Gary Marcus

It's nowhere near the heart of the project though -- this is a feat of dexterous manipulation. That said, it doesn't help that the PR conflates the problem solving (trivial) and dexterity (major accomplishment) aspects of the work.

& the paper is framed in way that conflates pure neural networks with hybrid neurosymbolic systems & conflates pure learning w system that learns given a strong innate basis. 2 critical architectural decisions, at the heart of many recent debates, presented in misleading fashion.

Please don't conflate OpenAIs motivation for theater with the actual relevance of this research. It's about hand dexterity. The Rubik's part is for show.

That said. It has already been shown that System 1 intuition based systems can also sole a Rubiks cube.

where?

Google 'DeepCube' and find UCI's research on this.

Is MCTS considered a system 1 approach? Deepcube definitely uses it in its algorithm alongside DRL which achieves the autodidactic iteration part.

DeepCube also uses A* search, as classical as they come. I find System I/II a bit fuzzy, but certainly would not count MCTS as system I; it requires an enormous amount of very deliberate computation. It's more like System II on steroids.