Access limits can reduce misuse but cannot eliminate it

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 9:22

   huge financial incentive incentive to be closed source so how do you ensure that there are few or no misuse incidents if there's uh no guardrail on the model yeah on the model layer at least yeah i mean i i think it's tough to get to like zero a chance of like any bad actors using it right i think you can just do your best and i think that we have some mechanisms and we're actually onboarding some additional mechanisms to kind of detect like bad actors or misuse in the future but yeah I think it's always going to be even for the frontier models I think that people think of unrestricted models as like kind of this gateway to misuse but like they're still jailbreaking still exists on frontier models but all you can do is like okay close patch it you know like best case scenario and etc etc so I think that it's always going to be best effort

2. 10:11

   and I think that's just the industry in general right is how do you stop crime or how do you stop hackers how do you stop anything it's always it's always going to be best effort and i don't think you ever get to a point where it's 100 but you always try and i think we're going to try just like everyone else is trying to detect and stop misuse i mean jailbreaking is kind of hard though like frontier models have gotten really good at not being jailbroken and um you know i have no insider knowledge of like the state of what is being used for offensive cyber but i think it's at this point a lot of it is like glm 5.3 type models rather than jailbroken mythos or astra yeah i mean that's probably true if you think about jailbroken mythos or astra because they do a really i mean they're they're really stringent on like who gets access right so

3. 11:00

   i mean if you're you're going to narrow it down to where like only like 100 companies get access then i mean yeah that's definitely going to be true right like if you that's that's kind of like that's what i was kind of saying it's like anything right like if you want to minimize something that in this room gets stolen right then if there's only two of us here right you're gonna have a way less chance right but if you if you let eight million people into this room your chances are going to kind of go up way up so I guess that is one way to to kind of minimize the risk I guess if that's the approach you think is best. So like going back to misuse instances if you were to have a misuse incident like if someone used obliteration models to do something bad like how would you be able to detect and mitigate that and respond to it yeah so i'm not going to go into

4. 11:43

   like the full internals of our system today but i think that um yeah like we haven't we haven't had that issue quite yet like it hasn't came to us no one we've never had an incident where someone's like used our models and someone's like oh this criminal has used your model too do something bad so that actually hasn't happened quite yet and then the second of all like we're like i said we're continuously improving our efforts and i'm not going to kind of give away all the internals of how our system works today yeah for sure so do you think that