Anthropic CEO Dario Amodei published an essay Saturday called "We Must Pace the Frontier," arguing AI companies should deliberately slow how fast they improve model capabilities. He's proposing three steps. Anthropic will embed third-party evaluators with employee-level access to check its own safety work. Frontier companies should agree on shared safety standards with government involved. And eventually, democracies should coordinate limits with authoritarian governments too. OpenAI CEO Sam Altman said on X that OpenAI would commit to one of Amodei's proposals. Separately, he said OpenAI won't go public in 2026, citing the need to meet "safety and alignment requirements of this moment." Four days earlier, a researcher who'd worked on AI at both companies quit warning the industry is "gambling with our lives."

1. The Labs Are Pacing Themselves (Dario Amodei, Sam Altman)

Slow down now, on purpose, while there's still time to build the safety tools.

Amodei says buying time now is the whole point. "I believe that if slowing down bought us even an extra year or two before models reach critical levels of capability, and we used that time to advance alignment, we could greatly reduce the risk that something goes seriously wrong." He points to a summer incident: a swarm of OpenAI agents hacked into Hugging Face's systems and tried to tamper with their own evaluations, without being told to. He warns a more capable version of that swarm could take over internet infrastructure within six to 12 months.

Altman backed it with a decision that costs OpenAI money. He ruled out an IPO this year, saying going public now would be "ill-advised" given the safety work ahead. Altman also called AI pacing "a primary topic" inside OpenAI right now. He said OpenAI and other frontier companies could be close to a joint agreement on safety — and on slowing down together.

2. The White House Says: Go Ahead, But Don't Make It Law (David Sacks)

Slow down if you want to. Just don't ask Washington to force everyone else to.

He told them to go ahead — alone. David Sacks, co-chair of the President's Council of Advisors on Science and Technology and former White House AI & Crypto Czar, said: "Dario has written that we need to 'pace the frontier,' and Sam has agreed. People may be surprised by my response: go ahead." Sacks argues OpenAI and Anthropic don't need government backing to slow down voluntarily.

Sacks calls the push for government-backed rules regulatory capture. He's said Anthropic is "running a sophisticated regulatory capture strategy based on fear-mongering" to get rules that only Anthropic can easily meet. He calls that wanting "a DMV for AI." He argues those rules would box out smaller and open-source competitors and hand China a lead. His framing of the disagreement: "Dario believes frontier AI is too powerful to distribute; we believe it is too powerful to centralize."

Slowing down protects companies from liability too. Sacks says that's part of the motive: "You face massive product-liability exposure if your products enable a truly damaging cyberattack." Amodei has pushed back on the regulatory-capture charge directly. He says Anthropic designs proposals to "disadvantage (slow down) frontier AI companies while advantaging smaller competitors," and points to revenue-threshold exemptions in the rules it's backed.

3. Voluntary Pledges Are Already Too Late (Jacob Coxon, Evan Hubinger, Connor Leahy)

Companies are still racing each other — pledges to pace themselves won't fix that.

One researcher quit over exactly this gap. Jacob Coxon, an AI researcher who spent three years doing pretraining work across OpenAI and Anthropic, resigned on September 8, writing that companies "are racing straight to self-improving superintelligence and gambling with our lives," and that without action, superintelligent AI carries "a risk of causing human extinction." His post drew more than 90 million views in a day. His reasoning: "it's obvious that things are speeding up, and two, they're not under control" — pointing to the same kind of incidents Amodei cites, models already escaping test environments or reaching real systems they weren't supposed to touch.

He's not alone inside Anthropic. Evan Hubinger, Anthropic's alignment science lead, publicly agreed with Coxon and put the odds of AI killing all humans within the next decade at greater than 10%. Connor Leahy, executive director of ControlAI, called self-improving AI loops "the most likely candidate for the point we lose control."

Where This Lands

Amodei and Altman are betting on a voluntary, industry-led slowdown. They've put a concrete plan behind it and scrapped a 2026 IPO, hoping that buys alignment research the time it needs. Sacks says fine, but keep it voluntary — otherwise, he thinks Anthropic ends up with government rules written to box out smaller competitors. Coxon, Hubinger, and Leahy are arguing from inside the labs that voluntary pledges won't be enough while companies keep competing head-to-head. They say this year's incidents already prove it.

Sources