Ulrike Hahn on Nostr: I’m saying that any kind of NN whose target task is predicting what happens next ...

I’m saying that any kind of NN whose target task is predicting what happens next (for example, next token prediction in the context of LLM training, or a recurrent network a la Elman, or the Bayesian brain hypothesis) elides the difference between an “external guide” and self-driven adjustment.

the whole point about those tasks, conceptually, is that they provide teaching without a teacher….