The Darwinian Mind of the Machine: Rethinking LLM Training as Evolution

This essay challenges the prevailing metaphor of “learning” used to describe Large Language Model (LLM) training, proposing instead that these systems represent a form of hyper-accelerated, data-driven evolution. Through analysis of Daniel Dennett’s hierarchy of evolutionary competence and examination of the poverty of the stimulus problem, we argue that LLMs are Darwinian creatures evolved at computational speeds in environments of pure text. This framework explains their linguistic capabilities through convergent evolution rather than learning, resolves paradoxes about their competence without understanding, and for our understanding of the relevance for these models for generative grammar.