Where a lot of this gets stuck is that AI software dev isn't a closed loop. They're still completely dependent on complex human written codebases to actually do anything. They're adept users of libraries and platforms but, at least for now, not so adept builders of the same. The ability cutoff being roughly where public documentation ends.

I don't think this will be a hill forever, after all folks have managed to get new novel mathematics out of these models, but for the medium term it seems like the good (software) business to be in is selling shovels to the people digging for AI gold.