Google DeepMind says LLMs can prove theorems but still cannot make scientific “jumps”
kylekabasares · x · 2026-07-28
Google DeepMind argues LLMs still lack the “jump” needed for scientific discovery
The post discusses a DeepMind paper that reframes scientific discovery as a cycle: sensory experience → intuitive leap to axioms → logical deduction. The paper argues that current generative AI has largely mastered:
- Induction: pattern matching over massive data
- Deduction: formal reasoning and theorem proving
But it still lacks abduction — the non-logical leap that creates new explanatory hypotheses.
Using Einstein’s notes on discovery and his formulation of general relativity as a case study, the paper argues that today’s LLMs may be able to prove consequences from given premises, but are structurally unable to invent the premises themselves. It also claims that the bottleneck in artificial scientific invention is translating simulation into formal axioms, and suggests physically grounded multimodal world models may be needed to bridge that gap.
More from AGI Musings
- NeurIPS 2026 MATH-AI workshop opens papers on agentic mathematical reasoning — KaiyuYang4 · 2026-07-28
- Compute, not algorithms, is the real moat in frontier AI — GavinSBaker · 2026-07-28
- AI critics say the world mostly sees slop, not transformative value — yacineMTB · 2026-07-28
- AI is becoming a ladder of abstractions, from copilots to agents that choose their own work — pzakin · 2026-07-28
- OpenAI economic research says AI is reshuffling tasks across occupations — soumitrashukla9 · 2026-07-28
- Jensen Huang says AI destroys tasks, not jobs, and creates new work — garrytan · 2026-07-28