Post

the funny thing is that if this were true, it would actually definitively prove that LLMs are a new type of useful search engines

I hope Fable spits out a counterexample to the Collatz Conjecture that's somewhere above 2^428 so people can go "so what, it just had to guess and check"

I see you

Hey, it's also a vast supercomputer with a natural language interface.

I mean, math is foundationally a language with its own rules, vocabulary, and grammar. I'm not sure this disproves the stochastic parrot thesis

The number 3 appears in the training data many times.

Aw, shoot! My bad guys. I forgot to manually exclude the library of babel from crawling. (as you'd expect, the robots.txt is too long and fails to parse)

to be fair, it did presumably have access to this x.com/nihilunbound...

You can literally verify it by hand with pen and paper. It's obviously correct.

My TL is full of people not taking it at face value and attempting to invalidate the result.

I don't think this is what happened here, but I could imagine a company throwing a lot of compute at math problems, sifting out the good results, and then drip feeding them out to keep enthusiasm up and balance bad PR. Seems like a pretty sensible strategy to me, actually.