What happened
Researchers tested whether AI language models can actually learn from experimental feedback in a real biological screening task, and found they do — but only once the underlying AI reaches a high enough capability threshold. This matters because it suggests AI systems won't be useful research assistants until they clear a specific skill bar, and that the feedback they get actually shapes what they do rather than just retrieving memorized patterns.
Why it matters
This is the first rigorous evidence that AI agents can genuinely adapt their reasoning in response to real experimental results rather than just recalling training data, but only if the base model is capable enough — it's a finding that cuts through both hype (AI can't learn) and doom (AI learns from anything), showing instead that there's a real capability threshold where feedback-driven learning switches on.