Gary Marcus

But it doesn't capture others (e.g., between "dog" and the smell of a wet dog, the feeling of touching a dog's fur, the problem of cleaning up dog hair, etc.). So "dog" is partly grounded for Google translate and it could be more grounded for a robot with touch and smell. 2/

But "dog" will never been completely grounded even for humans, because there are always aspects of dog-ness that an individual human has never experienced. The richer the set of relationships a system knows, the broader and deeper its intelligence can be. end/

That's a good explanation of the degree of "groundedness" of a concept. The reveals (1) experience with interacting with a concept increases groundedness (2) knowledge collected socially increases groudedness. This effectively tells you that concept's groundedness is not fixed.

The fact that groudedness is a spectrum and likely dependent on context, tears down the entire 'symbol-manipulating' narrative.

It just depends on which level of abstraction you can do an inference. May be you are asking why to do inferences? Well you cannot learn everything from experience like falling from a cliff. Sometimes inferences may save life. The question is how to do inferences in a neural net.

Isn't essentially everything learned from experience? That is, we learn by imitation and then we learn through language, in both cases its still an experience. There's no direct upload mechanism like we find in the Matrix.