How Far Are We From An AI Einstein? - Adam Brown

Adam Brown (Google DeepMind/Stanford) argues AI deriving general relativity from Newtonian physics is the terminal test of human-level intelligence, possibly 10 years away.

• Brown treats AI independently deriving general relativity from 1900-era physics as the definitive benchmark for human-level intelligence.
• He estimates this could happen within 10 years, calling general relativity humanity’s greatest intellectual leap.
• His private Stanford GR graduate exam eval: models scored zero 3 years ago, weak-student level 1 year ago, now essentially ace it.
• LLMs have read more than any human ever will, yet still underperform what a human with equivalent knowledge retention would likely produce in novel connections.
• Like chess engines that search more positions but evaluate less naturally than humans, LLMs have lopsided strengths relative to their raw knowledge base.
• Physics problems have two steps—translating words to math, then solving math—and LLMs now handle both well on standard graduate coursework.
• As LLMs have strengthened, evaluating them requires hiring PhDs to craft new problems; scraping internet math problems no longer suffices.
• Whether LLMs think natively in higher dimensions is a misconception; mathematical notation, not spatial intuition, is the real tool—same for humans and models.

2025-01-01 · Watch on YouTube