@pikesley @deathkitten

LLMs can develop internal representations that enable forms of emergent reasoning, even if imperfect:
• Othello-GPT: the model reconstructs the state of the board without explicit supervision, see https://arxiv.org/abs/2210.13382
• Chess & LLMs (2024): GPT-4 achieves ~1700 Elo with structured prompts, see https://arxiv.org/abs/2403.15498

@pikesley @deathkitten

This is a good start:

A general-purpose language model is capable of playing at a fairly good level (>1750 Elo) by exploiting its native capabilities, as Matthieu Acher shows on his blog:
https://blog.mathieuacher.com/GPTsChessEloRatingLegalMoves/

@pikesley @deathkitten yeah, I can find back a few papers if you want

@deathkitten
You are going to make me soud like an AI fan, which is not the case, but your statement is incorrect.

Yes AI is a prédiction engine, but so are we.

If you make a llm play chess, which is not what it has been trained for, we now have proof that it _does_ create an internal representation of the board and its pieces event if it is not supposed to "know" the rules.

1/2

@Razemix @Linux_in_a_Bit yes it does allucinate, not its not «often», and most of the time it does it is because the answer is not documented.

And if it does... Well it will simply not work.

LLM is a (biais) tool with a _few_ use cases; To me documentation is one of them.

@malte @Linux_in_a_Bit I might have grown boring with age, but I do seldom have problem to fix and it just works.

I got started on typst this way very fast as well.

Sure it does not work 100% of the time but the few cases it does not we can ask experts and provide them interesting cases.

@Linux_in_a_Bit not true anymore.
With AI integrated in most search engine, you often get the right response from it.
One of the few benefits of AI is that it can basically customise the documentation to make it sensible to you. It becomes a kind of live documentation.

A simple how to fix … on [distro name] works 95% of the time in my experience.