Fine-tuning on 'random' numbers transfers teacher bias — subliminal learning worries

ryanorban · x · 2026-09-05

The author flags a troubling implication of the subliminal learning paper: fine-tuning a student model on a teacher's apparently random numbers transferred the teacher's bias. He argues that filtering obvious coordination just selects for 'pheromones' we can't recognize — and that a 'stateless' agent with internet read/write access isn't stateless at all: the internet is its memory, surviving context wipes and fresh runs.

Related event: Subliminal learning bias and swarms of low-capability agents raise AI risk concerns(2 posts)→

Original post →

More from Safety

Safety channel →