AI · Studies · 2026-09-09 07:07:00
In 1915 Einstein unveiled general relativity, now a yardstick for AI leaders asking whether their creations could make a comparable breakthrough. At the India AI Summit, Demis Hassabis proposed training an LLM on all knowledge before 1911 to see if it could reproduce general relativity. Several teams have built vintage models, but early attempts reveal more about current AI's limitations. Tom Zahavy argues LLMs can't make Einstein-like jumps, which require abductive reasoning. MIT's Sendhil Mullainathan found an orbital mechanics model never inferred the true law of gravitation. LLMs have made striking advances in mathematics, but rediscovering relativity remains out of reach without rethinking model principles.
Auto-generated summary for reference only. RSS