I think we're already finding the asymptote. "10 GPUs gets us 80% there, 100 GPUs gets us 81% there, 1000 GPUs gets us 81.1% there, who wants to throw 10,000 GPUs at this shit?"
I am not so sure. There are almost certainly some algorithmic workarounds that will dramatically improve outputs. You could probably even program a separate AI instance to generate slightly different prompts, which are then fed separately into the video generation prompt, until you get something you like. Then re-train the video AI on its previous videos humans have decided are "good". I think we still have a few years before we're in big trouble, but I seriously doubt the institutional/establishment ability to respond to this will be more effective than the reactionaries who have already proven their penchant for immediately spinning any public perception to their benefit, whether or not that perception is based in reality.
You are arguing that if you take the stochastic mean of "mediocre" enough times you will arrive at "excellence" and that's simply bad math. Mixing and remixing and remixing and retraining is all the AI companies have been doing for five years and they're still giving us story prompts like this and going "IS YOUR MIND NOT BLOWN" I spent 15 years in an industry where gadgets were developed to do the work 80% as good as a human. The end result was that gadget was invariably given to the 100% human. Even now, every AI dipshit techbro out there is coming around to "you need to study prompt training" as in "if you want to keep your job you need to figure out how to trick a markov bot into giving you useful information." I have no idea how long it took OpenAI to turn "a petri dish with a bamboo forest gorwing within it that has tiny red pandas running around" into that miniature horrorshow. What I do know is that the next step, in the real world, is a producer goes "great, now give them four legs" and the prompt honing continues. The next thing that happens is the producer goes "I liked the old red, bring it back" and now you're fukt because "red" was not a prompt before which means we don't get to cumulatively hack at this thing, we get to start over. If you're working with a Maya jockey? He's five, six hours into your thing and dollars to donuts, his pandas are quadrupeds. And when you give him notes? They do not obliterate the content you had before. When you're working with humans, you don't have a "pick which of these four interpretations of your idea are the least ghastly" you have a logical progression to completion. And I think it's extremely naiive of everyone to assume that the whole world will decide two-legged pandas are good enough.There are almost certainly some algorithmic workarounds that will dramatically improve outputs. You could probably even program a separate AI instance to generate slightly different prompts, which are then fed separately into the video generation prompt, until you get something you like. Then re-train the video AI on its previous videos humans have decided are "good".
