This is by no means a comment about the quality of the project, but my god it's very uncanny in some frames. I feel like this would open up a lot of doors to creepypasta content. I'd love to play around with this
50+ hours on 256 H100s is considered impressively low training?
Really makes me wonder if any of this incredibly computationally expensive research is worth it, which seems only useful in potentially promising a future in which humans are given less opportunity to express themselves creatively - while delivering them an infinitely produceable amount of ai generated 'content' to passively consume
The main progress value is that Test-Time Training appears to work very well in practice. I think that as labs begin to test it as scale in LLMs, it will become commonplace in next-generation models.
Readit News
Really makes me wonder if any of this incredibly computationally expensive research is worth it, which seems only useful in potentially promising a future in which humans are given less opportunity to express themselves creatively - while delivering them an infinitely produceable amount of ai generated 'content' to passively consume
I'm wondering the same thing. 256 H100s were hot for two days straight to be able to make short clips of cartoons that almost don't look like shit?
It just isn't compelling to me.
Would be really cool to just use this (or parts of it) as one of the prompts and see what results.
[1] - https://www.newyorker.com/magazine/2004/04/19/cat-n-mouse
Deleted Comment