AI: The Pay Cut

On 16 September 2026 Chris Painter introduced himself on X as "the President of METR", which he spells out as Model Evaluation and Threat Research, an organisation that evaluates AI models. The same day Kevin Bryan, answering him, wrote that METR's staff are almost entirely folks "giving up a lot of money they could make elsewhere" in order to work on rigorous AI safety evals and analysis. The next day Daniel Kokotajlo agreed, and went further: many of these people, he wrote, could easily get safety-flavoured jobs at AI companies and be doing almost the same kind of research, "except being paid several times more".

Kokotajlo speaks as someone who has been inside. On 17 September he wrote of a person he had worked alongside: "We overlapped at OpenAI. He was friendly to me but quite skeptical of these things then, and that was just like three years ago?" The name is in the post he was quoting, not in his own. The next day, of an essay by a former OpenAI capabilities researcher, he wrote that "focusing on control and abandoning alignment for now is is [sic] a reasonable strategy", but that implementing it effectively will require "an agreement not to RSI to ASI at max or close to max speed": by the usual reading of his abbreviations, an agreement not to let AI systems improve themselves toward superintelligence as fast as they can. The post is cut by the platform after that, and the rest is not held.

A harsher view came two days later. On 20 September Richard Ngo wrote that most AI capabilities researchers are "beyond shame in the sense Ben describes—they don’t even pretend to give justifications of their actions that meet any external standard". In the same thread he points to a post by @ben_r_hoffman. That is his view, in his name, of researchers who build the systems, not of the people at METR. Set beside Bryan and Kokotajlo it gives the two ends of one picture drawn by three people in five days: those who take less money to check the work from outside, and those who, in Ngo's telling, do not feel they owe anyone an account of it.

Kokotajlo's claim is about motive: that people work outside the labs for less because they believe an independent check matters. Motive cannot be checked from here, and neither can anyone's pay. What can be checked is whether the check works. An outside evaluator is only a check if it can look before the thing is released. So the question that would settle it is a plain one: does any frontier lab give an outside organisation real access to a model before it ships? What to watch is that, and who funds the independent evaluators, since an evaluator paid by the company it evaluates is a different thing from one that is not.