Gary Marcus

There are theoretical impediments that even "general methods" cannot circumvent. If you search for a common point of two parallel lines, you can have the best search-and-learning method in the world and you won't find one. ML confronts such impediments to reaching GAI. #Bookofway

That's the whole debate, isn't it? Is the human ability to learn things like mathematics, geometry, formal logic, informal reasoning, and science the result of some ensemble of general search and/or learning processes operating in the brain, or something else?

nobody doubts search and learning are important, but those are so broad as to encompass almost everything that has been or will be tried. shocking as it may sound, http://rebooting.ai  strongly endorses learning, though of a form that is more compatible w abstract knowledge.

Wouldn't the disagreement between your Rebooting AI thesis and Richard Sutton's Bitter Lesson thesis be that Sutton believes the answer is to keep working on general learning methods like reinforcement learning and that you believe the answer is to pivot to symbol manipulation?

I don't think Sutton's argument is that if you add 1 million times more data and compute to existing machine learning techniques you'll get AGI. I think it's that researchers should work toward ML techniques that can efficiently leverage 1 million times more data and compute.

Why speculate on what Sutton means or meant? Do you @strangecosmos believe that any ML technique can solve any of the toy problems in #Bookofwhy or Primer, given all the data in the world, and no information beside data?

I can’t say what ML techniques will exist in 50 years, so I don’t know. I think Sutton’s essay (http://www.incompleteideas.net/IncIdeas/BitterLesson.html …) is pretty clear. I’m not aware of any ML expert who thinks no further fundamental research progress is needed to get to AGI, just more data and compute.

If by "ML techniques" you mean any future algorithm, then you are justified in saying "I dont know". But if by "ML techniques" you mean algorithms based on data only, we can tell even today that the answer is NO. We can't compute 3-dim volume from a 2-dim shadow. #Bookofwhy