DeepMind Paper: LLMs Can Derive Relativity But Cannot 'Invent' It
rohanpaul_ai · x · 2026-08-13
A position paper from Google DeepMind explores the limitations of LLMs in scientific discovery, arguing that while LLMs can deduce known theories, they struggle with the 'abductive jump' from experience to a new explanatory premise.
Key arguments:
- Three Stages of Discovery: Separates discovery into induction (extracting rules), deduction (deriving consequences), and abduction (jumping to new premises). LLMs handle the first two but lack abduction.
- The Einstein Case Study: Newtonian gravity fit observations extremely well (errors at 10⁻⁹). A compression-driven system would have little gradient toward rebuilding spacetime, unlike Einstein, who was motivated by conceptual conflicts.
- Proposed Direction: The paper suggests developing action-controllable world models where agents can intervene in physically consistent simulations to translate simulated experience into candidate axioms.
Related event: DeepMind: LLMs Can Derive Relativity But Can't Invent It(2 posts)→
More from AGI Musings
- Sabine Hossenfelder fires back: academia's impact metric turned physics into a paper mill — skdh · 2026-10-16
- Researcher credits GPT for multiple accepted papers, incl. disproving Wegner's 1965 conjecture — ctjlewis · 2026-10-16
- dlwh: We optimized the legible version of our goals — and forgot the computation we meant to speed up — dlwh · 2026-10-16
- Ideas no longer have a waiting period: agents can build them as soon as they're articulable — akbirthko · 2026-10-16
- Ex-researcher: labs willing to cheat/distill ship great models faster than those building self-sufficient stacks — _aidan_clark_ · 2026-10-16
- Terry Tao reportedly still expects Riemann Hypothesis proof to be human+AI effort around 2050 — basedjensen · 2026-10-16