Oliver Habryka of Lightcone Infrastructure argues internal AI models should be publicly deployed to expose safety flaws quickly
Academic Boaz Barak agrees internal use is high-risk.
I guess you don't literally mean all internal models or ASAP - you mean all models-that-one-might-plausibly-want-to-deploy (e.g., not helpful-only ones), as-soon-as-the-safeguard-learning-rate-decays-a-fair-bit, or something like that? I think it's bad to ship something with known-very-broken safeguards
Oliver Habryka@ohabryka@Miles_Brundage I think all internal models should be publicly deployed as soon as possible, so I think I am in favor of this. If the safeguards aren’t ready then it’s good for people to notice that! Most of the risk comes from internal deployment not external.
@ohabryka (agree we don't want too big of an asymmetry re: public/private knowledge, or public/private use, but there are levers to push on there besides "give it to everyone" such as expanded researcher access, transparency, not deploying internally either, etc.)
Miles Brundage@Miles_BrundageI guess you don't literally mean all internal models or ASAP - you mean all models-that-one-might-plausibly-want-to-deploy (e.g., not helpful-only ones), as-soon-as-the-safeguard-learning-rate-decays-a-fair-bit, or something like that? I think it's bad to ship something with known-very-broken safeguards
@ohabryka Gotcha, think we just have very different views on policy then (prob stemming from the "where all the risk is" part)
Oliver Habryka@ohabrykaNo I would support regulation that you can only deploy a model internally after you deployed it externally, so I do mean all internal models. I also think shipping helpful-only models would probably be net good? But it sure is tricky and less obvious. I agree that it’s bad to ship models with known broken safeguards, that’s why you shouldn’t deploy them internally where all the risk is!
#133Boaz Barak@BOAZBARAKTCSGenerally agree that public deployment is good. Internal deployment is one of the riskiest and highest stakes applications.
Oliver Habryka@ohabryka@Miles_Brundage I think all internal models should be publicly deployed as soon as possible, so I think I am in favor of this. If the safeguards aren’t ready then it’s good for people to notice that! Most of the risk comes from internal deployment not external.
#1902Nathan 🔎@NATHANPMYOUNGTake that I couldn’t have guessed in advance.
Oliver Habryka@ohabryka@Miles_Brundage I think all internal models should be publicly deployed as soon as possible, so I think I am in favor of this. If the safeguards aren’t ready then it’s good for people to notice that! Most of the risk comes from internal deployment not external.
No I would support regulation that you can only deploy a model internally after you deployed it externally, so I do mean all internal models.
I also think shipping helpful-only models would probably be net good? But it sure is tricky and less obvious.
I agree that it’s bad to ship models with known broken safeguards, that’s why you shouldn’t deploy them internally where all the risk is!
Miles Brundage@Miles_BrundageI guess you don't literally mean all internal models or ASAP - you mean all models-that-one-might-plausibly-want-to-deploy (e.g., not helpful-only ones), as-soon-as-the-safeguard-learning-rate-decays-a-fair-bit, or something like that? I think it's bad to ship something with known-very-broken safeguards