Stop Thinking of LLMs as Next-Token Predictors

4 points by square_usual


drakairos

I don't think I will. RL just tunes it to predict next-tokens to complete tasks, regardless if the training data is natural (pre-existing text) or synthetic (rewarded generation). I would consider both chess engines in the analogy "next-move predictors".

No one in this space considers any deep learning model, generative or not, to be "imitations" of training data. They have always been considered universal functions that learn an approximation of some real-world function. In the case of LLMs, regardless of training, is f(current_text) = next_token. We say "predict" because that's what you call the output of an approximation of a statistical function.

I don't even know why I'm bother commenting when this whole post was generated by an LLM.