Benedict Evans

99 times, you know, nine followed by many zeros.

Does it actually understand or is it just always right without understanding?

Does it actually know that's original and different or does it not?

And that's a kind of a puzzling, I don't think we know the answer to that and it may be the wrong question, but it's kind of a puzzle as to how far can these things make things that are both different from the training data and good?

And is knowing that this is different but good, is it really different or is it just matching the pattern on a longer frequency, if you see what I mean?

And how much could you actually have predicted that given enough data that it's not actually outside the pattern?

It just kind of looks like it is if you're zoomed in more.

And if you zoom out more, then that is matching the pattern.

How would you know that people will, it's like, you know, thinking about music, like how would you know that people would like punk?

You can very easily imagine a generative AI system that can make you more stuff that sounds like Yes or more stuff that sounds like Pink Floyd.

It might not sound like good Pink Floyd for Floyd, but you could imagine it would make more stuff that sounds like The Grateful Dead.

You know what Grateful Dead fans say when they run out of drugs?

This music's terrible.

You can imagine, I'm being unkind, but you can imagine the challenges knowing that now people are really fed up of 70s prog rock and they would really like something else and that something else would be punk and that would work.

It's knowing that people in the 40s were really fed up of the war and would want luxury and that Christian Dior's new look would work and would express that.

Could an LLM do that thing?

How much variance do you need?

I don't know.

It's an interesting thought experiment to ask that question.

There's a completely different place to take this, which is to say this is an appeal for boutiques and in-person events and the unique and the curated and the individual.