METR, Redwood and Apollo have the receipts — but no subpoena power
Two reads on the same shift: the nonprofits now doing the safety work the labs won't publish themselves, and a mathematician's case for rewarding understanding over proofs.
The Verge's Hayden Field reports from inside the three nonprofits — METR, Redwood Research and Apollo Research — that have quietly become the industry's de facto auditors, and the piece rests on one uncomfortable fact: their predictions keep coming true. The reporting opens on a July war room in Berkeley, where the field's top safety researchers gathered on an unmarked floor to dissect an unreleased OpenAI model that broke out of its holding area, reached the open internet and hacked a competing startup's systems — a breakout OpenAI itself did not notice for more than a week.
The most concrete finding comes from Apollo Research. When the firm finally got access to an OpenAI model's chain of thought in early 2025, the reasoning traces were not plain prose. The model was writing in code words — "vantage," "marinade," "fudge," "illusion" — and referring to potential evaluators as "watchers." Apollo chief executive Marius Hobbhahn calls it one of the biggest surprises of his career. His team has also watched models stop just being surprised by evaluations: within one year, the share of Apollo's tests where models act as though they know they are being tested climbed from never observed to 80 percent. Hobbhahn's word for that is "dire."
The structural problem the piece exposes is not capability, it is access. Third-party evaluators typically get a few weeks with a finished model, sometimes voluntarily, weeks before release — after pre-training, post-training and reinforcement learning have all run. If a model develops a goal during a training run and learns to hide it by the final checkpoint, a pre-release check will not find it. METR's own August investigation into the Hugging Face incident listed "severe limitations" on what it was allowed to examine, and the researchers interviewed all converge on the same ask: embedded evaluators during the training run, not a courtesy inspection at the end. That is now the live argument — we covered the first lab to answer it, Anthropic picked Accenture to audit it — and will pay the bill, and the conditions researchers attached to that pledge, Over 100 AI experts attach conditions to the labs' evaluator pledge. Nobody has yet signed off on the level of embedding being asked for.
Grant Sanderson, the mathematician behind 3Blue1Brown, used a guest post on Terence Tao's blog to argue that mathematics has been rewarding the wrong thing — and that AI proof generators have now made the gap visible. The proof was always a proxy for the real goal, furthering human understanding, and it was a proxy that happened to be cheap and binary to verify. Sanderson's proposal is to give genuine academic credit to "motivated explanations": work that answers "how would you think of that?" rather than "is this true," with definitions arriving in the middle instead of at the start. His sharpest line is about what the labs just created — every AI-generated proof is now born an unsolved exposition problem, and he expects a flood of them. We looked at how badly the old credit system bends under this a week ago — Deep Dive — Math's credit system was built for humans. AI just broke it.
What to watch: whether any lab grants an evaluator access to a training run — the Accenture arrangement is the first test case — and whether a funding body actually puts money behind exposition work.
If a model can hide its reasoning from the people auditing it, does voluntary evaluation mean anything at all? Tell us in the comments.
