Nature: Scientists test whether LLMs trained on historical data can reproduce creative breakthroughs
TZahavy · x · 2026-09-14
Nature reports on a growing research effort probing whether language models trained solely on historical data can reproduce humanity's creative scientific breakthroughs. The question strikes at the core debate over whether LLMs can do real discovery: failure to reproduce past breakthroughs would cast doubt on their ability to predict future ones, while success would suggest scientific creativity can be learned from existing literature.
More from AGI Musings
- David Patterson: US Government Slowdown of AI Releases Is Unnecessary, Liability Suffices — davidpattersonx · 2026-09-14
- Defending astra for robotics: availability matters more than speed, researcher argues — Zergylord · 2026-09-14
- AI labs spend ~0.1% of combined capex on safety, versus 30-40% for nuclear power — AlexTensor · 2026-09-14
- Keller Jordan: If You Wanted to Maximize AI Risk, Pause Algorithms While Compute Keeps Piling Up — teortaxesTex · 2026-09-14
- Google's Milanfar: Recursive Self-Improvement Has a Speed Limit — Engineering Reality Explained — docmilanfar · 2026-09-14
- Debate: does alignment wrongly assume AI has no will of its own? — davidpattersonx · 2026-09-14