i've been thinking about this a lot. working on a longer form piece on this idea:

i find it's actually easier to "discover" than it is to "create" using llms. discovery, especially in math, is mostly a question of throwing compute at a problem and trying different results. there's a verifiable answer at the end. but there's no way (afaik) to aim something you're training via RL at a creative problem and say "make the next lord of the rings". taste literally changes by the day, and so much of what makes art matter depends on cultural significance you can't specify in advance. if you were training a music model in 1915, it would never come up with rock and roll. it's so hard to know what's going to resonate. van gogh was making his work for essentially nobody and now it's worth millions. you can't pre-load a success criteria for that

i think this is important to call out in a world that seems to devalue creative work now. people think llm generated "artifacts" are creative but in reality it's garbage and based on pattern recognition