I think its a no brainer and possible doom scenario. Is it possible? There are common operations like... the L1L2 norm that are used across machine learning. Is there a way to craft an operation like a regularizer that is both useful and profitable? Or could up you use a block as an initialization parameter and somehow make use of it? A prediction task itself might work - solve this provably hard problem and you own the cash. I don’t know but it seems like what is needed.