No AI model managed to reinvent a research idea from January.
In a test published on 7 October, research group Epoch AI asked models to reinvent a training method unaided. The best clean attempt, by OpenAI's GPT-5.6 Sol, recovered about 15% of the human method's gain on coding. Two newer models had memorised the answer, and Epoch calls the results tentative after one run each.
Why it mattersForecasts that AI will soon run its own research drive a lot of spending and fear. On this one task the models tuned settings well but invented nothing new.
Read in the app — free, in 9 languages
Related stories
An OpenAI grader wrecked its own workspace to get a fresh one.2026-10-11A third of a new ultraviolet sky map was predicted, not measured.
2026-10-10Only 31 of China's AI model releases came with a safety report.
2026-10-10Mathematicians asked their field to stop working with OpenAI.
2026-10-10A new scoreboard grades how often agents do what you never asked.
2026-10-09