I've put my latest ideas here, in sort of a blog post @liuyao12@observablehq.com about a procedure for solving "tree search with geometric constraints", if that's a thing. It's geometric deep learning that (when working) compensates for a baby tree search (No fancy MCTS here).
I'll be adding more content, but as it now stands it might only appeal to those who are familiar with tiling *and* geometric deep learning. It could be a cold email for @tao@mathstodon.xyz though there's a lot on his plate already.
DeepMind folks might be interested, if I can do a proof of concept.