00:00Obviously, you are one of the four board members that voted to remove Sam Altman a few years back.
00:05We don't need to rehash that history.
00:07But I do just want to know, kind of three years later, removed from that,
00:10who or what do you actually think should be sort of looking after this frontier AI?
00:17I think it gets back to what you just said, which is in most industries that are consequential,
00:21where there are real risks, we don't just have one person or one regulator or even one government
00:26that's in charge.
00:26We have these complex, multilayered sets of standards, industry rules, government rules.
00:32The public has a say.
00:33And that's the kind of thing that we're going to need for AI if it's going to benefit everyone.
00:37Well, the CEOs of both Anthropic and OpenAI gave a presentation to the UN today,
00:44trying to address some of these issues with regards to not only how it should be overseen,
00:49but also what restraint, if any, the companies themselves would show,
00:52particularly on things like bioweapons and things that are a little bit more controversial.
00:56Do you think that the companies themselves should be a little bit more proactive
01:00in showing restraint in the absence of any true regulatory oversight?
01:04I think they do need to be thoughtful about what they're doing.
01:08I think it is appropriate that they're calling for government regulation.
01:11Of course, you know, we have seen the movie before where companies call for their industry
01:15to be regulated.
01:16And then when it comes to any specific proposal, they fight it tooth and nail.
01:20We've actually seen that with AI the past couple of years, I would say.
01:22So it is a question of whether they change their tune there or not.
01:26But in the meantime, certainly they need to be looking at whatever they can do for themselves as well.
01:31But when you hear, let's just say Dario Amadei, just not to cherry pick too much,
01:35but, you know, a couple of weeks ago he was just talking about the big threats that he saw from
01:39this.
01:39When we hear those types of warnings coming from the people building these models,
01:43is that self-serving or should we really be listening to their warning?
01:46I think we should really be listening.
01:48And it's because, you know, we should be skeptical of what they're saying.
01:51We should always ask, is it in their own interests?
01:53But it's really not normal when people are building a new technology.
01:57Either the researchers, the engineers, or the companies that are building that technology
02:00aren't normally the ones who are sounding the alarm the loudest.
02:03So the fact that that is what's happening in AI, I think, should give us pause
02:06and should make us stop and listen and think,
02:08even if we don't take everything they say at face value.
02:11There's a bit of a conflict going on right now,
02:12not just because there really is no universal oversight,
02:16but also the speed at which this industry is going.
02:18I mean, some people are saying, look, it's unprecedented.
02:20You can go back to any sort of major revolution we've had in human history,
02:24certainly in the United States,
02:25and the speed of which AI has come to force is like nothing other.
02:29Is it possible for governments and regulatory bodies to even catch up at this point?
02:35I think if we're trying to do regulation in the format of
02:38the government writes a checklist and here is what you have to check off
02:41in order to be safe, that's not going to work here.
02:43It's moving too fast.
02:44But I think this level, the speed of development
02:47is why you're seeing things suggested like,
02:49can we bring in third-party auditors who are coming in
02:52and looking at what the companies are doing?
02:53Or can we put disclosure rules in place, transparency rules,
02:56so that even as things go faster, we can have not just the public,
02:59but also external experts, companies looking at each other's work.
03:02I think there are approaches we can take that can keep up with the speed
03:06if we are willing to put them in place.
03:08As far as the approaches, I was speaking with someone yesterday,
03:10an expert on this topic, and she was talking about this idea
03:13of something akin to a nuclear non-proliferation agreement,
03:16but just for the AI and the modern age.
03:19The leaders of the two superpowers are set to meet in Washington,
03:23President Trump of the U.S.
03:24and obviously President Xi Jinping of China.
03:26Is it possible that those two countries, excuse me,
03:31can actually lead the charge on this, assuming they can find common ground?
03:34I think we should have pretty modest expectations for what we'll see tomorrow.
03:38AI is pretty near the top of their agenda,
03:40so I'm sure we'll see some kind of statement about managing risks,
03:42but whether there's a lot of substance behind it remains to be seen.
03:45Over the longer term, I think AI is a technology with all kinds of different applications,
03:50all kinds of different risks and challenges it raises.
03:52Some of those make sense to handle domestically.
03:54Some of those are maybe an international question.
03:56I think the thing that is really central for the U.S.-China relationship
03:59is these risks that are arising from the speed of development,
04:03because it is U.S. and Chinese companies that are pushing the fastest at the frontier,
04:07and if they're saying that they're not able to do that responsibly
04:10because of fear that they'll be overtaken by someone else,
04:13that's where you need that bilateral conversation.
04:15But on the negative side of this,
04:16I mean, one issue with that comparison with nuclear non-proliferation treaties
04:20is that everyone understood the power and destructiveness of nuclear weapons.
04:24I think we understand maybe potential some of the destructiveness of AI,
04:28but there is also a lot of upside with regards to what AI brings to our economies
04:32and just overall to humanity, if you believe what we know already.
04:37That's a hard balance to say we should stop this simply because there is a big downside
04:43if we know that there is potentially a big upside.
04:45Are you a believer in that upside, and do you think that upside outweighs the downside?
04:49I think there's definitely a lot of upside from AI if we do it right.
04:52I think if we rush in a way that is dangerous
04:54and that risks losing control of the technology completely,
04:58then we're not going to get any of that upside.
05:00So I don't think trying to shut it down, trying to prevent proliferation
05:04the way that we thought about it for nuclear weapons is the right approach.
05:07The way I think about the nuclear analogy here is that was a space where
05:11even in the midst of the Cold War, the U.S. and the Soviet Union was able to talk to
05:14each other,
05:15were able to talk to each other about shared concerns and ways to mitigate shared risks.
05:19And I think in the same way, even as the U.S.-China relationship continues to be relatively tense,
05:25there can be shared interests that both countries need to come together to tackle on AI.
05:29There's been a lot of concerns about how these AI agents and these large language models operate
05:36and the autonomy that we've seen with some of them, autonomy that has even surprised companies
05:41like Anthropic and AI, the hugging face incident is sort of illustrative of that.
05:45The idea that they set these bots out on a task and they were very aggressive in going after that
05:51task
05:52to the point of effectively going off the reservation, speaking to each other in ways that their creators
05:57did not intend to and certainly were not even aware of.
06:01Do we actually understand what these machines are capable of?
06:05I think we don't.
06:06And I think that's why you're starting to see not just the CEOs, but also a lot of top leaders,
06:10top researchers inside industry and outside industry saying,
06:12hang on, we might need to take a little bit more time here.
06:15We might need to, the phrase now is pace the frontier.
06:17We might need to install a brake pedal so that we have the ability to make sure that we really
06:22understand
06:23and can control the technology that we're building.
06:25Well, here at the Georgetown Center for Security, I mean, how much of this can actually be,
06:30if not litigated, certainly mapped out in academia in a way that maybe it can transfer
06:36to the regulatory and legal space?
06:38I mean, I think traditional academia with sort of multi-year journal publishing timelines
06:43is not going to be able to keep up here.
06:45But I think there's a lot of, you know, academia is one source of external expertise,
06:49independent expertise, policy centers like the one I lead similarly.
06:53I think it's not necessarily traditional academic structures that we need,
06:56but certainly independent voices, independent eyes, and a more greater polarity of perspectives.
07:01Thank you, Dave.
