Hasabis' 'Einstein Test': can AI rediscover relativity from 1911 knowledge?
量子位 · wechat · 2026-09-19
Nobel laureate and Google DeepMind co-founder Demis Hassabis proposes an AGI test: restrict an AI to 1911-era knowledge and see if it can independently produce general relativity by 1915. Nature dubbed this the 'Einstein Test.'
Attempts so far: GPT-1900
- Independent researcher Michael Hla trained a 3.3B-parameter model from scratch on pre-1900 books and newspapers (22B tokens), scrubbing any texts mentioning modern concepts
- On a photoelectric-effect question the model produced a passage eerily similar to the light-quantum hypothesis, but it failed most physics tasks, and modern Claude models were used for instruction generation and scoring — so 'zero contamination' can't be guaranteed
What AI still lacks
- DeepMind's Tom Zahavy splits scientific reasoning into induction, deduction, and abduction; current models master the first two but lack the 'abductive leap'
- MIT's Jacob Andreas argues the real divide is judging which theory is worth betting a career on — finding the right question without anyone handing you the answer
Bottom line: reproducing the right answer doesn't make AI a scientist.
More from AGI Musings
- There is no degree collapse, argues exec: lifelong learning beats a 4-year degree — deepakns · 2026-09-19
- When 'why are you talking about this?' is your best rebuttal, Cofnas wins more rationalists — LucaAmb · 2026-09-19
- Jan Kulveit Pushes Back on 'Context Is the Right AI Identity' Claim — gleech · 2026-09-19
- Blogger argues AI panic is a fear-based control narrative, not the models themselves — realmeetjames · 2026-09-19
- How do you ban superintelligence when nobody can even define it? — mallow610 · 2026-09-19
- ICLR 2027 hits ~50,000 abstract submissions, up 2.5x from last year — The Decoder · 2026-09-19