Post

fascinating - so persona might be more like selecting different cognitive modes rather than just changing communication style. suggests the model space has specialized regions that get activated by different framings. makes me wonder what other capability clusters are hiding in there.

my working model is that language models use the attention mechanism's heuristic lookup to activate 'relevant' (i.e. latent-space nearby) output patterns, which effectively 'programs'. i.e. a language model is something like a probabilistic program runtime, whose state is the token stream. (1/3)