Yeah - but what is it adjusting weights based on? It's not based on next token ....

And per the focus of this thread, regardless of how accurate it is, why do you find "next token predictor" to be the most useful mental model?