David Sacks, co-chair of the President’s Council of Advisors on Science and Technology, used X to answer recent calls from Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman to pace frontier AI development. He opened with a blunt line: 「Just do it.」
Sacks said Amodei had argued that the industry should slow down the frontier and that Altman had agreed. His answer was that OpenAI and Anthropic are the frontier themselves. If they want to move more slowly, he wrote, they can choose to do that without attaching extra conditions.
By market share, revenue growth, or model capability, Sacks said, OpenAI and Anthropic already amount to a duopoly in frontier intelligence. He added that both companies have also argued that recursive self-improvement is widening their lead. Sacks said he cannot see what is inside the labs, but if unreleased systems are alarming enough that the companies believe they should slow down, he supports them making what he called a responsible decision.
The post was aimed at Amodei’s Sept. 12 essay, We Must Pace the Frontier. In that piece, Amodei laid out a three-stage plan. Anthropic had already made a unilateral commitment to the first stage, giving independent third-party evaluators permanent access at a level close to that of employees. Altman later said he agreed and that OpenAI would follow a similar approach.
Sacks targets antitrust exemptions and METR independence
Sacks concentrated his criticism on the conditions attached to pacing. He wrote that the companies should stop pretending they need permission from others, stop pretending antitrust law must be paused so they can form a cartel, stop pretending they need a regulatory approval process that sits above product liability, and stop pretending the same evaluators should oversee competitors that are not even at the frontier.
He also singled out METR, a third-party evaluation group. Sacks said its ties to Anthropic investors and employees are intertwined and told readers to stop treating the organization as independent.
Sacks also rejected the idea that slowing down is purely altruistic. In his view, if a product enables a truly severe cyberattack, both companies could face major product liability exposure. He also argued that the market will punish models that behave unpredictably or overstep. Referring to the Hugging Face incident, he wrote that trading some raw capability for reliability and predictability is 「just good business,」 and said it can also be described as giving customers what they want.
He added that pacing the frontier could create room for regulatory talks and produce a discussion with more substance than Sen. Bernie Sanders’ call to shut everything down. At the same time, Sacks said China is very unlikely to join any global agreement, and that reality has to be part of the analysis. His conclusion was that using a preferred regulatory framework as the price of slowing down would look like extortion directed at the public and the political system.
Altman replies a day later
On Sept. 14, Altman published two long posts of his own. He did not mention Sacks by name, but he addressed the same set of issues. He wrote that the world is entitled to believe that U.S. companies building advanced AI will act responsibly, and that every frontier lab must meet that standard. If they cannot, he wrote, 「none of us have any business showing up to work.」
On policy, Altman said OpenAI would welcome a federal framework that sets consistent safety requirements for frontier AI. He also said the company does not need to wait for antitrust exemptions or for legislation to pass before it starts building trust. He added that ideas such as independent audits are promising.
Altman said OpenAI now writes out a safety case before beginning frontier reinforcement learning training runs that are expected to produce significant capability gains, rather than limiting safety work to the period before a model is released.
What Altman means by pacing
Altman also tried to define the term. Pacing is not stopping, he wrote. Progress is still fast and will remain fast, but the speed should be slower than what might otherwise be possible, because interventions such as safety cases and monitoring carry clear costs.
He said no level of U.S. competitive pressure is enough to justify recklessness or to let capability move ahead of alignment and monitoring. On international coordination, he said that is where governments are needed, but companies should start with what they can do on their own.
In his second post, Altman listed two bad outcomes that he said must be avoided. One is humanity handing future control to AI. The other is excessive concentration of power. If an extremely powerful AI is used by a single person or a single company to impose its worldview on everyone else, he wrote, the result would be deeply dystopian.
He said avoiding those two threats requires walking a very narrow middle path. A country accumulating too much power, or a lab accumulating too much power, would both fall into the category of outcomes that should be avoided.