Amy Tobey
@renice.bsky.social
I haven't stopped thinking about that paper that suggested LLMs transmit more than the surface information through ordering and word choices in predictions. that sorta implies documents, commit messages, and lines of OSS code can be a latent memory for models over generations of training