The Einstein test: Nature asks if AI can rediscover general relativity on its own
mikeflache · x · 2026-09-20
Philip Ball's Nature feature probes whether language models trained on historical data can make breakthroughs on the level of Einstein's general relativity.
- The idea stems from Google DeepMind's Demis Hassabis, who proposed at the India AI Summit training an LLM on all knowledge predating a cut-off date to see if it could independently re-derive gravity as curved spacetime.
- The piece frames a retrospective benchmark testing whether AI can move beyond accelerating pattern recognition to formulating radical concepts like the equivalence principle.
- Ball argues scientific creativity lives in long conceptual reasoning chains, so evaluating AI this way pushes past equation-solving and data ingestion toward measuring genuine originality.
More from AGI Musings
- After 15 years of better dashboards, agents can now execute finance transactions themselves — LexSokolin · 2026-09-21
- If your checkout isn't agent-friendly, agents will just recommend your competitor — jeff_weinstein · 2026-09-21
- AI is 3% of GDP by spending, yet 80% of today's GDP hinges on it — moderate scenario — teortaxesTex · 2026-09-21
- Kevin Roose: "Stochastic parrots" and "marketing hype" claims were disprovable by 2023 — kevinroose · 2026-09-21
- "We are feeling machines that think": Damasio's take on bias resonates — pwlot · 2026-09-21
- Will 60,000 ICLR Submissions Kill Academia? Insiders Say Reputation, Not Papers, Rules — JundeMorsenWu · 2026-09-21