Altman and Dario both call for slowing down AI; Trump tech adviser Sacks: If you're really afraid, stop yourselves—don't drag the whole industry into it.

动察 Beating AI News Flash: Anthropic CEO Dario Amodei published a long essay calling for slowing down capability improvements in frontier AI. He believes that AI is increasingly participating in next-generation AI R&D, and combined with several recent incidents of Agents going out of control, safety research is failing to keep up with capability growth.

Dario proposed a three-step plan: first, let third-party evaluators gain long-term access to AI companies with near-employee-level permissions; then have frontier labs in democratic countries coordinate safety standards and, when necessary, slow down capability growth together; finally, push for international coordination. Anthropic has committed to implementing the first step first.

Sam Altman subsequently publicly supported it, saying that OpenAI has also been discussing "pace the frontier" in recent weeks, and announced that OpenAI will also bring in similar long-term independent evaluators.

Former White House AI and cryptocurrency lead and current co-chair of the President's Council of Advisors on Science and Technology David Sacks then hit back: "If you really think internal models are dangerous, then stop yourself." He said OpenAI and Anthropic already dominate frontier AI and can absolutely slow down on their own without dragging in other companies.

What Sacks dislikes most is several competitors discussing together that "everyone should slow down a bit." Dario himself also admitted that this kind of coordination may require antitrust exemptions. Sacks believes that if government approval is still needed to let the largest AI companies sit down together and set speed limits, it looks very much like a monopoly alliance.

He also named METR, questioning whether it is suitable to serve as an independent referee for the entire industry. METR has long conducted external evaluations for companies such as Anthropic and OpenAI, and recently Anthropic safety researcher Joe Benton left to join it. However, METR made clear that it does not accept funding from these frontier AI companies, only model access and free tokens.

Sacks also believes that slowing down itself aligns with the commercial interests of OpenAI and Anthropic. The more easily models go out of control, the greater the legal liability; customers are also more willing to buy stable, controllable AI. As for a globally unified slowdown, he judges it impossible that all countries would be willing to join.