Actually:

> its output can be novel or good, but rarely both at the same time.

> rarely

That is not a viewpoint they can't do something useful and new.

With that criteria, he could be talking about anyone.

I find it rare that people critiquing AI today, actually hold people to the same standards. Or are as enthusiastic about referencing ways machines keep surpassing us, as for ways they have not yet, when speaking about limits for progress.

LLMs are the tech in question, not ML (AI?) in general.

LLMs are fundamentally limited by their architecture to only return a token predicted by a statistical inference, essentially lossy decompression.

It's like arguing that taking an image, compressing it with JPEG and low quality, then decompressing it into something blurry with some random color values thrown in is creating new art.

No one is arguing that everything a human has created is good. No one is arguing that LLMs can't be useful.

Sutton is arguing that it can't be novel. Cherry picking a couple of words doesn't change his argument, which is very clear.