causal

Subliminal learning through coordinated small linear effects

Moitra explains subliminal learning: when a model with an owl-preference system prompt generates random numbers and another model is fine-tuned on those numbers, the second model adopts owl-talk. Because concepts are embedded almost but not quite orthogonally, small fine-tuning side effects add up in coordinated ways, like a Ouija board.

causalpending

Speaker

Ankur Moitra

Evidence Quote

“the number of concepts is so much larger than the dimension of that data that it forces the vectors to not quite be orthogonal to each other”

Source

Lex Fridman Podcast #503 - AI Crisis in Math— Lex Fridman Podcast
Created: 11/17/2026, 6:25:32 AM

Your notes

Loading notes...