Gary Marcus

Of course, computational power isn't the real limiting factor for general #AI, anyway. We will need fundamental advances in understanding common sense knowledge and reasoning to achieve extrapolative AI rather than our current interpolative AI.https://twitter.com/Grady_Booch/status/1081740548627550209 …

But it could well be that we achieve those advances using even more vast amounts of data

Anything is possible, in the realm of the unknown. But there is no reason to believe it - data-reliant methods have not achieved (to my knowledge) even a tiny advance in common-sense reasoning. @GaryMarcus? @witbrock?

"Common sense" is knowledge about the everyday world. Do you think it is innate, or is it learned from experience? I'm betting on experience, at least for common sense about the physical world. 1/

some of it is obviously learned, but why would you think it is all learned, despite a billion years of evolution shaping neural development and a significant cognitive development literature on innateness?

These are good points. I guess I was thinking about whether it COULD be done with a pure learning approach. Of course the most practical approach might be related to the Belbute paper I mentioned: wrap ML around a physics engine.