AI of Deepmind discovered 4x4 x 4x4 non-comm. multip. algorithm with just 47 mul.
https://www.nature.com/articles/s41586-022-05172-4
This of course will destroy all sceptics who did not believe that neural networks can do math in symbolic or even pure form. Considering oneMKL is only so fast because of libxsmm library (and by itself outperfoms everything else) I will report it here first. THIS ALGORITHM is non-commutative, i.e. has A elements on one side and B matrix elements on the other of EACH "*". THAT means it can be used recursively to multiply block matrices.

Sounds good! Looks more like a discussion than an issue to me.
Also, this is mod2. So may not be useful immediately. But this is: https://arxiv.org/abs/1904.07683
You can even adapt code from here, slightly changing https://stackoverflow.com/a/10836199/11173412
All results. Both modular and normal arithmetic. 7 MB file. equations.txt