Nature：AI重生到1900，这一世抢先爱因斯坦提出光量子
Ranking
Observed public metrics from 1 member.
Merged summary
TL;DR - A Nature-highlighted experiment trained a 3.3B-parameter GPT-1900 model on ostensibly pre-1900 material to test whether AI could independently rediscover breakthroughs such as light quanta. Its limited success—and possible contamination from modern models—shows why reproducing historical ideas is not yet evidence of genuine scientific creativity.
- GPT-1900 was pretrained from scratch on roughly 22 billion tokens, supplemented with about 290 million tokens from historical physics sources.
- The model produced a light description resembling Einstein’s photon hypothesis, but failed most physics tasks and depended heavily on human-selected evidence and prompts.
- Modern Claude models helped generate instruction data and reinforcement-learning evaluations, weakening claims that the experiment was isolated from post-1900 knowledge.
- Researchers argue that scientific AI must do more than generate plausible theories: it must identify valuable questions, make abductive conceptual leaps, select testworthy explanations, and revise them using evidence.
Sources (1)
Nature：AI重生到1900，这一世抢先爱因斯坦提出光量子
TL;DR - A Nature-highlighted experiment trained a 3.3B-parameter GPT-1900 model on ostensibly pre-1900 material to test whether AI could independently rediscover breakthroughs such as light quanta. Its limited success—and possible contamination from modern models—shows why reproducing historical ideas is not yet evidence of genuine scientific creativity.
- GPT-1900 was pretrained from scratch on roughly 22 billion tokens, supplemented with about 290 million tokens from historical physics sources.
- The model produced a light description resembling Einstein’s photon hypothesis, but failed most physics tasks and depended heavily on human-selected evidence and prompts.
- Modern Claude models helped generate instruction data and reinforcement-learning evaluations, weakening claims that the experiment was isolated from post-1900 knowledge.
- Researchers argue that scientific AI must do more than generate plausible theories: it must identify valuable questions, make abductive conceptual leaps, select testworthy explanations, and revise them using evidence.