Stop Thinking of LLMs as Next-Token Predictors
4 points by square_usual
4 points by square_usual
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.
No one in this space considers any deep learning model, generative or not, to be "imitations" of training data.
Many lay people do. I think that's the easiest steelman interpretation of this article.
I don't even know why I'm bother commenting when this whole post was generated by an LLM.
Funny, this morning I complained about people who can’t see the LLM writing for what it is. So if you’re right, now I’m that person. I think this doesn’t read like LLM prose, it reads like moderately polished human writing, which is quite different.
LLM writing sounds superficially clever, but uses a tremendous amount of words that say very little. It tends to misuse specific words when ordinary ones would work, shallow attempts to build suspense, and “eyeball-kicks”. I don’t see that here.
I could be wrong though. What makes you think this is LLM written?