"Use LLMs for Shannon work, humans for Kolmogorov breakthroughs Analyzing how Donald Knuth used LLMs to make progress on a hard combinatorics problem, the speaker offers a precise decomposition of human-AI collaboration: LLMs are extraordinarily good at exhaustive search across a known solution space — Shannon-style correlation finding — while humans are needed to synthesize those findings into a new compact causal representation. "The LLMs were extremely efficient at doing the Shannon part of it... but eventually he used the solution and he came up with the proof" — meaning Knuth had to create the new conceptual framework. The practical implication: use AI to rapidly explore and surface patterns in a well-mapped domain, then invest your own cognitive effort in the synthesis step that produces a genuinely new model of what is happening. Source: Vishal Misra, No Priors (Martin Casado), 'How LLMs Actually Work: Bayesian Inference, Causality, and the Path to AGI'"