Ask HN: Can LLM's generate a sudoku that has not been created before?

[bnormative]

or can a LLM generate a sudoku which it has not been trained with?

LLM's are neural networks that are trained with a massive amount of data. Their training data can contain all the publicly available sudokus that in the internet.

So, can it really be creative enough to come up with any sudoku that it doesn't even see before?

I would expect that they could not because they are not trained with. Because they are just next word predicter.

[krapp]

It doesn't take creativity to do that. There are tons of sudoku generators already, all you would need to do is run one until it produced a novel result.

Novelty isn't equivalent to creativity. If it were, a random number generator would be the most creative force in existence.

[bnormative]

However, consider LLM's and the way they actually work. If I want it to create a unique LLM, is it intelligent enough to come up with one that it is not trained with? Does it have enough understanding to compare all the other ones it is trained with and come up with a new one using some probabilistic randomness?

[krapp]

I would assume it could come up with one it isn't trained with, but a correct, much less novel, result would be due to random chance, not an intelligent awareness of what a sudoku is, and what sudokus already exist.

It might just as well copy half of a New York times sudoku, or make up entirely nonexistent words. As far as I know LLMs can't actually do the math necessary to generate a correct sudoku. It would just give you some "sudoku-like" text.

I'm not an expert on LLMs but based on what I do know, that's what would happen.