October 10, 2026
AI models test hypotheses better than people, but generate them worse
Per Peter Gostev’s observation, models grab the first decent idea and check it thoroughly. Searching for alternatives is noticeably weaker than verification.
Peter Gostev
@petergostev
Models are superhumanly good at testing and validation, but because of that they are worse at research and idea generation. They just want to latch onto one decent idea and test it until it's dead. This is one of the blockers for using models in research and science, but judging by the results, Bell may handle it better.
· 3.6K views
Per Peter Gostev’s observation, models grab the first decent idea and check it thoroughly. Searching for alternatives is noticeably weaker than verification.
There are no measurements of idea search yet, though Gostev hints that Bell handles it better.