When the bubble bursts, this tech isn't going away.

Last year some blog recreated GPT-2, the first LLM 'too dangerous to release!,' for twenty bucks. Corpus to model within the hour. The billions spent on research since then have let newer models of similar size benchmark a hair below the big-iron cutting edge.

I recently found out Chroma, once a leading image model, was made by like two people, for presumably $200,000. Not nothing... but within reach of even small organizations. Recent pressure has been toward making models even smaller and less structured than that. Several updated video and editing models have quietly declined to release their weights, because they're small enough and powerful enough that most people would just run them locally. That desire, and that potential, will not vanish for wishful thinking.

If models had to train on nothing but public-domain, BSD, and CC0 materials, they'd still work the same way. If the money dries up to-morrow, then what's possible with hobbyist research alone will continue. If even they hit a brick wall, barely past current capabilities, that's still a couple gigs of linear algebra that'll try and do anything you ask.

Meanwhile - people will still make art. People make art even when it's illegal. Photography didn't stop anyone from drawing.