KAN author Ziming Liu: LLMs learn alchemy, not the science of AI

ZimingLiu11 · x · 2026-10-24

Responding to a debate on whether LLMs can do AI research, KAN author and MIT researcher Ziming Liu offers an alchemist analogy: today's training tricks work like potion-making—heuristic tricks learned from data beat random chance, but they are correlations, not the underlying causal mechanism (like rain helping a potion when water vapor was the real cause).

He extends the point: an LLM trained in Aristotle's era would know ancient Greek but not modern math or physics, because those languages hadn't been invented yet. Similarly, he argues, we haven't yet invented the language of a "Science of AI"—so LLMs couldn't have learned it. An accompanying three-panel comic illustrates how empirical heuristics → hidden correlations → scientific understanding enables systematic control.

Original post →

KAN author Ziming Liu: LLMs learn alchemy, not the science of AI

More from AGI Musings