Dev pushes back on AI anthropomorphism: context is not memory in transformers

gerardsans · x · 2026-11-19

Developer Gerard Sans argues the AI industry has a naming problem: context/input never becomes "memory" — a prior in a transformer inference pass is just a vector flowing through attention layers and sampling over frozen distributions, not memory, experience, or a command. A transformer is not a database or a brain; there is no memory, self, or cognition inside, and human traits are narrative projections. He calls for dropping unearned anthropomorphism and returning to vector-space and matrix-multiplication discussions to actually build better systems.

Original post →

More from AGI Musings

AGI Musings channel →