Stackexchange format is not always so great for cutting edge research. noticed this years ago, its a factor thats contributed to my "retirement" from Q/A involvement on the site. research inherently intimately involves unanswerable questions. re the idea of matrix permutations vs GI, theres obviously some deep connection, is it described/ surveyed anywhere? this also reminds me of my old idea of FSM equivalence testing connection to GI, seems maybe some deeper links there to be explored also.

re GI hardness generation: think its a great/ very worthwhile idea to study this problem empirically, have promoted that direction for many yrs, great to see some others finally taking it seriously, but theres a key way its different than P vs NP problems. "worst case problems" from that study tend to take exponentially increasing time to solve. but for GI, if the best algorithm is polylogarithmic, then the worst cases do not "stand out" as much vs execution time from the "easy" cases...

...

Growing and Retaining AI Talent for the United States Government / Raff

Cellular automata as convolutional neural networks / Gilpin

> Our results suggest how the entropy of a physical process can affect its representation when learned by neural networks.