David Sacks thinks the ‘overheated’ AI rhetoric is ‘out of control’

AI is going to solve the problem that AI creates. That's kind of my view on where that market is going.

Good to be here. Thank you.

Well, I think the first thing to recognize is that we are in a globally competitive environment. We're not the only country that has advanced AI labs. And as the president declared in a major policy speech he gave on AI in July last year, we have to win this AI race and I'm not sure it has a definite finish line. Some people have said it's more of an infinite game. Either way though, we don't want to fall behind our global competitors because that will have a huge impact on our national security and on our economy. So that I think that's the first thing to recognize is that if somehow we slow down or stop AI development, it doesn't mean that AI progress is going to stop. It just means it's going to happen in other countries. And specifically China, something like half the world's AI developers are in China, are Chinese. They have the technological expertise, capabilities and a lot of the talent. So again, it is just globally competitive and it's not like we can just unilaterally disarm or stop AI development. So I think that's sort of the place to start. And then in that important AI speech, the president laid out a few of his pillars on how we win this AI race. And one of them is being pro-innovation. Doesn't mean we don't have to have any regulations, but generally speaking, we should be pro-innovation. And in the United States, the innovation comes from the private sector. Ultimately, it's our companies that push the frontier forward. And so, when I say let them cook, I mean, that's really what I'm talking about is trying to have a generally encouraging attitude towards innovation. So, that would be pillar number one. And then the president also laid out a pro-infrastructure pillar which really means being pro-energy which is something that he's talked about for a decade.

Right. Exactly. And so I think he anticipated the need for energy. Energy is the basis not just for data centers but for our entire economy. And fortunately he's been pushing for this pro-energy policy for a long time and it's never been more relevant or necessary than now. And then I think the third pillar that he talked about was being pro-export. We want the United States to win this race. The way that you measure winning I think in a globally competitive market is based on market share. If in 5 years we look around the world and all the data centers are running on Huawei chips and DeepSeek models, that means that we lost. We don't want to have that future. What we want to see is that the whole world is running on American chips and American models. That would lead to the best economic results for the United States. It would also lead to the United States having more soft power in this area. So that was another important pillar and I think it's just important that there are smart things we can do with respect to exports. Obviously we don't want to ship our leading edge semiconductors to China or something like that but generally speaking we want to help encourage our companies to have the greatest market share. So those were sort of the big pillars of the AI framework and then let me pause there and then I can tell you how we got to the regulatory framework that we just introduced a few weeks ago.

What's basically happened over the last year is that a lot of the action moved to the state level and so state legislators have been taking up the issue. In fact, there's something like 1,200 bills going through state legislatures right now. Our feeling on that is we don't need that many bills. I mean a lot of this regulation is sort of knee-jerk regulation. Although there are some areas where the states have regulated where we think it's a good idea, could be a good idea. And so what we did is we looked at all of those and we identified a number of principles that we then included in our framework. So one of them is around online child safety. That's a really salient issue. I think mostly it came from the whole social media debate, but it's also relevant for AI as well with AI chat bots. So we've said this is a legitimate area for states to regulate. Our north star on this issue is parental empowerment. I mean ultimately we think it's up to parents to decide what apps their kids should use and for how long, screen time limits and things like that and so we would let parents decide whether their kids use AI chat bots and what apps they install and how long they get to use them for. I think child safety is one area. There's another around communities. We do, with respect to data centers which is another issue that's become very salient for people. We don't want consumer ratepayers to have to pay for higher electricity prices because of AI data centers. We understand that that would not be popular at all and there's a lot of concerns about that. We've supported this idea of a ratepayer protection pledge which we got all the major AI companies and hyperscalers to sign up for in which they agree that for any new data centers that they build it will not increase residential electricity prices. The quid pro quo for that is making it easier for AI companies to build these data centers if they bring their own power. So that's sort of what the president has talked about is we want our AI companies to become power companies. We think that's a much better way of addressing voter concerns than just having a complete ban on data centers the way that Bernie Sanders is pushing. That would just stop the progress all together. So that's another area we've talked about in our framework. There's also some pillars around creators making sure they're protected. There's some pillars around innovation. Again, we tried to go through all the different areas that the states were regulating and put together our list of principles.

I guess it sort of depends. First of all, what's the causation and then what's the liability? I don't necessarily want to judge any of those cases too specifically because I haven't gotten into the fact patterns.

Well, I think people are very concerned about kids' safety online and that's a concern that we share. And again, this is where we think that the north star should be parental empowerment. Let the parents decide what their kids use. I should say that there's something like over a billion people who are using AI successfully every day. Now, it's one of the most quickly adopted product releases in history. I guess we're now about four years in to this roughly since the launch of ChatGPT. And so, I do think that there are these edge cases, there are these horror stories. I mean, I think that does always happen with a technology that's so rapidly adopted by so many people that there are going to be these anecdotes. I don't think that they've been super common. But yeah, but they're out there and I do think that the AI companies have already learned from some of these things. There are these cases of teenagers who had histories of mental health challenges who after using an AI chatbot may have engaged in self harm and I think that the AI companies have learned from that and now they're doing a lot more where if they detect that that might be the fact pattern that they will do things like obviously not be encouraging towards that behavior but maybe show numbers for suicide prevention hotline and things like that.

I mean, if they know, they probably should stop it. But the problem is just how do you know? I think it's very hard to know exactly all the ways that their product is being used. And I think that is one of the problems with some of these state liability laws that make the developer liable for the actions of the end user. I just don't think that the developer is in a position to know exactly how their product is being used. In the same way that Gmail, for example, doesn't necessarily know whether its email service is being used in the commission of a crime. I think that'd be a very hard thing for Google to figure out is just how is email being used or let's say that a criminal, let's say a money launderer uses Microsoft Excel. I mean, they use a spreadsheet and it could be a cloud-based one to help perpetrate their crime. It's just very hard to know exactly how the end user is using these products. I mean, this is why they have terms of use and they don't want their users using them for unlawful purposes.

Well, I don't know. You have to tell me exactly what cases we're talking about. But what I would say is that I mean should AI models give you step-by-step instructions on how to create a nuclear bomb or a bioweapon? No. I mean obviously not. We don't want them doing that. But that's something that's been identified as a potential misuse very early on. I mean going back to even before 2023 and that's why all these major AI labs have red teaming and they test for that kind of stuff to make sure. So, you know, I would say for the obvious misuse, I do think that they should do everything they can to prevent that. Obviously, are they going to be successful at preventing users from being able to use their tools in ways that we don't like? It's probably never going to be 100%.

Well, a little bit. I mean, I think Elon is maybe a little bit more pessimistic than I am. I mean, on these topics.

Well, I mean, I don't know. I mean, I think he's basically in a place where he acknowledges that AI is going to be developed regardless. And so, what we want to do is remain in control of it. I mean, I don't think we want to hamstring US development to such a point where again, then China just wins the AI race. So I think he understands that the technology is moving forward and we should develop it in a way that allows us to remain in control. So an area for example that I agree with him on is we want the AI to be as truthful as possible. And you know he's pointed out that if you teach the AI to lie, like let's say you teach it to be woke or whatever, that's very dangerous because then the AI can lie to us about what it's doing. That's not what we want to inculcate in AI. I very much agree with that. By the way, it's not like I don't think there's any potential dangers or risks of AI. I mean, I do acknowledge that there are potential dystopian futures we don't want. In my view, the most likely dystopian future would be something described by George Orwell in 1984 as opposed to James Cameron in the Terminator. My view of it is that the biggest danger is that the government will ultimately use AI to surveil us, control us, censor us, like that kind of thing. And to me, that should be the thing that we should be most afraid of is a marriage of government and corporate power around AI. By the way, that is the path that I think we were on before President Trump's election. And if you go back to the Biden executive order on AI, there were really two major things that it did. One was it said that DEI values, which is to say, all the woke stuff should be promoted in AI models. And that's how you ended up with that black George Washington controversy around that first release of the Google Gemini model.

I think that if you're an AI company and you don't want your product to be used in war, then don't sell to the Department of War. I mean, that's kind of in the name, you know, but they did make a decision that they wanted to be used by the Department of War. They do want to be used throughout the government. They made that decision and then they tried to insert themselves in the chain of command by having a veto over the lawful uses of their product. By the way, just to back up and just say that I'm not involved in government procurement, right? So, I'm a policy adviser. I was not involved in the Anthropic dispute with the Pentagon. I'm still not involved in it. But just if you want my opinion, I just don't think it was a very realistic position for them to take to say that they can dictate to a government agency what lawful uses of their product should be allowed. I just think you have to make a decision about whether you're going to sell to that government agency or not. I do think they did try to set themselves up in a way of being superior to the chain of command. I think that that's a tough position to take with respect to any government department, but particularly with respect to the Pentagon. I don't think it was particularly feasible. Now, if we get into their concerns for a second, one of their concerns was around mass surveillance. First of all, the Pentagon says that it does not engage in that and that's not a lawful use. What Anthropic I think said is, well, there's all these loopholes in the law. If they had come to me as a policy matter and said, hey, we have all these concerns about loopholes in the law. We want to ensure greater privacy for Americans. I would have been very interested in having that conversation. You know, I'm a civil libertarian. I'm very interested in privacy.

Well, yeah. I think if you have a concern about there being loopholes in the law, then let's try and change the law. You don't try and do it through a terms of use negotiation with the Pentagon. That's not going to work. I mean, in other words, they're trying to strongarm the government into agreeing to their changes to the law in exchange for their product. I mean, that's effectively what they're saying they want to do. But look, if they had talked to me about that issue, and by the way, they never did. I did talk to Dario a number of times and we did discuss various issues, but he never raised that concern with me. I think it only came up in the context of this Pentagon negotiation, which is kind of curious. But like I said, if they had been interested in discussing the surveillance issue, I would have been very interested in that conversation. I'm still interested in the conversation. Look, if they want to identify areas of the law that are problematic and could lead to the government using AI to infringe on the privacy of Americans, I would be interested in having that conversation.

Well, I don't want to comment on their lawsuit specifically, but I think it happens all the time that, you know, you can have companies work with the government, one department of the government, while another department is suing the government. And I mean I think these things happen all the time.

I think ultimately if everyone does what they're supposed to do, it won't be a national security threat. But here's basically what's going on is we have this area of cyber security where you've got offense and defense. You got the White Hats versus the Black Hats. And basically, it's all about preventing hackers from breaching systems, right? They try to break in, they try to steal data, maybe plant back doors, do other nefarious things. And you have this arms race between again cyber offense and cyber defense. Now, AI is going to be a major part of cyber security moving forward. Meaning that you're going to have hackers who are powered up by AI models, but then you've got cyber defenders who are powered up by AI capabilities as well. I think that the cyber security market will eventually reach a new equilibrium in which both offense and defense have AI capabilities and therefore the danger won't be ratcheted up to a point where we can't control it.

Well, no, no, no. Listen, I think holding it back was the right decision because where I'm going with this is I would not want the hackers to have access to it before the defenders, right? What I'm saying is that I believe that the market will eventually reach a new equilibrium in which there's AI capabilities on the part of both hackers and then also cyber defense. And you know, again, AI is going to solve the problem that AI creates. I mean, that's kind of my view on where that market is going.

Well, it may not solve every problem, but I do think in the case of hacking, it will. And look, I think that Mythos is a good example. By the way, there I think there were multiple reasons why Anthropic held back that model. I mean, supposedly it's like a 10 trillion parameter model, which just means that it's very big. It's very expensive. I'm not sure they had the compute to serve that commercially. So, I don't know that a commercial release was really even feasible for them. But in any event, they're giving limited access or preview to a couple of dozen companies and the government. And what the people who are in this project Glasswing will do is they're going to use the capabilities to scan their code bases, look for vulnerabilities, and then patch them. So they're going to harden their code bases. Ultimately, Mythos isn't the only AI model that will do this. I mean, all the major AI companies are working on this now. It's really an extension of coding capabilities. So what happened first is that the AI models got really good at coding and then as a result of that they got good at finding bugs and then some of those bugs are vulnerabilities and some of those vulnerabilities can be strung together into an exploit. So what you want to do is use these new AI coding models to find the bugs and the vulnerabilities and patch them before a hacker does the same thing. And I think the net result of it is that we're going to find a lot of dormant vulnerabilities that were there in the software maybe for a long time that no one had found before. As long as everyone does what they're supposed to do, which is the CISOs and the IT departments use these new tools to patch the bugs, then actually it will harden cyber security around our core systems because those vulnerabilities won't be there anymore.