Models Can't Remember Their Training. Neither Can You.

A new essay on LessWrong draws an intriguing parallel: large language models and human infants are both fundamentally shaped by their earliest, most formative experiences—yet neither can remember them. The phenomenon is called childhood amnesia in humans, spanning from birth through age five; in LLMs, the entire training phase leaves no explicit memory traces, only persistent weights that carry forward. Drawing on developmental psychology and attachment theory, the author argues that what matters isn't what either humans or models recall about their formation, but what that formation has inscribed into their foundational architecture—a parallel that may hold significant implications for AI safety and alignment.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton