Clever, I wonder what would happen if you did this in reverse? Like, if we froze the weights then added some layers trained on a bunch of weird edge cases then tried to convert those weights back into branches.

QA through entropy lol