Thread by @Smerity

@Miles_Brundage @jackclarkSF @OpenAI @NvidiaAI (a) Released after I'd already spoken to the journalist
(b) Fragmented information that's important to the narrative
(c) A week of 256 TPUv3 cores which means it's trivially reproducible by larger companies especially with code ..? What's the responsible disclosure narrative?

@Miles_Brundage @jackclarkSF @OpenAI @NvidiaAI I had literally not even seen the week training estimate yet it was in reply to a post that uses my tweet as a source for the 256 TPUv3 core number (via @katyanna_q's article). I'm at the center of this web and I personally have every and no clue as to what's going on.

@Miles_Brundage @jackclarkSF @OpenAI @NvidiaAI @katyanna_q Journalists spoke to me about this as it's my area of work and as I have good awareness of @OpenAI's technical work yet I now feel I gave @OpenAI too much of the benefit of the doubt and that's sad for me, both as I feel like I misled others and also as I love @OpenAI.

@Miles_Brundage @jackclarkSF @OpenAI @NvidiaAI @katyanna_q I know that's harsh but that's getting to more and more how I feel. The responsible disclosure feels like a purposely crafted PR narrative rather than a practical one. Worse, it's on an important topic to discuss. State actors could sneeze and have this model trained.

@Miles_Brundage @jackclarkSF @OpenAI @NvidiaAI @katyanna_q If this truly was a threat (LMs at this scale are this good) then even the paper was dangerous. Yet the paper was released as well as the code and the model only takes a week to train and @NvidiaAI have done similar work in four hours. None of the narrative makes sense to me.