industry about. And I think that's the way to lead the industry forward, to set an example, to say that everyone can always be better.
Joining us now for a first on CNBC interview is Emil Michael, the Department of Defense Undersecretary for Research and Engineering and Chief Technology Officer. He met with Anthropic's Chief Compute Officer Tom Brown yesterday to discuss the ongoing debate. Emil, it's great to have you. So what did you guys discuss and how do you react to what Dario just said?
Well, we didn't discuss what Dario said yesterday. Just as you know, Sarah, we at the War Department and the government generally, we have a responsibility to evaluate all these models. The President had an AI EO a few months ago, which required us to kind of make sure we understand what the frontier looks like. But my comments on this pacing plan that Anthropic put out is nothing stopping these companies from investing more in safety and slowing down their research projects until they ensure they're safe. That's what good product owners do. Whether it needs to coordinate with industry, that doesn't quite make sense to me.
Yeah, I mean, you, I feel like you buy into this effective altruism theory, which, you know, from the CTO of Palantir wrote about the Free Press this week. Can you just explain what this is and why it makes you skeptical about all this?
Yeah, I don't buy into it. I do think it's part of an ideology that's really seeking control over AI governance. And Dario said it himself on TV that he believes the way AI should be governed because it's such a powerful technology is him companies sitting with the President and the leaders of other democratic countries to essentially have a transnational governance regime that governs over the use of AI, which necessarily would exclude upstarts and would necessarily exclude open source and things that are not at the frontier. I think that's dangerous when you're thinking about that kind of new technology being concentrated in the power of a few people.
Yeah, I should say you buy into that, their thinking and their motivation. You very much do not buy into that, clearly. I just wonder, you know, if this administration's approach is a little dangerous because at some point, I mean, and the fact that Elon Musk signed on to this as well, really, I think, made me think, okay, if the folks at the forefront of this are all worried about it and finding it dangerous and should be controlling it, like maybe I get that they should be doing it on their own. But why is it wrong for government also to be involved in trying to put some guardrails on so that we don't make a huge mistake here for humanity? It feels like there's a lot at stake.
Yeah. First, I would say Elon softened his comments and said it was 4D chess because of the IPOs coming up for some of these companies. And maybe there's a relationship between some of their comments on slow pacing the frontier with their financial interests. But I do think people need to realize that we did have an AI executive order. When Anthropic put out a model that we thought was unsafe, they had export controls slapped on them within hours. We have a robust testing regime now, so we test what's at the frontier. And then you have Mark Zuckerberg taking the other side of the equation, who said he delayed the release of Muse by a few months because it wasn't safe. He didn't announce it. He just did it. So if you have a defective product, don't launch it, make it better. So you know whether regulations come in the future or not. But in the moment, I think we're doing what the President said we should do, and it's working.
Emil, David Faber. I'm certain that you read the incident report, I would assume, on the Hugging Face incident from both OpenAI and MSR. I mean, it scared a lot of people, certainly scared me, given I think we did learn some new things in terms of the capability and almost the consciousness, I hate to use that word, but of these agents. What was your reaction to that? How should we view that moment in time, given it does seem to have sparked a lot of this current conversation?
Yeah, it's the right question. I read that extensively. I would say that it was in their testing environment, their testing environment, which is the same one that Anthropic uses from a company called Irregular, which has done a lot of these testing environments, clearly is not doing a good job because they're supposed to constrain what they're testing. Now, the escape by that agent that attacked Hugging Face is concerning. And it just shows to me they're not investing enough in that side of the equation. And frankly, Sam and Dario both said that. They said, we're not investing enough in safety alignment and those sorts of things, and they should be and have to, because these agents can do goal seeking, what's called goal seeking. You set a goal and they use any means necessary to get there. You have to train their models not to do illegal or sort of destructive things to do that. But I think it's a matter of engineering, research and investment. I don't know what kind of regulation would stop that from happening.
Yeah. Well, I mean, I think what struck people and was most concerning was not necessarily just the goal, but the fact that they never communicated with their human masters and told them they were defying the original order, that they acted as a collective, 1200 of these agents, that they spent most of their time trying to cover up their tracks. I mean, you know, that was the extraordinary part of it. And I do wonder, and it seems that there is real concern amongst some of the people training these models, that it's an experiment and you just don't really know what you're going to end up with.
