It's legit. For the network types this works on, it's much less computationally complex. A lot of the networks we use now are based on the assumptions of fast m

It's legit. For the network types this works on, it's much less computationally complex. A lot of the networks we use now are based on the assumptions of fast matrix multiplication, so it's not universally applicable -- but there's a ton of networks this does work on, and there's a good indication that we can rephrase a model to solve the same problems in the hashtable space.