Geoffrey Irving noted capability halting is an existing safety proposal.
Interesting logic I haven't seen mentioned: - if Chinese models are better due to distillation, & - better models cause x-risk, but - you need to race to get to better models due to competition with china, then - best way to prevent that is to stop making your models better
@krishnanrohit I feel like I have seen people propose not making models better as a mitigation, tho.
Interesting logic I haven't seen mentioned: - if Chinese models are better due to distillation, & - better models cause x-risk, but - you need to race to get to better models due to competition with china, then - best way to prevent that is to stop making your models better
@krishnanrohit How you can win if you both have models of similar quality but china will win on scale and having the actual power to run the stuff with low margins?
@krishnanrohit so the safety case is just... don't train
@JohnGal43951639 If it's actually of similar quality then it's hard, down to brand or UX or security concerns etc
@krishnanrohit
Geoffrey Irving noted capability halting is an existing safety proposal.
Interesting logic I haven't seen mentioned: - if Chinese models are better due to distillation, & - better models cause x-risk, but - you need to race to get to better models due to competition with china, then - best way to prevent that is to stop making your models better