Not a great impression to have your demo video demonstrate how one of your 'most advanced' AI models loses to the most common check-mate pattern in all of chess.
For an LLM, just being able to play an entire game of chess without illegal moves and without inventing pieces that aren't on the board is an achievement. Even more so for a live model. Then again, who knows how much the harness helped here
Seems more than good enough for a live model though! I can imagine this demo being extended to be a lot nicer to play with. You can just feed the model engine analysis and it can make as high of quality moves as needed. No longer any correlation between the model's understanding of the position and the moves that would be made but I think that's still a really nice improvement when thinking about this as adding live voice interaction to existing chess vs computer functionality rather than adding chess to possible interactions with the latest live voice model.
Genuine question, from someone with a non-math background - at what point does the constraint become the number of unsolved conjectures remaining, instead of the ability for LLMs to actually solve for one?
Math is very hit-or-miss; the complexity of a question does not give much of an indication about how complex the answer will be. Look up the formula for solving a degree-4 polynomial equation to get a purely visual idea how this can look (and then degree-5 suddenly forces you to use complicated new functions). And there are problems (like the Collatz conjecture or P vs. NP) that there doesn't seem to be any promising angle of attack for over at least decades.
I would wager that this is a fundamental part of the structure of math that has been a constant from ancient Greece till LLMs. There are even some formal results, similar to Gödel's theorems, that say that the maximum necessary length of a proof grows arbitrarily fast (e.g. more than exponentially, double-exponentially, or any function with a formula) with the length of the statement being proven.
Point is, math will most likely never suffer from this particular problem.
Not entirely following the question, but there are an infinite number of conjectures, the blinding majority of which serve nearly no purpose to humanity. Consider that for every executable program one could create a conjecture, and therefore a mapping exists from executable programs to conjectures. Now, consider the infinite possibility space of executable programs…
Anyway, unless you mean conjectures that humans have already posited, or ones that are particularly famous, that list is much shorter, but also contains conjectures that I’m not convinced can be solved before the heat death of the universe using all available compute power. P != NP is a conjecture, for example. Also a lot of prime number conjectures that are extremely computationally expensive.
the aligner already supports french (it's meta's massively multilingual speech model under the hood), so it would all work today, but i started on standard ebooks, which is english-only.
but gutenberg and librivox both have french libraries so i could absolutely do that next!
reply