On Apr 28, 2025 09:47, Russ Allbery <[email protected]> wrote:
> We could get some protection by retraining the model from the base > training data For the moment, nobody in this thread considered retraining. I say that in some case, given some sponsors, it should be doable. I of course do not envision runing 300k+ GPU like Facebook does, though nodes with 8x H100 aren't completely out of reach. There are a few ways we could get our hands on them. Either asked for VMs sponsored, or run a crowed funding to get the cash owned by Debian. This may be the same project as building a cloud for Debian. Please don't say it requires too much skills or work, IT IS doable with enough sponsors. IMO we should give it a try if we really need it. Thomas Goirand (zigo)
