> That’s a statistical association not a concept. Try asking it questions that mix concepts like “Can you eat Apple share price?” which aren’t in its corpus.

ChatGPT: > No, you cannot physically eat an Apple share or any other stock share. A share of a company's stock represents ownership in that company and is typically bought and sold on stock exchanges. Share prices fluctuate based on various factors such as supply and demand, company performance, market conditions, and investor sentiment. While you can buy and sell shares of Apple on the stock market, you cannot consume or physically eat them.

That seems perfectly reasonable to me. The answer correctly identifies the problem with the question.

I would recommend looking up the Othello paper. Chess may be beyond the current level of LLM capability, but that doesn't mean they aren't manipulating things at a level higher than tokens.

Obviously it gets such a simple case correct, the grammar makes the subject clear. I was illustrating the approach using your wording for clarity, Chess was the actual example.

The Othello paper is hardly a counter example. Researchers created an Othello specific model that almost learned the grammar of Othello not how to play well. Yes, there was largely correct internal game state built up from past moves. No it didn’t actually learn the rules so it would make strictly legal moves nor did it learn to make good moves.

I don’t bring up this inaccuracy because it actually makes much of a difference to playing Othello, but rather to illustrate how these systems are designed to get really good at faking things. There’s approaches that allow AI to actually learn to play arbitrary games, but they differ by having iterative feedback rather than simply providing a huge corpus. It’s like science vs philosophy, feedback prunes incorrect assumptions.

Obviously you can use interactions with prior iterations to train the next iteration. But it’s a slow and adhock feedback loop.

> The Othello paper is hardly a counter example. Researchers created an Othello specific model that almost learned the grammar of Othello not how to play well. Yes, there was largely correct internal game state built up from past moves. No it didn’t actually learn the rules so it would make strictly legal moves nor did it learn to make good moves

It is, though. Nobody said anything about playing well, or learning the rules. The very fact that it had a valid internal representation of the game state means it's extrapolated beyond token-level. Which is the point.

> The very fact that it had a valid internal representation of the game state means it's extrapolated beyond token-level. Which is the point.

The paper said it was making incorrect moves thus it has an invalid representation of the game.

So an LLM when specifically trained on Othello, a game with very simple and completely mechanical rules, failed to abstract what those rules actually where. This means at a purely mechanical level it doesn’t understand the game when that was exclusively what it was trained to do.

It’s a clear illustration that these things are really really bad at abstraction. But that should be obvious because they are simply manipulating arbitrary tokens from their perspective. It doesn’t intuit that the game should have simple rules and therefore it doesn’t find them. People on the other hand have a real bias regarding simple rules.

Except that incorrect moves don't imply an incorrect representation.

The internal representation was a literal 8x8 grid of piece locations they could externally change and have it generate moves consistent with the changed position. It's about the clearest example of a learned higher-level internal representation I can think I've seen.

The fact that it didn't also perfectly learn the rules while it was doing that is entirely uninteresting.