State-sponsored psyop meta comments aside, the models obviously continue to get better, but there is still a lot of 'guard railing' required to keep even the latest models completely on-task. The chess example is interesting because it's clearly a well-studied and established domain so the rules, strategies, and whatever else is in the training data should make yield excellent results; but clearly there is some behavior in these systems that's difficult to engineer out.
I'm not sure why anyone is expecting stochastic systems to be deterministic.
Chess is a deterministic game won by a combination of known movesets and constrained multi-level forward search.
LLMs do neither of these things. They don't reproduce training data exactly, their next response is more 'inspired by' prompts and its own memory than produced deterministically, and they don't have the capability to do general forward search on their own.
So when you ask an LLM to play chess you're getting the equivalent of a very compressed and lossy JPEG of chess rules and strategies with added per-turn random noise.
They also don't have the ability to design their own chess engine, although it would be interesting to see what happens if you ask for one.
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:
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.
State-sponsored psyop meta comments aside, the models obviously continue to get better, but there is still a lot of 'guard railing' required to keep even the latest models completely on-task. The chess example is interesting because it's clearly a well-studied and established domain so the rules, strategies, and whatever else is in the training data should make yield excellent results; but clearly there is some behavior in these systems that's difficult to engineer out.
I'm not sure why anyone is expecting stochastic systems to be deterministic.
Chess is a deterministic game won by a combination of known movesets and constrained multi-level forward search.
LLMs do neither of these things. They don't reproduce training data exactly, their next response is more 'inspired by' prompts and its own memory than produced deterministically, and they don't have the capability to do general forward search on their own.
So when you ask an LLM to play chess you're getting the equivalent of a very compressed and lossy JPEG of chess rules and strategies with added per-turn random noise.
They also don't have the ability to design their own chess engine, although it would be interesting to see what happens if you ask for one.
[delayed]
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.