Robert Ta discusses three AI governance proposals backed by Anthropic's Dario Amodei: banning chip sales to China, cracking down on model distillation, and requiring safety testing for all capable models. A speaker questions whether mandatory testing could favor large labs over open-weight initiatives and whether China could actually sustain an AI advantage.
CEO of Anthropic, Dario, says there are 3 things that he actually would support. Number one, we should not sell powerful chips or chip-making equipment to China, and we should crack down on the rampant smuggling and workarounds used to obtain access to such chips. Number 2, we should crack down on industrial-scale distillation operations. Distillation is a much more compute-efficient process than training models from scratch. It allows China to build much better models than its number of chips would ordinarily enable, and thus partially evade chip bans. Number 3, all sufficiently capable models, open and closed, should go through mandatory safety testing. The best way to address threat number 2 around the cyber and biological threats is just to test directly for those threats and alignment risks before release. He thinks that this idea is actually close to a consensus. He says, I have been heartened both that the Trump administration has moved in this direction in recent months, and by recent industry proposals that would apply such testing to the most capable models regardless of their country of origin or whether they are open or closed. What are your thoughts on his three measures he does support there that he publicly is saying like him and his company Anthropic are supporting? The don't sell chips to China, crack down on distillation. I have some thoughts on that, and then the all. capable models should be tested.
I mean, on the surface, they're all well thought out and seem like really reasonable suggestions for where we're at in the world. I think, you know, a couple of them contain embedded beliefs like that the CCP could be an asymmetric power that could gain advantage for an extended period of time. That's something that's not really happened in history. It's not historically informed. Is it just like fearmongering, or is there actually a reason to think that that position could be held? The industrial-scale distillation, I think, is a good distinction. Like, individuals and startups should have the ability to distill, but doing it industrial scale cannot. I don't know how they enforce that. That's on them. But that sort of makes sense. And I think that's something that, you know, the people who are advocating for open weight models would say that's reasonable. It doesn't like get in the way of the things that we think open weight models are going to be good for.
The third one is all sufficiently capable models, open and closed, should go through mandatory testing. So this is like the industry standards, like proactive voluntary testing approach.
Yeah, this one's, I think, pretty controversial because it's in the direction of regulatory capture. And because most people who make models may not even be able to, you know, sort of have enough resources to be compliant with these things, right? So that's been the thought is that OpenAI and Anthropic were trying to get these safety testings in order to like lock out other competitors. And I think this is on the institutions in our country, like to figure out like what is actually a threat and to define that. I think this is where Dario is accused of fearmongering and the fearmongering tactics. It's reasonable to think that these things might be dangerous, but there's actually not evidence that LLMs as technology can break through and create these kinds of problems that he's warning of yet. Yeah.
