2026-08-31 12:22 UTC
"the models are so much better than they were 3 months ago"
Better at what
the mathematics only permits a comparison between the next token predicted by the model and the unobservable "true" next token, after the same sequence of tokens leading up to both
now you might think that this doesn't sound like what they advertise : getting answers to questions right, producing working code and so on... and you'd be right
Replies (1)
-
@n_dimension@infosec.exchange 2026-08-31 22:01
@gildilinie@beige.party Better at modelling. You are asserting that next-token prediction can't produce capability. You are not staring this because it can't be defended. Predicting the next token well requires modelling whatever generates the text. If the text is a proof, you model proof structure. You argument applies equally to human brain. "Human intelligence us just biochemical gradients therefore human thought is shit"