A lot of people in my bubble are mad at Anthropic right now.1 Nate Soares recently argued (compellingly! Thank you Nate!) that Anthropic is failing to live up to its ideals.
Inspired by Nate, I’d also like to flag something that is really annoying me about Anthropic. And OpenAI, you’re not off the hook either.
Many people recently got excited about the idea of pacing the frontier, including two guys with unusual influence over whether the frontier gets paced: Dario Amodei, who wrote a blog post called “We Must Pace the Frontier”, and Sam Altman, who said “I agree with Dario that we need to pace the frontier.”
Given the breathtakingly short timelines to Extremely Dangerous AI that both companies espouse, I naively assumed they’d urgently start doing things to make pacing more likely.
Things such as:
Doing something similar to AIFP’s pacing plan. It looks like a solid plan to me, but if you don’t like it, please tell the public why. And then you should use your thousands of employees and bajillions of dollars to come up with something better.
AIFP published this in August, which was years ago in AI time. Please hurry up.
You’ve already shown you can quickly reallocate staff focus towards security and alignment when needed (which is great). Do that here, and make the details public.
And no, hiring a single Technical Program Manager for Pacing doesn’t let you off the hook. We’ll need a lot more than that.
Not redacting so much of METR’s report. The full version could persuade reasonable skeptics that pacing is urgently needed.
Of course, OpenAI should give embedded evaluators enough access to produce a report that is at least as good.
Not putting out terrible proposals that could make it harder for policymakers to intervene if they think AI companies are about to create AI that could overthrow the government.
To their credit As a bare fucking minimum, Anthropic says here that by January 15th 2027, they’ll either “publish a proposal [on pacing AI development] or publish an explanation of why [they] decided not to.” It better be a damn good proposal, because we are short on time.
They’re also mad at OpenAI, but that has been the case for a while.
TBC, I’m not endorsing any of these specific measures (though prima facie they seem good); I’m more annoyed that there has been barely anything substantive done since the initial hurrah over pacing.





Why even continue the race at the enormous cost required? We know that almost no one has run up against limits that a vastly more model would surpass. The bottleneck now seems to be learning how to use the tools we have.