Speaker 28:30
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. Stop pretending meter is independent when it is intertwined with Anthropics investors and staff.
Speaker 210:00
The only controller guardrails that AI needs is a strong and smart high IQ president and The USA has that in spades.
Speaker 114:30
TLDR, the only tangible new fact is that OpenAI and Anthropic are going to have embedded third party evaluators from unknown organizations with Daria floating meter as a possibility.
Speaker 218:30
If you're pacing the frontier, that's gonna compress margins because everyone else is just gonna catch up and it's gonna become more oligopolistic.
Speaker 220:30
No one's really giving them credit for that. Everyone thinks like, no. This is cynically what gets them to 10,000,000,000,000.
Speaker 223:30
Figures
| Length of OpenAI's paused large RL training run | two weeks | 4:30 |
| Time Dario posted | 7am Pacific Time Saturday morning | 5:00 |
| Time Elon endorsed Dario's essay | 8am Saturday morning | 5:00 |
| Time Sam Altman endorsed | 9:30am Saturday morning | 5:00 |
| Time David Sachs posted | 8:15am Saturday morning | 5:00 |
| Change in August AI spending | slightly down, mainly because of price cuts | 20:30 |
Glossary
- pacing the frontier
- The idea of a coordinated slowdown in the development speed of the most frontier AI models.
- PBC
- Public Benefit Corporation, a structure that lets a company weigh public benefit goals alongside shareholder interests.
- Sherman Antitrust Act
- The core US antitrust law prohibiting firms from conspiring to restrain competition.
- Section 230
- The US legal provision granting internet platforms immunity from liability for content.
- METER
- An AI safety evaluation body with personnel flows to and from OpenAI and Anthropic.
How to listen
Founders, investors and engineers watching AI policy and regulation, especially those who want to understand who is pushing this "AI slowdown" discussion, who opposes it, and where the legal obstacles are.
The opening chit-chat about rhabdo and the US Open can be skipped.