Missing Paper: Learnable Nature Conjecture

Type: gap Slug: gap—learnable-nature-paper Sources: nobel-prize-lecture-accelerating-scientific-discovery-with-ai—hassabis Last updated: 2026-05-13


What’s missing

The 2024 Nobel Lecture articulates a conjecture that many natural laws are learnable by ML from data alone, but no peer-reviewed paper formalises or tests this claim. As of May 2026, this remains the most important idea in the corpus without a proper scholarly home.

Why it matters

If true, the conjecture reframes the relationship between AI and science: AI is not merely a tool but a discovery engine. If false, AlphaFold’s success is domain-specific and does not generalise.

What a paper would need

  • Formal definition: what counts as “learnable”? What is the failure mode?
  • Negative cases: domains where ML fails to discover regularities despite adequate data
  • Comparison: learnable vs. physically-simulated approaches on the same problems
  • Theoretical framework: connection to statistical learning theory, computational irreducibility

Connections

  • Claim: claim—learnable-nature-conjecture
  • Theme: theme—AI-for-science

Honest Gaps

  • The Nobel Lecture extraction is only ~8K chars — Hassabis may have articulated the conjecture more precisely than what is available.
  • It is possible that a paper is in preparation at Isomorphic Labs or DeepMind but not yet published.