[EigenLord]

Interesting. I've been slowly coming to the conclusion that the way forward with machine learning is actually less "machine learning" as we've grown accustomed with it, less pretraining, less data, less search, more direct representation, symbolic processing, more constraint-satisfaction, meta-learning etc. All those things we need less of (pretraining, data, etc) are messy, brute force, and contingent. Working with them, you'll always be dependent on the quality of your data, which is fine if you want to data-mine, but not fine if you want to model the underlying causes of the data.

From my (admittedly sketchy, rushed) understanding of what they're doing, they're essentially trying to uncover the minimal representation of the solution/problem space. Through their tracking of the actual structure of the problem through equivariences, they're actually deriving something like the actual underlying representation of the puzzle and how to solve them, rather than hoping to pick up on this from many solved examples.