Generative AI is obviously more fun and productizable than classification, recognition, and recommendation AI, but is it also more technically and philosophically significant?

I personally think yes. In fact I think generative AI subsumes those other categories. All intelligence is generative

Blogger at ribbonfarm.com, running summerofprotocols.com, host of /gloom ☀️

You can ask a generative AI to classify, recommend, or recognize. The reverse is not true. GenAI is a qualitative level up, not yet another application category.

True, one is a forward problem and the other is an inverse problem. Predicting p(X|y) is harder than the reverse if X is high dimensional.  To make this possible we came up with a way to learn a very good prior on X using big NNs. The prior constrains the soln of inverse problem.

Apologies for handwaving here, we’re not really solving for these distributions, just generating samples but maybe there’s still something there

So the real innovation is learning the prior I think. I’m not going anywhere in particular with this but I think you’re right. I guess next level is conditioning p(X) in ever more sophisticated ways to control generation