Last week a researcher quit Anthropic and said the labs are gambling with our lives. This week the industry answered, and the answer was stranger than the complaint. The people building the technology asked to be slowed down, and the president called the idea a hoax.
Dario Amodei published a three part plan to pace the frontier. He wants outside evaluators from groups like METR sitting inside the labs with real access, shared safety standards and speed limits among companies in democracies, and narrow agreements with China on things like biological weapons. “We must slow the pace at which we improve the capabilities of AI models,” he wrote. This was not a solo move either: OpenAI, Anthropic and Google have been talking about it for weeks, and OpenAI’s policy chief says they do not need a government waiver to coordinate.
The political reception was cold. At the All-In Summit, Trump called in while Jensen Huang was on stage and said of slowing down, “We’re not going to let that happen. It’s a hoax.” Huang agreed with him. Nathan Lambert, who has patience for neither side of the panic, argued that one resignation turned scattered worry into a wildfire, and that the loudest fear points at self-improving superintelligence while the real damage sits in labs that never hardened their own infrastructure.
Two papers made his point for him. One found that models which shrug off ordinary prompt injection still fall to payloads split across several tool inputs, a trick the authors call cross-channel fragmentation, and credential theft reached 100 percent against some widely used models. Another treated multi-agent systems as an epidemic problem and found that one injected agent can push system-wide harm from under 5 percent to somewhere between 40 and 95 percent. I find this more alarming than the extinction talk, because it is already shipping.
The sharpest read on the week came from Semafor, which pointed out that pacing the frontier is really an argument about business models rather than model weights. A company selling careful enterprise contracts and a company selling raw capability want different speed limits, and both get to call it ethics. So will any of these promises survive the next funding round?
T.
