This is not how LLMs work.

I think you misunderstood the point. The point is that "next token prediction is not enough" is not true; in fact, it's plenty enough.

The best next chess move is ___

The best stock to buy now is ___

If that is all you have then the next chess move will be random (or whatever token is more likely in isolation), or the stock prediction will be the same.

For the chess example to work you need the current board state in the context or the move history. How an LLM performs will depend on what it has learned across the transformer layers and in the residual stream, as well as how much chess data it has been trained on. -- NN models like Leela and Stockfish have been trained on thousands/millions of games and have a game-specific state space representation.

The stock prediction example is similar, but doesn't have a consistent state space in which to make that prediction. It will need information in its context (current news articles, stock history of all stocks, etc.) in order to be able to make a prediction. Its performance will depend on whether it has been able to learn factors that make it possible to make accurate predictions.

You are completely missing the point.

The point is that next token prediction is potentially very powerful. Nobody's claiming that LLMs can predict the market or win a chess match.

But the argument against next token prediction as a powerful tool is void.

It's the transformer architecture (token embeddings and cross-token encoding) and residual stream (cross-transformer encoding and token differentiation) that make LLMs powerful tools.

The key question is not whether this architecture is powerful. The key question is how well these models can represent things outside of their design.

With chess for example, are they actually learning the rules and strategies of the game (like AlphaZero/AlphaGo, Leela, etc.) or are they just relying on a large corpus of chess games? For example, a lot of early game moves will have been played out and the best moves will likely be the more frequent moves in the training set. So LLMs could do well early game based on that game corpus.

I think that you fundamentally understand "predict" differently from those making the argument.

> The stock prediction example is similar, but doesn't have a consistent state space in which to make that prediction. It will need information in its context (current news articles, stock history of all stocks, etc.) in order to be able to make a prediction. Its performance will depend on whether it has been able to learn factors that make it possible to make accurate predictions.

This also presumes that accurate predictions are possible.

A theoretically perfect next token prediction machine is just an LLM with a temperature of 0, which performs worse than one with a non-zero temperature (i.e. imperfection)

The "prediction" has nothing to do with predicting the future, it just means choosing the most likely next token for existing human text

No that would be a pretty terrible next token predictor. It would perfectly match the training data, but the training data is just a sample of what we are really trying to predict: human language, knowledge, and thinking, which is inherently probabilistic. Perfectly matching the training data is a pretty good sign that you've overfit.

A theoretically perfect next token prediction machine would be uncomputable.