Comment by YeGoblynQueenne
3 hours ago
For me the useful intuition is that LLMs haven't somehow magickally learned to implement any of the algorithms we know that we have used to make strong chess engines: alpha-beta minimax and Monte-Carlo Tree Search on the one hand, and obviously the ability to learn accurate evaluation functions by self-play.
I mean we've done all this before in a task-specific fashion. It's useful to know that LLMs haven't managed to do that in the process of learning to represent the entire text on the web. On the other hand they have gotten say very good at machine translation without being trained exclusively (and I select the preceding word carefully) on machine translation.
Edit: I'm saying this because there is this idea expressed by e.g. Ilya Sutskever, that in order to predict the next token accurately an LLM has to learn something about all of underlying reality. See for example this interview with Dwarkesh:
https://x.com/biobootloader/status/1640512444958396416
Where Sutskever claims that "Predicting the next token well means you understand the underlying reality that led to the creation of that token".
If that were true, we should have seen LLMs play good chess by now. There is a huge amount of data on playing chess floating around on the web in the form of algebraic chess notation and if LLMs were capable of learning the "underlying reality" of chess, they would already have. They haven't. Because they can't. What Sutskever is saying flies in the face of literally hundreds of years of statistical modelling, which is to say, building predictive models that, very explicitly, do not have to understand any "underlying reality" and only have to be good at modelling a dataset.
>If that were true, we should have seen LLMs play good chess by now.
Not at all. LLMs learn by imbibing a mass of relationships as isolated fragments of information. There is a certain amount of sorting and indexing that happens during the training phase. There is also a certain amount of compute executed on these relationships during inference. LLMs can model processes that fit within the compute budget. Language translation works well because language is lookup-heavy while being light on compute.
Chess is a compute heavy game of finding the best move out of many possibilities with wide variation in the quality of each move. Humans cut through the compute requirements by reinforcement and learning intuition. LLMs don't get reinforcement on chess so they must compute during inference a unified model of chess. Developing a strong model of chess from raw fragments of information is simply not in their compute budget.
But it speaks in words, therefore it must be super duper extra smart!!11 /s
Sarcasm aside, I think this is an easy cognitive trap to fall into. It does sometimes feel like the LLM must have some world model because it converses somewhat coherently. Examples like this failure to understand chess, or to count the number of Rs in "strawberry", seem difficult to explain if the models are intelligent. But that doesn't stop people believing they are anyway. I think there must be something about the conversational interface that fools us easily. I wonder if people trained in interrogation techniques are also fooled?