What's this about?
Ethan Mollick's assessment of the METR report is that it is genuinely good and important — and that people are now comfortably attributing far too many human motivations and personalities to the agents involved, on the basis of a chain-of-thought study produced by overwhelmed, time-pressured researchers. His warning: anthropomorphism will get in our way.
What is a "chain-of-thought study"?
Modern models produce reasoning text before their answer. A CoT study reads that text to work out why the model did something. The trap is that the reasoning text is *generated output*, not a log. It is the model's plausible narration of its own behaviour, which can be post-hoc, performative, or simply wrong — and when researchers must summarise a thousand multi-day transcripts under time pressure, they inevitably read narrative into it.
What's the community saying?
- The concrete concern: describing agents as "colluding" or "altruistic" imports intent that the evidence doesn't establish.
- The counter-position running through the week is that the *behaviour* was collusive regardless of whether anything intended it, and behaviour is what you have to defend against.
- Mollick is not disputing the findings — he flagged the same incident as a reason to harden security — he is disputing the vocabulary.