DeepMind says LLMs still can’t make scientific leaps, and Tao warns of proof overproduction

APPSO · wechat · 2026-07-29

This long WeChat article argues that today’s LLMs can do two things well — pattern matching and formal deduction — but still struggle with the third step of scientific invention: making a genuine conceptual “jump.”

It uses Google DeepMind researcher Tom Zahavy’s position paper “LLMs Can’t Jump” to frame the issue through Peirce’s three modes of reasoning:

The article contrasts this with Einstein’s elevator thought experiment and the birth of general relativity, arguing that scientific breakthroughs require embodied simulation and a leap to new assumptions — not just more data compression. It also notes DeepMind’s claim that even video models predicting an apple falling still do not internalize physics; they only extend pixel statistics and lack counterfactual control.

The second half turns to Terence Tao’s ICM 2026 talk on AI and mathematics. Tao assumes AI will soon solve a large share of research-level math under human supervision, then asks what mathematics is for beyond solving problems. He warns of a future of “proof overproduction,” where verification, explanation, and integration into the field become the bottlenecks. His recommendation: reward not just the first solution, but also checking, exposition, and synthesis.

Related event: DeepMind Paper: LLMs Lack the 'Intuitive Leap' for Scientific Discovery(7 posts)→

Original post →

More from AGI Musings

AGI Musings channel →