François CholletVerified account

A way to describe intelligence is that it is the power to produce abstraction. AI in the true sense would be Autonomous Abstraction. Current AI consists of recording abstractions generated by the human mind (via hard-coding rules, training ML models on human-labeled data, etc).

All generalization must by definition come from some form of abstraction. The abstractions that can be learned by DL models (often reflections of human abstractions injected via labels) are fairly weak and shallow, which is why DL is only capable of local generalization.

If AI today isn't capable of autonomous abstraction, it isn't because it's a particularly hard technical problem. Rather, it's a subtle conceptual problem. It doesn't seem like many people are looking at it. You can't produce the right answers if you're asking the wrong questions

AI today is to human intelligence what the phonograph was to human musicians. Superficially, it may seem like it has a comparable output & is ready to replace musicians (which was a real fear back then). But it has radically different capabilities: it's merely a recording device.