François CholletVerified account

Makes me wonder, if you have the right foundation, cognition becomes an emergent property of that system

It does, just not exactly how we currently do supervised learning. Humans generate inferences by mapping space x to space y given any density sampling of x-cross-y.

The problem with calling learning through labelled data "supervised" is that it has no relation to how we are supervised by our environment during our continued learning process, trying to find stability in an environment of constant impulse feeds. So in short: agreed.

Maybe online learning, with continuous feedback on whether the mapping is a useful one is a closer match.

I disagree. I believe generative models are paving the way towards cognitive processes. Single shot and zero shot learning are supervised learning tasks.

When we are children and are told object names, does this not map space X to space Y? Does experientially learning its properties not then produce a higher dimensional embedding?

And too explain the criticism: we do not understand our own thinking and intelligence well enough to make any pro or con statement here. I'd give it a good chance that these mappings are maximally representative for minor parts of cognition but that's not more than a guess.