Gary Marcus

um, wasn’t that close to my assessment, too? i wrote “the system is (very impressively) trained to do the perception and manual dexterity parts, but cube solving algorithm itself is innate, and symbolic, not acquired via training.”

ask 500 people on the street what they think “solving” a rubik cubes mean.

see other reply that starts “because”; they misled on the two key intellectual questions that are confronting current AI & also because made it sound like they had done something sexy that they hadn’t done. you think their framing is as accurate as alternative titles i suggested?

Just to chime in, though it's possible to solve the cube with RL, OpenAIs thing would be way more impressive if truly end to end (it learned to solve AND manipulate at same time), so it is fair to point out that 'our system learned to manipulate a rubik's cube' is more accurate.

Even if they had rubiks cube solution being done via a neural net, if solving NN was seperated from manipulation NN that's injecting human prior via architecture. Which is exactly what they do by separating vision from manipulation already. Lots of caveats in paper actually...