Anthropic CEO Dario Amodei has published a long essay calling for frontier AI's capability gains to slow down. His argument: AI is increasingly doing the work of building the next generation of AI, and after a string of recent incidents where AI agents went rogue, safety research can't keep up with capability.
Dario lays out a three-step plan. First, let third-party evaluators work inside AI companies long term, with access close to that of employees. Second, have frontier labs in democratic countries coordinate on safety standards and, if necessary, slow capability growth together. Third, push for international coordination. Anthropic has committed to the first step.
Sam Altman then backed him publicly, saying OpenAI has spent the past few weeks discussing how to "pace the frontier," and announced OpenAI would bring in long-term independent evaluators of the same kind.
David Sacks, the former White House AI and crypto czar and now co-chair of the president's tech advisory council, fired back: "If you really think your internal models are dangerous, just stop." OpenAI and Anthropic already dominate frontier AI, he said, so they're perfectly capable of slowing down on their own. No need to drag other companies along.
What bothers Sacks most is competitors getting together to agree that everyone should go slower. Dario himself concedes that kind of coordination might require an antitrust exemption. And if the government has to give the green light for the biggest AI companies to sit down and set a speed limit, Sacks argues, that starts to look a lot like a cartel.
He also called out METR, questioning whether it's the right fit to serve as an independent referee for the whole industry. METR has long run outside evaluations for Anthropic, OpenAI and others, and Anthropic safety researcher Joe Benton recently left to join it. METR says it takes no money from frontier AI companies — only model access and free tokens.
Sacks also argues that slowing down happens to serve OpenAI's and Anthropic's commercial interests anyway. The more likely a model is to spiral out of control, the bigger the legal liability; customers would rather buy AI that's stable and controllable. As for a globally coordinated slowdown, he figures China will almost certainly never sign on.
https://twitter.com/DavidSacks/status/2098973625252708460