Adobe will fork over cash for clips to train text-to-video AI

[alexahn]

https://www.theregister.com/2024/04/12/adobe_buying_videos_ai/

[gmerc]

Until Sora is out.

Same pattern: - RHLF -> RLAIF using GPT4 - Firefly using Midjourney

When there's a "good enough" AI solution, the synthetic data will come from it.

[leereeves → gmerc]

Nitpick: RLHF and RLAIF provide preferences between candidate data from another source, they don't provide synthetic data themselves. Of course someone can train a model using synthetic data generated by another model, that's just not RLAIF.

And a serious issue: that's probably not copyright safe in the way Adobe wants.

[gmerc → leereeves]

The fact that they trained on Midjourney kind of defeats that argument.