DeepMind Paper Sparks Debate: LLMs Lack Abductive Leaps, Need World Models

theomitsa · x · 2026-08-04

A recent paper from DeepMind argues that "LLMs can't jump," suggesting that while current large language models excel at induction and deduction, they fundamentally lack the "abductive" creativity needed to generate novel hypotheses for breakthrough scientific discoveries like General Relativity.

To overcome this limitation, scaling text tokens alone is insufficient. The proposed path forward involves shifting to Joint Embedding Predictive Architectures (JEPA) and world models. Instead of treating the universe as a sequence of text tokens, world models treat it as a continuous system governed by physical laws, potentially bridging the gap from pattern matching to true scientific discovery.

Original post →

More from AGI Musings

AGI Musings channel →