> The thing is it's like a DJ mixing records it could generate some amazing new mixes but if the pattern is not already out there it's very unlikely to find new patterns.

What does this mean in practice?

Hypothetically, let's say I ask a future (1-2 years out) model to write me a brand new fantasy book series, and tell it what all my favourite books are - and it writes me something that is stellar, 5/5. If someone comes to me and says "yes but is this TRULY original?" What does that even mean?

I think some people are very confident that LLMs cannot find new ideas, but I don't know where they get that confidence from - LLMs have continuously exceeded the thresholds proposed by their critics, and now it feels like we're getting into the esoteric. It's a bit of a... God of the gaps situation to me.

Hypothetically, let's say a language model solves a math problem that has never been solved before - would that change your mind? Do you think that's even possible?

I mean this would be a great point if ai wasn’t being used for new research where there’s not a lot of patterns, or if people were original but if you read or watch a lot you see most stuff is already just remixes of something else