alt Hacker NewsThe fact that a model trained on the internet, on which the correct rules of chess are written, is unable to determine what is and is not a legal move, seems like a sign that these models are not reasoning about the questions asked of them. They are just giving responses that look like (and often are) correct chess moves.
It's a sign that they are 'reasoning' imperfectly. If they were just giving responses that 'looked like' chess moves, they would be very bad at playing chess.
(And I would hazard a guess that they are a primarily learning chess from the many games that are posted, as opposed to working things out from the rules. Indeed, if you make up a game and tell chatGPT the rules, it tends to be even worse at following them, let alone figuring out optimal play. But again, it will do so significantly better than random chance, so it's doing something with the information you give it, even if it's not doing so very well. I think it's reasonable to call this thinking, or reasoning, but this mostly becomes an argument of semantics. either way they do it significantly better than random chance but still not tremendously well. If your expectation is that they cannot work with anything novel then you're going to be continually surprised, but if your expectation is that they're as good as a human that has 'learned' from all the material its been given, especially material that's in-context and not in the training data, then you're also going to be disappointed.)