The restriction aims to prevent rapid cyber capability proliferation.
Depends how much open-source AI could be accelerated when using their models to do the work
If a lot of the gains come from good post-training, then preventing acceleration makes sense from a safety point of view because you don't want open-source catching up to Mythos-style cyber capabilities super quickly
@BlackHC Do you think they’re actually worried about them? I can’t see Ant being worried about eg Ai2 or Arcee
@BlackHC Hmm, ok, that’s interesting.
Depends how much open-source AI could be accelerated when using their models to do the work
If a lot of the gains come from good post-training, then preventing acceleration makes sense from a safety point of view because you don't want open-source catching up to Mythos-style cyber capabilities super quickly
The restriction aims to prevent rapid cyber capability proliferation.
Depends how much open-source AI could be accelerated when using their models to do the work
If a lot of the gains come from good post-training, then preventing acceleration makes sense from a safety point of view because you don't want open-source catching up to Mythos-style cyber capabilities super quickly
@BlackHC Do you think they’re actually worried about them? I can’t see Ant being worried about eg Ai2 or Arcee