AI agents still fall short of scientists, yielding far fewer insights
typewriters · x · 2026-10-03
A research thread argues AI agents cannot yet discover mechanisms that generate new scientific insights the way Isaac Newton did. Even the strongest agents approach human predictive accuracy, but the mechanisms they uncover yield far fewer insights than those found by human scientists, and many agents still fail common-sense scientific constraint tests.
More from AGI Musings
- Ben Horowitz: a personal agent's biggest win was canceling someone's NYT subscription — gethackteam · 2026-10-03
- repligate: even GPT-3 was probably conscious, and it makes everything click — repligate · 2026-10-03
- Andy Masley: we owe humans positive duties and conscious AIs negative duties — AndyMasley · 2026-10-03
- If LLMs are conscious, you must accept clonable, pausable, reversible experience — argued again — wordgrammer · 2026-10-03
- Frontier LLMs seem pricey today, but 100x cheaper models in a year will reshape everything — NickPassig · 2026-10-03
- Debate over EA discourse: are mild philosophy arguments too edgy to say publicly? — AndyMasley · 2026-10-03