Well there are certainly heavyweight researchers who believe the DL can organize and recognize higher principles. Andrew Ng and Geoff Hinton are among them.

But DL nets can be very fragile solutions, that only curve fit to the training data.

You dont want to recognize the same picture of the puppy, you want to recognize all puppies.

Others who disagree are trying to integrate logical approaches with DL into novel archetectures.

I see it as more of pendulum moving back and forth.

But Pearl's work is in Causal Statistics and is a hard core Bayesian.

https://www.quantamagazine.org/to-build-truly-intelligent-ma...

It's not like his complaints aren't being addressed in the field, they are just not easy problems (transfer learning, world models)

In a neural analogy, you don't haven't built a brain, you have a reflex.