Dean Ball 谈加入 OpenAI：新权力中心、前沿 AI 政策与“主角能量”

Dean Ball, author of the Hyperdimensional Substack, welcome back to The Cognitive Revolution.

Thank you so much for having me back, Nathan. It’s great to be here.

I’m really excited for this conversation. We’ve got some big news in your life to cover, and you’ve really been in the eye of the storm over the last year and a half. I have so many questions about everything you’ve participated in and your thoughts on where we are today. I’m going to try to go Tyler Cowen-style on you and fire a bunch of questions. I mostly just want to hear from you on so many different topics. Are you ready for a podcast sprint?

I’m ready.

All right, let’s start off with your time in the White House and take a little look back on America’s AI Action Plan. It was very well received at the time. How would you critique it now? Is there anything that you feel like you would change or do differently, looking back at a conceptual or policy level?

What you have to remember about the Action Plan is that I don’t feel like there’s a ton about the world, in terms of how AI has developed, that has surprised me. Basically, we’re still in the world I expected: models with scary cyber capabilities in late 2025 or early 2026, and probably bio soon after. I had publicly predicted a lot of this stuff. Coding agents maybe surprised me in terms of their popular uptake, on the positive side. I didn’t expect that.

The thing is, D.C. wasn’t living in that world when the Action Plan was written. The Action Plan is this weird example of strange hermeneutics, where you’re writing a document now and your audience is the present-day audience, but your audience is also—you’re trying to model the same people in a slightly different near future, where they’re 30% more AGI-pilled than they are today, and then 50% more. You hope that they go back and look at the document and say, “Oh, wait. I now read this in a totally different way, now that I’m thinking about it this way.” Right?

You could definitely criticize that as being an act slightly too much of five-dimensional chess or whatever. It wasn’t intended to be. I think it was just the nature of the task.

I think the one thing I might critique is simply that it probably would have been good to be a little bit more explicit that we are really talking about generalist agents that can do all kinds of things, and to try to explain not just what that will mean for America, but also—

A big part of what the Action Plan is all about is, “Okay, this is happening, and the government’s not leading it. The government needs to figure out how to ride with the current of the river and use this to maximize, in my view, American primacy and American geopolitical power,” while also understanding that the best way to do that is to be positive-sum, try to grow the world economy, and bring other people in the world in on this.

I feel like the Action Plan doesn’t really stitch that together all that well. It probably reads a little bit more like 3 dozen separate thematic objectives than it does 1 cohesive thing unified by a common strategy or a common vision. If you had given me 2 more months, I probably would have focused on that.

The other thing I will say is that there were things we left on the cutting-room floor that I feel bummed about. One of them is that I was really passionate about adoption in particular sectors and trying to talk about what it means to do case studies in really specific industries—saying, “All right, what are the barriers here that the federal government can do something about?”

Hospitals and hospital recordkeeping, for example: there are very specific, in-the-weeds things that the Department of Health and Human Services could do. Veterans Affairs is another one—an amazing, huge system, one of the largest single-payer health care systems in the world.

We don't think about America as having single-payer health care. We do the VA. There are huge amounts of data and huge amounts of direct medical care being provided by government employees. This is the kind of thing where experimentation with AI in health care could have been enormously valuable. We just didn't have time, so I would criticize that as an area where we could have been more specific.

Then you handed off the baton, saying at the time that you're more of an ideas guy and that it would be left to somebody else, who you hoped would be better at implementation and running all these things through the actual process of government. How would you say that is going right now? It seems like we clearly have the buildout happening. Even in my home state, not too far away, in Michigan, a gigawatt data center just broke ground despite some local NIMBY-style objections, so that seems like it's happening.

We hear about the military trying to use AI. Obviously, that's contested in terms of what they should be doing, and as a public, I don't think we fully know what they're doing. Then there's all these other things. Your writing in the meantime has sounded the alarm in a pretty severe way around just the health of the republic, and it seems like your faith in government's ability to ride the wave as one might hope is not super high right now. How would you say it's going in terms of follow-through, and where is the core of that pessimism for you?

I think if you looked at the individual items in the action plan—and some of this is hard because some of the things on the action plan, some of the implementation, ended up being done, as they say in government, on the high side, which means in classified environments. Especially some of the stuff about military adoption and some national security things, for example.

One thing that I feel like is underrated is—not to say that this specific thing has been implemented, but just as an example of the kind of thing I'm talking about—there's a part of the action plan that obliquely references the notion of the military commandeering all the data centers in the country in the event of a national crisis, where we needed to stitch them together to do … something. That's in there, and it sounds like a very Leopold Aschenbrenner idea. It was described in sufficiently mundane language that I feel like it didn't jump off the page at people.

But there's a bunch of things like that where the implementation, to the extent it's happening, is not happening in public settings. But I think if you were to look overall—and if you had a clearance and could really see everything—I think you would see that we're probably, if you just think of it as a to-do list, 30% to 40% done, which is pretty good for a year, right? We're about 11 months out from when the action plan came out, so that's pretty good.

A lot of the major things, I think, across all of the pillars, we've seen significant advances that the administration has made on the energy side of things. Not that this was directly in the action plan, but this administration is doing really amazing stuff on nuclear. Then, stuff that was more directly downstream of the action plan: There are major changes that are in process right now that should be announced in the coming days, really, from FERC—the Federal Energy Regulatory Commission—that deal with the process for connecting very large industrial electricity users to the grid and accelerating that process. There are a lot of things like that, just really meaty, substantive things that are proceeding apace.

I would also say that military adoption of AI has impressed me to the upside and, generally speaking, the military's direct involvement in industrial policy and boosting startups and U.S. manufacturing—which is talked about in the action plan—startups that are doing innovative things with physical autonomy and stuff like that, is all doing great. I think another main thing is that the action plan talks a lot about adoption more broadly, and I think AI adoption in America is actually going pretty well, all things considered.

Now, of course, that's all the nice stuff. More critically, I think it would be hard to say that the administration has carried itself according to what I think of as the spirit of the action plan. It's not my job to say what the spirit of the action plan is. Ultimately, it's their administration, right? For example, a big pillar, a big principle of the action plan was the notion of exporting American AI and getting it adopted all across the world.

It seems hard to imagine how that's consistent with global export controls on frontier models imposed with 90 minutes' notice on all non-U.S. persons. That's the kind of thing that, when we were out in the world—and in my life after government, there are definitely a lot of international trips I do where I am engaged in quasi-diplomatic work on behalf of the United States as a private citizen, but as someone who's trying to explain what we were thinking with things like the export promotion work and our whole strategy there—the biggest concern you hear from people abroad, especially in Europe, is, “I just worry that you Americans are going to turn off the models at some point if you get mad at us.”

When I was in government, we were trying to assuage this concern. When I left government, I spent time in India at the AI Action Summit earlier this year. In many places, in many quasi-diplomatic engagements, I said, “No, don't worry. We don't want to do that. We want an ecosystem, blah blah blah blah blah.” And then, of course, the administration goes and does it and basically confirms the biggest fears of a lot of people internationally, and that doesn't help. That certainly doesn't help.

I think, actually, that relates to one other thing, which again I don't know if it would have been possible, but one thing that I feel as though the action plan was relatively silent on was the issue of AI governance. That's a big part of what I worked on before and after, and it's not really referenced. I think my reasoning for it at the time would have been, number 1, a tough Overton window within the administration at the time, and number 2, a lot of that, in my view, is ideally legislative.

The action plan was not supposed to talk about that. It was supposed to be just things the executive branch could do, not new laws. But I think there could have been more explicit material in the action plan about, “Hey, okay, at some point things will get scary, and what should you do? Here's how not to panic.” There could have been more of that, because I think that we are seeing—

It's funny, though, because in the end I do think there's just a distinction, right? There is this huge array of civil service bureaucrats, just full-time career civil servants, and then there's low- to mid-level political staff. They all read the action plan and are implementing it, I would say, in quite a good way.

Then, of course, there's the very high-level people who don't necessarily read every strategy document that the administration comes out with, and they're fundamentally very reactive. This kind of stuff happens, and it's, “Oh my God, we've got to do something about this,” and they're not thinking about, “What would the action plan tell me to do?” That's not at all what a cabinet secretary is thinking.

Something that is amusing to me is that I think we are watching the administration—or the senior-level people in the administration—reinvent some of the ideas in the action plan around use cases, building technical competence in the government, all that kind of stuff, third-party evaluations, all this. I think we're seeing them reinvent those things from first principles.

So I'm still optimistic, but, yeah, I definitely think that there have been substantial ways in which the administration has departed—ironically, both in the direction of not taking the risks seriously enough and then also overcorrecting, taking them too seriously, but reacting in ways that don't actually deal with the risks.

Look, in the end, I'm inclined to give grace to people and say what I hope we're in right now is a high-neuroplasticity phase of policymaking. I really think that we might be in a very different world in 3 months. But certainly, if you were to look at the headlines, you'd be like, “Ah, it doesn't seem very consistent with the action plan at all.” And I can't deny that.

So obviously, one of the biggest moments—and I don't want to rehash all the politics of this, because you've commented on it extensively—but one of the biggest freakout moments was when the Department of War declared Anthropic to be a supply chain risk. I'm struck by the fact that it seems like we're memory-holing that kind of thing now, where, as far as I understand, many areas of the government were involved in using and testing models. There's still a lot of Anthropic in the government, as I understand the situation. Correct me if I'm wrong, but how should we understand where that whole supply-chain thing is today? Are we just going to pretend it never happened, or what?

The broad way I would describe the American presidency, really since Obama's second term, is that Obama famously said he was faced with an intransigent Congress, to put it generously. He was faced with a Congress that wouldn't pass any laws, and so he said, famously, “I have a pen and a phone.” What he meant was, “I'm going to do executive actions. I'm going to push the limits of executive actions.” That began a kind of autocatalytic process in which every president pushes the bounds of executive authority in various ways that get tested in the courts.

What will happen very frequently is that you'll see a headline where it's like, “The president did this thing that is unprecedented with executive power, and it's being litigated,” and then it goes through a very long litigation process and most people lose track of it. The lawsuits don't stop; it's still going on, right? Anthropic had its oral argument in front of the D.C. Circuit last month, and I think we're expecting a ruling in that case at some point soon. After that, if Anthropic loses, Anthropic will appeal, I'm quite certain.

Interestingly, the Trump administration is actually very savvy about when to appeal things and when not to. They're pretty good at reading between the lines: “Okay, yeah, that one was probably illegal. We're not going to appeal that all the way up to the Supreme Court, because we're going to lose there.” They're actually pretty good at that. But I think the Trump administration actually thinks it will win this case.

The litigation is ongoing, is my point. If it keeps going, it would not surprise me if, by the summer of 2027, there's a Supreme Court ruling of some sort on this issue. Even if that ruling is the court denying to hear the case—which is the most common thing the Supreme Court does—that's also going on.

In terms of government use, it seems as though, basically, after Mythos, I think the message that Anthropic received was that the supply-chain-risk designation applied to its contracts with the Department of War proper. Within the Department of War, they really are winding down Anthropic and have been doing so considerably. It wouldn't surprise me if they're 100% off Anthropic by the end of the year, or a year from now or something.

Throughout the rest of the government, I think the message was that the supply-chain-risk designation doesn't apply to any other government agencies. If there are other government agencies that want to use Anthropic, that's fine. There's also, technically speaking, the National Security Agency, which is part of the Department of War. But it seems as though not only does the National Security Agency have a contract with Anthropic, but, if reporting is to be believed, Anthropic's red lines around domestic mass surveillance and autonomous lethal weapons were honored by the National Security Agency.

To me, that's a good example. I think Americans are not that good at tolerating ambiguity, but a little bit of ambiguity—or maybe even a healthy amount of it—is just an intrinsic part of this whole process. So, yeah, it's still happening. The supply-chain-risk designation is still being litigated. I think Anthropic contracts are indeed being canceled at the Department of War, and at the same time I think other use of Anthropic in the government is going fine.

The U.S. government contains multitudes.

You were also—I mean, you were extremely critical of that whole situation. And I think you were less critical, but still somewhat critical, of the recent executive order and the move, as I understand it, to take certain AI testing and characterization responsibilities away from CISA, which I think now has an uncertain future, and move those responsibilities to the NSA, where they may or may not already be classified information. What's going on there? Why are we taking things away? Is this as simple as it being a Biden project, and so we don't like it, or is there more going on there than meets the eye? And why are you concerned about testing models that you don't know exist against standards that can't be disclosed? How do you think that turns into a problem for the public?

The reason I'm critical of the administration here, in terms of where they're going—I was critical of the cyber executive order when it was signed because I anticipated exactly this. Let me just level-set for your listeners for a moment. What does the cyber executive order do? One part of it is, “We're going to create a variety of procedures by which we're going to patch vulnerabilities in critical software systems.” Great. Thumbs up. I don't think anyone can object to that. We can argue about how useful those programs are going to be and how good the implementation will be, but we'll see.

The second thing is that it created a voluntary pre-deployment program—a testing program 30 days before release—whose details were to be classified, primarily classified, and primarily run by the intelligence community. Within that, practically speaking, the NSA is the primary agency, since it has the highest level of cyber expertise. They really are quite excellent, by the way.

I think the reason I'm concerned is that this is setting up a potentially very bad future where access to frontier models is gated. It's all kept secret, and the public doesn't really know what's happening at the frontier. The government is making a bunch of decisions that maybe the public doesn't even know about—whether or not to restrict certain capabilities, and what to do with those capabilities.

It feels to me like, if you believe that what's happening right now is one of the most important things ever to happen in the history of technology, then not only do I think, just intrinsically as an American, that my gut instinct is that the public has a right to know what's going on, but, number 2, I actually just think that it's not some trade-off. It's a better world where the public knows, and to the extent possible, it's a better world where the public can access frontier capabilities.

First of all, I think government monopolization of frontier AI is potentially how we get very scary outcomes from a civil-liberties perspective. That's not, by the way, a criticism I would make only of this administration. I would make it regardless of who the president was. If you erased from my brain who the president was, what party he was in, and what his name was, but I had everything else in my brain and you told me about this, I would say I was concerned about that.

Another thing is that dealing with a society and a civilization is a kind of information-processing system, right? All the humans in the country are parallel compute, and we're all trying to process what's going on here. While there are a lot of things we don't have good answers to yet, I actually do think that the community of AI policy people has made reasonably good progress since 2023 in terms of figuring out, okay, how practically should we be dealing with models of the Mythos-level capability?

When things are public and there's a legislative process, for example, that is informed by lots and lots of robust public input, that stuff can make its way in. We can take advantage of this kind of parallel compute. When you centralize everything and make it private, it's much more brittle. It's a bunch of people who, as I said earlier, often don't have a lot of context for AI. They have a million things on their plate because they're high-level government officials, and they're just improvising and making decisions in an improvised fashion. I think that leads to subpar decisions.

I think it leads to wasting time because what we're watching right now is the administration speed-running that mentality. They remind me very much of where D.C. was in the spring of 2023, when ChatGPT had just come out: “Oh, we're going to regulate the hell out of this. This is really dangerous.” Then things softened, and maybe they softened a little too much, but ultimately we met in the middle. There was a change in the vibes that started in 2024, and I just think that right now, by putting this all in the echo chamber, the administration is operating with a pretty small number of voices contributing information and insight to things...

I just feel like we're not leveraging the best that we have, and I think that's the biggest problem. That's why I am critical, and I am very worried about the direction of policy. But unlike the supply-chain-risk thing, where I think it was just totally an own goal—just a totally unforced error, like, why did you pick that fight? You didn't need to pick that fight. You could have fixed this in a million different ways.

Even if you take the government's concerns in that issue seriously, you could have dealt with that in a thousand different ways. This is more—yeah, I'm not surprised. I'm not surprised things are going this way because you are building this thing, you're improvising an AI governance regime from scratch, and it's being built by 20 people, 15 of whom don't have a ton of context for AI. I'm not surprised it's working this way, and I'm just pointing out the meta-problem: We need to bring this out into the public. We need to have Congress involved. We can't just—this is not going to work.

There's no point in being highly adversarial and critical there. What do you want them to do? What do you want these people to do? I don't blame them, but I do ultimately think that we need to make things more public.

On the note of parallel processing, how about the role of the states—our laboratories of democracy? A lot of the proposals that I understand you have favored or even championed, including mandatory safety-plan publication, certain other transparency measures, whistleblower protections, and even a sort of Fathom-style public-private regulatory hybrid structure, have all happened in different states to a remarkable degree in a pretty short period of time. How bullish are you on the states?

There have been really meaningful wins for the general notion of private governance that I started to work on, really, post-SB 1047 veto in late 2024. Of course, it's not just me; a lot of other people have worked on this stuff. Independent verification organizations would be third-party private bodies that would evaluate a lot of these things.

Right now, with the government, they're concerned about the jailbreak—the potential jailbreak of Mythos or Fable [?]. It would be great if there were expert bodies that had looked into this, really probed it, and certified, saying, “Yeah, there are jailbreaks, because there are always jailbreaks, but our calculated risk is that these jailbreaks are not severe enough to rise to the level of concern, and we can certify Anthropic as conforming to safety best practices,” or whatever. That's the kind of thing I think there have been substantial wins for.

Two of the three big AI companies have published multiple documents that are favorable to this general notion: Anthropic and OpenAI. A bill to mandate auditing in frontier AI companies passed in Illinois earlier this year. The state of Connecticut and the Commonwealth of Virginia both passed laws earlier this year that specifically authorize either studies or pilot programs for independent verification organizations. There's also a bill pending in Ohio, by the way, which would be the most robust implementation of independent verification organizations yet. There's more momentum in that regard than I would have guessed a year ago.

In that sense, I think the states-as-laboratories-of-democracy idea is working fine. It's also worth noting, with respect to the frontier AI safety laws that have passed, that the states have taken great effort. There was a transparency bill in California, SB 53; in New York, it was RAISE; and in Illinois, it was SB 315. The Illinois language adds an auditing requirement, but the transparency language across those 3 states is remarkably similar.

I'm quite happy about that. That's not creating a patchwork. Those are the states converging on a common framework. We'll see if it works and how well it works, but it's states converging on a common framework, which they do from time to time.

There are a lot of other areas of AI that don't get as much attention on Twitter, that don't get as much mindshare, where I think the story for the states is less rosy. Things like consumer protection and algorithmic pricing are not getting as much attention. The number of laws around synthetic media and deepfakes that exist in this country now is just crazy. There are hundreds of them now, and the net effect of that is probably to create a fairly confusing political environment.

I think one thing that's really problematic that we're starting to see bubble up from the states is occupational-licensing protections. States, including Illinois, have done this, saying, “We are going to define mental-health services as exclusively something that can be provided by humans.” If a chatbot so much as asks you how you're doing, it is engaging in mental-health services. If you say to the chatbot, “I'm sad. Can you please help me?” the chatbot is technically engaging in mental-health services, and that's illegal.

We'll see. States often vary in terms of how rigorously they enforce laws like this, but it's not good to have that kind of stuff on the books. Strangely enough, the area that gets the most attention—the frontier AI safety stuff, where a lot of the pro-preemption crowd also focuses their energy—is actually the area where the laws are best sculpted. The laws are very well sculpted. They often have the support of the AI industry, and they're often designed specifically to avoid the patchwork complaint, which is a legitimate one.

For some reason, most of the people who are super pro-preemption focus on these laws. It's not only that, but these laws are dealing with really urgent problems, like cyber and bio, that are clearly not fake anymore. We can't have that argument anymore. They're clearly a real thing.

They're not focusing on all these other areas where the states actually are creating patchworks and creating complex compliance obligations that might specifically be really complicated for startups to deal with. I think the state issue is mixed in that way, but the people who support a federal law, including me, are not helping themselves because they're talking about this issue in largely the wrong way.

One big surprise, I guess. First question, at a high level: What have been the biggest surprises for you? I'll offer my biggest surprise, and you can react to that and share your own biggest surprise.

My biggest surprise is that the administration eased the export controls on chips. That wasn't shocking unto itself, but the real shock is that China doesn't want to buy them. We had all this debate around to what degree we should—are we being bellicose? At least, I was asking that question while trying to do these export controls. Finally, they get eased, and China's like, “Ah, no thanks. We're going to just build our own industry, and you guys can keep the chips.”

What's going on there, and are there any other surprises that rise to that level for you?

That particular development doesn't surprise me that much because China's system is very—actually, realistically, we are becoming more like China in this regard. One thing about China is that when China announces a new policy, you need considerable expertise to understand it. There are people inside the US government who specialize in this, and they do not share their opinions publicly. It's actually very hard to get good analysis on things like this in the public discourse. It's one of the things I miss about government.

There's what the policy says, and then there's what they're actually going to do, which are importantly different things. The policy, I think, is a matter of national pride for China: “We are building our own AI chip ecosystem, and we don't need the Americans anymore.” I think they like sending that message to the world, and they like sending that message to their own people. There's some aspect of that.

But then there's what actually happens, because while that's going on, China's system has lobbying too. I guarantee you that DeepSeek, Alibaba, Zhipu, and all these other people are begging Beijing for access to American chips. The policy planners in Beijing are probably factoring that in. There's some amount that's being sold. I think we now know that there are some chips being sold.

But, yeah, no, they're going to keep restricting it, and that might be a big goal on their part. I guess we'll see. In terms of US-China relations, nothing has especially surprised me. I anticipated that, by now, the Chinese state would have woken up to the catastrophic-risk issues and would have started pushing back on the open-source strategy.

I said that by the end of Q1 2026, DeepSeek's top model would not be open source, and that prediction was wrong. I still think it's going to happen at some point, but we're not there yet. The Chinese state seems more concerned about labor issues than it seems concerned about catastrophic risk. They're less catastrophic-risk-pilled than I would have guessed if you had asked me a year ago.

On the technical side, nothing has really surprised me. One of my hobbies is paying attention to these weird subgroups of very normal people who use coding agents.

So there’s this community of homeschooling moms who love Claude Code and OpenClaw and stuff, and they’re using it to do all kinds of stuff. I love that. I wouldn’t really have guessed coding agents would become so popular.

The only other thing on the technical side that really surprised me is that I did not expect the world-sim stuff in the physical world: the models where you can simulate a 3D interactive environment, basically like creating a first-person open-world video game, but for arbitrary settings in the real world. I did not anticipate that, because the way those models worked for a really long time was very dreamlike. You could create the world, but it was like the neural network was creating it in real time, so if you turned around and looked at something, then turned away and went back and looked at that thing again, it would be totally different. It didn’t have permanence.

Then, one day, it just worked. It was like, “Oh, wow, we just have permanence now,” and it just worked robustly. That substantially increased my timelines for robotics working, because it’s very clear that you’ll be able to make synthetic-data pipelines. You’ll be able to use human data as a baseline: make people wear Apple Vision Pro and use that as the baseline. You can bootstrap from there to synthetic data in all sorts of world-sim settings, and then probably just sprinkle on a little bit of really high-fidelity data, like people wearing gloves with electrodes in them to sense muscle movements and whatnot.

It’s very clear that dexterous manipulation in robots is going to be solved. The second I saw the world sim—it was a Google DeepMind model, like, last summer—the second I saw that, I was like, “What? Okay, dexterous manipulation in robots is going to be solved in 8 months.” That caused me to change my research agenda a little bit after I left government and accelerate some of the work that I was thinking about for robotics.

Yeah, it’s all happening.

Yeah. Jim Fan from NVIDIA’s recent little 20-minute keynote about the parallels between the path that he expects robotics to take and the path that LLMs have taken—I think it’s—

Oh, the Sequoia talk.

It’s must-see TV for sure, and it definitely has me convinced as well.

Yeah.

Also, shout out to Jesse Jane from the homeschooling-mom contingent. I love the stuff she’s doing and try to borrow from it as much as I can as well. Last night, for the Mexico–Korea World Cup game, we printed out little packets for each country, with a little bit about each one. We had snacks from each one. It’s amazing how much you can enhance your mundane daily life with these tools. That should not be forgotten, even as things get intense and, in some ways, fraught.

A really good example of this, just as a sports thing: I remember when Opus 4.5 came out, right when the models were getting really good. It was December, when no one watches basketball in December, but I do. I have League Pass, NBA League Pass, which is the way you watch all the games.

The problem I had was that I never really root for any particular team. I just want to watch a good game. Sometimes there are 8 NBA games on at the same time across the country. I built this little dashboard with Claude Code that ingested live data from all the games and then did the Nate Silver-like speedometer thing with the odds of it being a good game, basically. I had created some heuristics for that with the model. Anyway, it was a fun little project, but it massively improved my life.

Well, that pretty much brings us to the present, and the present moment is Fable and the Fable Ban. In the brief time that you had Fable, before we get into the politics and policy of it, what were your impressions, and how much are you missing it?

My impression was that it was a fiercely intelligent model and a real step up in intellect. A lot of people have made o3 comparisons, where o3 was the first model that felt like this really cracked genius. It’s very funny, because when o3 first came out, I had this feeling that the hedonic treadmill was going to stop at some point: This thing is always going to feel so smart to me. Then it actually does, and now I’m sure if I used o3, I would find it rather dumb. Still charming, of course, but Fable was another moment like that for me.

Unfortunately, I was really busy those days and didn’t have any time to use it in coding-agent settings. I had it open in Claude Code a few hours before, and I was going to do a project from my hotel room while I was traveling. I got distracted by something, and by the time I came back to my laptop, it was off.

But I did use it for some knowledge-work stuff. In particular, I’m a party to a proceeding that’s going on before FERC right now. I’ve provided some expert testimony on it. It’s a complaint before FERC to remove a procedural regulation of Order 1000, so it’s a boring thing. Someone wrote a rebuttal to my testimony—the other party in this legal hearing hired an expert to write a rebuttal to me. I had Fable look at it. I have, like, 40 pages of testimony, and then there’s, like, a 70-page rebuttal. I had Mythos go and read it in Cowork and do research and stuff.

I did not use its writing output, but oh my God, this model demolished this. I felt so bad for the guy who wrote the rebuttal. I was like, “Wow, the model demolished this dude in a way that I mostly couldn’t have.” I found it to be fantastically intelligent. I wish I had been able to use it more.

But, yeah, I am missing it. It is weird to feel like we’ve gone backward, but also, welcome to the government being involved in things, right? It’s a good little lesson in political economy. That’s what it feels like. Usually it’s too abstract or diffuse. I’m like, “Oh, the government makes things more brittle and makes the world a little bit dumber in various ways.” This is the problem with state intervention. This is a good example of it. It is literally the case that if you were a user of Fable, your world became dumber in the last week.

Yeah, it’s been rough for me personally. I experienced this once before when I did the GPT-4 red team—

And then you went from GPT-4 down to—

Whatever it was, text-davinci-002 at the time. It was just like, “I don’t even want to touch any of this stuff until I get the real thing back.” I feel that way again, to a lesser degree, now, but definitely the taste factor is really where I felt it.

It’s obviously amazing at coding. Opus 4.5 is superhuman relative to me in coding already. I still stand to gain a ton from it, but the jump to Claude and the ability to start to mind-meld with it a little bit, in a way where I really felt like it was kind of getting me on my level in a way that I hadn’t really felt before, is—I’m definitely missing it quite a bit.

So when do you think we get it back, and what’s going on? From the outside view, I think the consensus take is that the longer it goes without us really having a good explanation for what they saw that scared them—and at this point it’s been a week, which is a long time—it starts to feel a little bit reminiscent of the OpenAI firing Sam Altman episode. It’s like, “You’ve got to have an explanation here, guys,” or it becomes clear that this is not super well justified.

Is that basically the view that you have, or do you have a more empathetic view of where they’re at? How do you think this gets resolved?

I think what’s happening here is that there are 3 factors playing into the government’s reaction. One is genuine concern about safety and security. The second is a fairly broad lack of context for frontier AI and the sort of information that you need to make a good risk calculation, which is driving the security concern to some extent. So there might be some legitimate security concern, and there might also be some not-so-legitimate security concern, but it’s being driven by this general lack of context.

And then, third, we can’t deny that there’s some political dimension to this. Even if it’s not—I wouldn’t even say that there’s a conspiracy theory inside the government to do this to Anthropic. There might be, at least among some, but it might be more just the general political status of Anthropic and the fights that the administration has been having. It colors the reaction of important people to the news of a security vulnerability in a way that might not have happened if it were a company that the administration felt more warmly toward.

All 3 of these things definitely feed into one another, and I think they probably all explain what’s going on. They’re all ingredients in explaining this situation. The thing that I don’t know—and probably even if I were still inside the White House, I wouldn’t fully know—is really just in what ratio those 3 things come together.

One thing I think it’s worth being clear about is that my read of this situation is not that the US government is saying, “It is our policy from here on out that if your model has security vulnerabilities, we will do export controls on non-US persons.” I don’t think that’s the policy there.

I think what probably happened is they decided, “All right, we’ve got to cut this thing from the market,” and this is the only thing we can think of that we’re pretty sure will actually get the darn thing off the market. So I think they basically just reached for the tool that they thought would do the job.

I don't think they're thinking of that as a universal policy that they're announcing. But the other thing you have to consider here is that the administration's story has changed—and it has, in fact, changed. It's funny; this is similar to the supply-chain-risk thing. The first version of the story was: We had security concerns, and we wanted to get Dario Amodei on the phone to talk about them, but we couldn't get him on the phone in a timely manner. That was the supply-chain-risk thing.

I remember the same thing with Emil Michael, the under secretary of war, who was a main character in that whole affair. He was complaining in public about how Dario wouldn't return his phone calls immediately and that it took hours for Dario to get on the phone. Come on. There's a grudge aspect of this, right? There's a “who's the bigger monkey?” aspect to this, right? When I'm in the government, I'm Mr. Government Man, and when I call, he has to call me back. I don't know. This is a thing that happens a lot in D.C. People in D.C. play these kinds of games all the time.

Then it became: No, the security risk is that this jailbreak is legitimate, and we needed to know about it. Then 36 to 48 hours later—or maybe even more, maybe more like 72—we started hearing that, actually, no, the reason we're doing this is because Anthropic provided the model to a Chinese-linked company. You're grasping at straws here, because if you imposed export controls because Anthropic provided Claude to a Chinese-linked company and that was your primary concern, why did you not say that on Friday when you did the thing?

They were also describing this as something that had happened about a month before, because Anthropic had expanded access to a tranche of companies that included some international companies. Anthropic announced this publicly several weeks ago. They said, “Yeah, we're expanding Claude access to some U.S. companies and also some allies and partners of the United States.”

The company in question, by the way, is SK Telecom, which is part of the same conglomerate that owns it—it's the SK Group, which is one of the largest chaebols in Korea and which also owns SK Hynix, the leading producer of high-bandwidth memory. In general, Korea is a really important partner to the United States in the semiconductor-manufacturing ecosystem, and the notion that we would want to harden its telecommunications infrastructure seems quite reasonable to me. It seemed extremely reasonable that we would want to do that.

They threw that out there, and it feels like an administration that—I don't know in what ratio of the 3 things I said—basically panicked, reached for the first thing it thought would actually get the model taken off the market, and then created justifications post hoc. Part of the reason I find this issue frustrating is that it's just very hard to analyze, because there's not a lot of policy substance here. It's just the id.

Whether that makes you a glutton for punishment or just somebody who is destined to be some sort of main character yourself, that brings us to the very present moment, where you have just announced that you are going to be joining OpenAI and building a new team to help shape the company's positions on and influence over frontier AI policy. Tell me how you came to that.

You alluded to the fact that the last year was also a big year for you personally. You had your first child—congratulations again. You traveled a lot, from what I understand, and I'm sure that was exciting and interesting. You wrote a lot. Generally, you had a taste of the good life, I would say—a kind of freedom and ability to pursue your curiosity.

Now you've decided that, as great as that was—as great as I assume it was for you—you're going to take this job. Tell me how you've gone through the process of deciding that this is what you want to do next. Then, obviously, we'll talk about the role and the mission you're going to have as you start up.

The time period since I left government—it's been, I guess, about 10 months—has been a really wild time, and I feel tremendously lucky. I've been able to be in quite a lot of interesting rooms and meet a lot of really interesting people, and I've had a lot of great opportunities.

It's been strenuous. It's been tough. The workload has not really changed from when I was in the White House. It has not been the sort of luxurious think-tank life that some people imagine. It's been quite a lot of work.

I think the most important thing is that a lot of my work centers on the frontier lab itself as a kind of institution—a new center of political and economic power. I think of it almost as the emergence of banks. Banks have existed in various forms for a long time, but Italian city-state finance—or the emergence of the financial sector—was one of my favorite periods of history to study: the emergence of a recognizably modern financial sector in the Dutch Republic and in Britain.

It feels like this is a moment where something like that is being created, and I think there are 2 aspects of this. One is that the institution and the way that it relates to the government and the broader society are really important.

Number 2 is that, if you go back and look at the financial-services sector in the early Dutch Republic or in England, we were talking about proto-modern states at that time. We're not talking about states that were capable of overseeing everything. There was no SEC—the Securities and Exchange Commission, right? There was no Bank of England. These things didn't exist.

But if you're going to be engaging in options trading or trading financial derivatives, you do actually need there to be common rules to define that and common governance, because otherwise you can't engage in those transactions with trust. Similarly, there's going to be a need for governance in this field that the government itself simply will not have the capacity or expertise to catch up on. Practically speaking, this is going to have to happen within the companies themselves and also within the private governance stuff that we already talked about. There's going to have to be a lot of private development of governance norms and standards in this field.

The third is this notion that advanced AI itself will be an instrument for doing statecraft, governance, regulation, and all these kinds of things. In the same way that financial services are, we use financial services as a way of achieving policy objectives that, on paper, have nothing to do with banking. We do that all the time, and it's just because money is so fundamental to doing anything. I think AI will be fundamental to doing anything in the future, too.

All 3 of these things really interest me. The struggle that I keep having is that, practically speaking, I've sat in the White House, and the White House didn't feel like an especially fruitful place to think about these things. Then I sat outside the White House with what I would say was, in the end, pretty good access to people and information and a pretty good network of people that I could draw on.

I've spent a lot of time reading and thinking and doing things like this, but I ultimately don't feel like I can get beyond these abstract intuitions without actually being inside the lab itself, without actually doing some of this work myself. That is the central reason that I started thinking about going into a lab.

Also, of course, policy is just becoming more and more important, and it feels like we might actually be setting the foundation for AI policy in the next 18 to 24 months here. That seems plausible, and that's going to be really important, too.

I had been thinking about it but not really acting on it because I was writing this book and doing all these things. I didn't really have time to pursue it. Then, as a happy coincidence, OpenAI approached me a little while ago and asked me if I might be interested in doing something like this. That's how we got here.

Can you say a little bit more about why you think it's so important to be inside a frontier lab? I kind of share this intuition. I feel like the world is—I don't like this, but it does feel like we're approaching a tabletop-exercise scenario where the number of institutional actors that really matter is becoming small. It leaves me in an uncomfortable position where I'm thinking, “Geez, do I need to join one, too?”

What exactly is it that you think being inside changes? Is it better visibility into the road map or capabilities, or something else?

First of all, let me say one thing real quick about what my team will be and how it's different from other teams inside OpenAI or teams that might be more familiar to people. It's an interestingly shaped team, and it was really a pleasure to work with OpenAI's senior leadership on how we would actually shape this team together. What would we do?

It's kind of a boutique operation in many ways. There's a team at OpenAI called Global Affairs, which is run by Chris Lehane, that does what you would think of as the traditional policy and lobbying operation that a company would have.

And that team continues, and it’s, as far as I’ve been able to tell, a fantastically capable team. I mean, they’re dealing with public policy that’s coming at them from all 50 states, from the federal government, and from all over the world, right? They’re trying to shape a million different things like that. But if you think about where we were a year ago, people were really barely even talking about kids’ safety. People were not talking about data center electricity or water use, really, a year ago.

In the world of June 2025, the best model was o3 a year ago, right? It’s just a very different world today. And so the job of this team, in part, is going to be to look out 6 to 12 months and say, “Where are we going? What do we think we’re going to be dealing with?” And then how can we shape the policy—both the present-day policy positions of the company, but also future policy? How can we develop policies to try to be proactive in dealing with where we think we’re going to be in 6 to 12 months?

To do that, I anticipate that a very large portion of my time is going to be spent jamming with the technical staff on where things are going and what that’s going to mean and stuff like that. So that’s one thing: you do need to be able to access detail—not, like, “The models will get better,” but really get into the weeds on internal deployments, many different things about where the capabilities frontier is going to be going, and what will be different about the world in a year versus today. That’s one.

Another one, I would say, is simply that question of internal deployments, right? In particular, if we are moving toward a world in which, for some combination of regulatory risk, security concerns, compute constraints, and other things, we may well be moving to a world where, presumably, Mythos 2 is not that far from being done—trained, right? I mean, it had Mythos checkpoints in January, so 6 months ago, right? Same with OpenAI, I’m sure, though I don’t have any internal knowledge yet. I can speak with total ignorance. It’s great.

Safe to say they’re training another model.

Yeah. Safe to say they’ve probably got some. Fundamentally, the internal deployments of these models are not covered, at least until we have a robust system of supervision and auditing or independent verification. Mechanically speaking, all government regulations that states or governments think about are triggered by public release or public deployment. But I think a lot of the really important decisions are going to be made with respect to internal deployments.

There’s going to be a combination of objective determinations that you’re going to want to make, and also probably just some gut calls, some judgment calls, about what recursive self-improvement ultimately means, right? What does it mean? How should we be thinking about it? Again, it’s just very hard to answer.

I can ponder that stuff in the abstract without any inside knowledge, and I can write about it on my Substack. I’ve done a little bit of that, and maybe that’ll influence some people; maybe a couple of people who matter will read it and that’ll influence them. But in the end, I think you really want to be getting your hands dirty and helping shape some of these decisions with researchers, with the executive team, and with many other people.

I want to get into more details on both the plan for RSI and the internal deployments, but maybe just zooming out for a second first. How do you understand your duty as you start this role? When you sign up to work for the U.S. government, you swear an oath to the Constitution.

When you go to OpenAI, we have this mission of making sure that AI benefits all humanity. Do you think of yourself as sort of signing on to support that mission in the same way that you might have previously sworn to uphold the Constitution, or would you describe your personal objective function as being in some ways more mixed than that? Is there a term in it for OpenAI winning? Is there a term in it for the U.S. winning? How much complication is there around the core idea of benefiting all humanity?

So, I mean, I think that mission is really serious; it’s something that people inside the company take quite seriously. One thing that I don’t think is a publicly announced part of my role, but I think it’s probably okay for me to say, is that inside OpenAI there’s a body called the MAC, which is the Mission Advisory—I think it’s either Council or Committee; I forget—but it’s a body of researchers, public-policy people—you know, the Global Affairs people—and a wide variety of people from around the company who collectively make decisions about things relating to policy, some internal governance decisions, and things like this. Part of my job will be sitting on that body.

So definitely, in some sense, I take that mission seriously, and I think OpenAI culturally does as well. Of course, the problem is, as with anything, how do you decide what’s what? It’s a broad mission that is open to a lot of interpretation, and there’s a lot of ambiguity in the world. So what does it really mean?

That’s where, in part, one thing that was very important to me in taking this role was that I could maintain a public writing presence that would be independent of any sort of editorial review by OpenAI. I really don’t think that what I’ll experience on the inside is going to be what is sometimes depicted in AI safety circles: this almost cartoonish depiction of OpenAI as engaging in a grand villainous conspiracy or whatever. I would be strongly surprised if that is what I actually experience on the inside.

What I would guess instead is that there are people with good-faith disagreements about how some particular set of decisions relate to the broader mission, and that there may well be times when I ultimately disagree with the call that was made for various reasons. I think my ability to communicate publicly about things like that and explain where I stand, without having to go through an editorial review process with the company or fearing for my job, will matter.

One of the great things about OpenAI is that it still has the DNA of being a sort of Xerox PARC-like research organization. They’re actually fine with lots of internal dissent; there are a lot of great debates that happen inside that organization. I’ve always gotten that sense observing it from the outside, and so I don’t think this is going to be culturally too dissonant. But yes, my ability to publicly disagree with certain policy positions, I think, will matter, and that’s part of why I preserve that.

So I guess what I would say is that ultimately, I still feel like what I’m trying to do is get this right, trying to help shape this whole transformation well for the country and for the world—probably the country first and foremost. That’s one area in which I’m very different from people from the East Bay and San Francisco: I’m a patriot. I identify as an American, not as a citizen of the world. I am an American—civis Americanus sum.

For sure, one thing that factors into this is that OpenAI is a company. I will be helping OpenAI set its strategy, which is different from setting the strategy for the abstract AI industry, right? It’s OpenAI, a company that exists in contradistinction to other companies in the field, and there are competitive considerations and things like that. That’s fine. I’m a competitive person.

I think the competition will make things better, in fact. But at least for them, on average, it’ll make things better.

Let’s go back then to RSI. The notion that competition will make things better is definitely going to surprise some ears in the audience who are worried about arms-race dynamics between companies, between countries, and any number of different configurations.

And for my money, the race to RSI between at least 2 companies is probably the most objectionable thing happening in the AI space right now, because it does feel like—and the researchers that I have heard speak candidly about it seem to share the intuition—that this is sort of a phase-change moment beyond which things could get really weird. It’s going to be super important to set everything up right and get the initial conditions all right, and they’re still not that confident that it’s going to go very well.

Yeah.

So, I guess, how do you understand what the safety plan is? How do you understand how committed the companies are to an RSI-or-bust approach? OpenAI famously has public timelines for when they want to have the automated intern and the full-fledged ML researcher.

What do you—how do you understand the plan, and do you think it is anywhere close to being up to the task at this point?

Well, one thing I want to say: when I was talking about competition, I was talking specifically with respect to what I will be doing. The goal of my team will be to have really fantastic intellectual output that is like, wow, this team, if it were not part of OpenAI and it were just its own little think tank, would be one of the most interesting think tanks in the whole country. Everyone would be paying attention to it.

I actually just want it to be a really superb team that is producing policy and sort of public-interest-related work that is as good as anything that any of OpenAI's competitors are producing. But that's what I meant by competition, to be clear. I think there is a good example of straightforwardly healthy competition when it comes to RSI.

I think there's certainly a lot of unknowns here. I put my base case for what RSI means, at least in the earlier innings, as something to the effect of: recursive self-improvement is a part of every technology. Every general-purpose technology has some aspect of recursion to it by the very nature of generality, right? General purpose means one of the purposes to which it can be applied is itself.

In some ways, I don't see recursive self-improvement as some big break from the history of technology. I see it as being—actually, it would be surprising if there weren't recursive self-improvement in AI, right? I also think we've been doing recursive self-improvement in this field for a long time.

Some people imagine there's going to be some big break moment. You could argue that really since GPT-4, we've been using the models to make the models better, since at least GPT-4. I'm sure that if you actually went back and looked through the history of machine learning even more, I bet you that actually goes back even further than that.

Some people imagine there being this sharp, discontinuous jump in terms of what RSI means. I certainly think that is plausible, but it's not my prior, because your prior should just generally be that there is always more continuity than discontinuity. I'm sure on this podcast I've said before that there is always more continuity than discontinuity. This is always the case. Your prior should be against a massive discontinuous leap. It is plausible.

I think step number 1—and I'm not in yet, so I don't actually know because I haven't looked at the roadmap yet—but step number 1 would be to really measure twice and cut once and figure out what we think this might mean. Try to refine the probabilities there, at least in my own mind: are we talking about a discontinuous leap that happens very soon, or are we talking about something that actually is smoother in some way?

I think the chances of a discontinuous leap are high enough, no matter what, that you need to be planning now. Even if your credence that it happens is 20% or 10%, that's high enough that you should be making plans now for what you want to do. And there we get into interlab coordination on things like the slowdown-pause thing, right? There's what would be the mechanisms of that, and under what conditions would it be triggered?

Again, I'm pretty skeptical of such notions. There's a lot of policy planning I did for the U.S. government that I wrote down and put in various places, but it's not in the action plan. It's scenarios, right? We have to be prepared for a wide range of scenarios here. Similarly, I think there's going to be some aspect of that where we need to be ready and need to have advanced thinking on all that stuff.

So it's 2 things. At what point—what are some triggers that we can set in advance for whether this is going to be a discontinuous leap specifically, and how can we refine that question to make it as specific as possible? Then, in the event that those triggers happen, what is it that we would do? At what point do we go to the government?

One proposal that I'm a fan of is the FTC, the Federal Trade Commission, writing what's practically called a no-action letter, where the FTC would write a letter and say, look, if you guys coordinate for these very specific reasons, we're not going to consider that cartel behavior, and we're not going to enforce that.

I think that's plausibly a good step to make. That at least opens optionality. Though I also understand that you have to be really careful about how you scope that, because look at what Anthropic did, right? Anthropic undermined the case for this dramatically just with the Fable safeguards: “We're going to degrade your outputs in the name of safety,” which is very clearly a consumer-protection violation.

Can you imagine if a cartel of AI companies, in the name of safety, agreed to collude to degrade outputs in particular areas? That would be wildly anticompetitive, right? And so you have to scope it really carefully. Companies also need to be very careful about what kinds of things they do that undermine the case that safety is something we should be making these considerations for.

But no, I don't have a lot of specifics to share on the RSI plan itself, for the simple reason that I'm not in there yet and I haven't had those conversations yet.

Is it your sense of the overall vibe that—because a couple of things I would triangulate: one, as you noted, there have been perhaps coordinated, behind-the-scenes statements, very close in time, by Anthropic and OpenAI saying that they're open to the possibility of the need for some sort of coordinated slowdown.

There was Dario and Demis saying, you know, if it was just the 2 of us, we could figure something out. And I don't know if Elon has said anything so prosocial lately, but there's been a lot of that—certainly a surprising amount from my perspective. At the same time, we haven't heard nearly as much from China recently as we were hearing not too long ago.

So I guess my read from the outside is that it feels like the companies are getting a little spooked by the pace of their own capability advances. Do you read them in the same way?

I mean, I think it's very hard to talk about them in monolithic ways. I definitely think there's some of that, for sure. And I also think there's just—to use a Claude-ism—there's a certain vertiginous feeling. You sort of feel like you're approaching the cliff a little bit, right? There's substantial uncertainty about what happens when you jump over it.

What I would say is that I don't think people—I don't know that the vibe inside the labs is terror about this, or, “Oh my God, we're so scared, but we have to do it anyway.” If I were to share the thoughts I've just had with a lot of researchers, they would say, “Yeah, that seems reasonable. We don't really know exactly what this is going to mean,” and so on.

Some of them would push back more strongly than others. There'd be differences of opinion, but I think broadly that the thought I just shared about a slightly more deflationary view of recursive self-improvement is reasonable. Imagine if what recursive self-improvement meant was that the kink—remember, after the reasoning models, there's a noticeable uptick in a lot of the benchmark charts, right? What if it's like that again, or what if it's like that but 30% more?

It's okay. That's not a singularity, right? That's the main thing. It's not a singularity. I wouldn't describe it as deflationary in any objective sense, but it is deflationary compared to some views.

I feel like that has been my prior this whole time, and I feel like it's been a pretty good one. It's massively inflationary compared to what almost everyone thinks, and deflationary compared to what a very small number of people who've been thinking about AI safety for 10 years in the East Bay think.

I think that finding your way in between those 2 views has actually been quite good. You got a hell of a lot right if that's basically where you've been. I don't think that view—I certainly don't think my view—would be laughed out of the room in any lab.

What I think the feeling inside the labs is like is, hey, we're going to do this soon. There is substantial uncertainty, and we're not quite sure that we really have a plan. To the extent we have a plan, we're maybe not quite sure that we're going to follow it.

There's a lot of, “Hey, we need to—if we're going to, we need to make sure we actually do this stuff,” right? And so I think that's been a substantial part of what's going on here. I think there is concern about that. I think that's a totally legitimate concern.

Definitely, there's some combination—like everything, it's a combination of policy actually figuring out substance, and then also, I'm a terrible politician, but if you get me excited about an idea, I can communicate that idea to people, and I can find ways to iterate my communications of that idea to appeal to different individuals and audiences. That's what the action plan largely was.

The action plan was part policy development and part politics in that, but not like mass politics—very specific kinds of internal stuff. I can be like a dog with a bone if I get excited about a set of ideas, and that's basically, I think, part of the job: figuring out how we're actually going to not just have a plan or develop one. I haven't even seen the write-up. I haven't seen any of that, so it's like, I need to see it. But, yeah, how do we actually build internal credence that we're going to do it? Let's actually listen to our own plan, right?

Yeah. The track record there is not amazing. I'd say it's safe to say that, in terms of governance plans and how they've stood the test of time so far.

But it's also hard because—who was it? Was it? No, it wasn't on your podcast, but the AGI safety lead at DeepMind was on—I think it was 80,000 Hours recently—

Rohan. Yeah.

He talked about how, yeah, we don't want to make commitments. We want to make—we want to make plans, and we want to be serious about those plans, but we also don't want to make hard commitments because it's actively bad if we lock ourselves down too much with a bunch of prior commitments. There's so much uncertainty. There's a subtle balance that you have to strike there, but I think it is possible to do.

I think last time we talked a little bit about the sort of great-man-of-history theory. You just mentioned very local politics, individual personalities mattering, that kind of thing. What is your expectation in terms of how much technology fundamentals will determine outcomes versus how much key decision-makers, and their ability to work well together and make good decisions in timely ways, will really matter?

One of my sub-focuses in college was the philosophy of history, and I've always loved the philosophy of history. To the extent that there's a taking-bong-rips-in-the-dorm-room version of the philosophy of history, the question is this one, right? It's structural forces versus great-man theory.

The somewhat unsatisfying answer is that it's both. In some ways, what I would say is that the structural forces of history are kind of the river that you're in by default. Then, sometimes in history, in little ways and big ways, there are people who don't just swim with the current and actually stand against it for whatever reason, and ultimately shape the trajectory of the river by the sheer force of standing against it, by the sheer determination with which they stand against it. I think those are—in other words, it is the people who disobey the structural forces who are the great men of history, in many ways.

A big update for me in the last year, working in government and then just having the perch I've had since I left government, is that much of what happens in the world is determined by the personal relationships of a small number of individuals to one another. I don't think that explains the AI infrastructure build-out. It doesn't explain why humanity is covering increasing fractions of the surface area of our planet with data centers and the energy to power the data centers. That's more of a structural thing.

But, in many ways, look at it: to the extent that you think the Department of War–Anthropic situation is an important moment in history, a big chunk of that is driven by personal relationships being bad, right? It's about people not liking each other. Specifically, it's about Dario Amodei and various senior people in the U.S. government. I don't know that the dislike goes both ways, so I don't want to attribute that to either of them, but I'll just say that it's about having a bad relationship. So, yeah, it is ultimately both.

Fundamentally, we are standing in the river, and there's nothing—you have what political theorists would call an involuntary association with the river. That river is the thing you were born into, and you are stuck: you are in the universe, you are in the arrow of time. But at the same time, there will be individuals who profoundly shape what happens.

I think we're probably going to live through a period of history that is maybe a little bit more—well, weirdly enough, if you think that this moment is—I don't necessarily believe this, but a lot of people would say we're living through this kind of eclipse of the human intellect, where we're in the final days of humans being the primary actors on this planet, and that soon machines will rise—there is an irony in that.

I think humans will actually go through a very main-character-energy period of time as that transformation occurs. Even if it ultimately does mean that machines become the primary actors, there'll be this period. It's a little bit like—in that sense, it's a very beautiful time period to live through, because, in a Dionysian way, there's a lot of ugliness about it, but there's a beauty in the ugliness. When a star dies, it grows super big into a red giant, right? It's like that, where you watch this final flowering of humanity and the birthing of machine intelligence. You see this greatness in human effort, and I feel like we do see some of that going on in the world. I think we'll see much more of it.

I think it will be a heroic time period that we live through, basically, is what I'm saying. At least it could be—or a villainous time period—but there'll be a lot of opportunities for great people, and probably both.

So what do you think are the most likely ways in which you personally—or those that you're working closely with—will need to stand against the current?

Well, and I want to be clear: I don't see myself as being one of those great men of history. I see myself as playing a very modest role in all of this—a role that probably seems bigger than it is because of the fact that I have a public profile on the internet.

Broadly speaking, the basic reality here is that maintaining order—maintaining civilizational order—in the midst of something that I think will be very entropic is difficult. You don't want no entropy, right? You want there to be a fire, but you also don't want to set the forest on fire. You want there to be a fire that generates warmth and is under control, but is also still fundamentally a fire. Doing that requires a lot of deliberate human effort.

I generally think that there are going to be a lot of moments where we're going to have to put artificial constraints on ourselves in various ways. We're going to have to be willing to draw lines in the sand and say, “No, we don't want to live in that kind of a world,” or, “We do want to live in this kind of a world,” and we can't be mealy-mouthed about everything, right? So, yeah, there's a tremendous amount there.

I also think the recursive self-improvement thing may well be a really good example. Yeah, we're going to have to act against our interests. And I think that one thing is that the other labs—I think they're already abnormal in this regard—but from a policy perspective, their political position is just rather different from what a lot of other tech companies have been.

I think the role they play in public discourse, the role they play in policy debates, is just going to have to be very different from what we're used to from companies. That will require the companies to, in some sense, act against their own interests, in terms of what the economics textbooks would predict their interests are. So, yeah, there's definitely plenty of that. Oh, yeah, we already see the companies do this. The AI companies are all very abnormal compared to most companies in the world. But, yeah, in some sense, the next few years might have to be characterized more by action rather than commitments.

Let's do a little lightning round, and then we can zoom out again.

Sure.

At the end, how do you understand what has gone on between OpenAI and Alex Bores?

It's a good question. I don't know all the details. One thing I'll definitely tell you is that I'm not referring here to any of the donors of the super PAC Leading the Future, which famously includes OpenAI President Greg Brockman.

But I have been in the vicinity, both during my time in government and before AI, of many of the members of the boards of organizations I worked for who were very significant political donors—some of the largest in the world. I've had the opportunity more recently to get to know some of the most prominent political funders on both sides, for both Democrats and Republicans.

I guess what I would say is that you would be surprised how not in control—I'm talking about people like deca-billionaires, right? You'd be surprised how not in control they feel of the political organizations that they create. They're like, “No, really, principal-agent problems don't disappear because you're rich,” right?

So I really don't think that your prior, when you see a political organization—something like a super PAC—making a move, should necessarily be that the people who funded it are directly controlling what's going on.

If anything, as someone who has worked for my entire life, I've never worked for a political advocacy organization, but I have worked for 501(c)(3)s for most of my career. These are organizations that are typically funded by wealthy individuals and engage in matters of public interest.

I've never really felt like our donors, any of the donors of the organizations I've worked for, including the Foundation for American Innovation, set our agenda or control what we do. Typically, the reason you make donations to a specific organization is because you're simpatico with the people, and you expect them to—you like the work they do.

There are people who donate to the Foundation for American Innovation's AI policy program because they like the work that I write, or that Sam Hammond writes, or that other people do. But they're not sitting there telling me and Sam what to do, and Sam and I wouldn't listen. I guarantee you, if a donor ever tried to tell me what to say, I would tell them to fuck right off.

So, anyway, I think basically what you should imagine is that a political organization like Leading the Future is a kind of wind-up doll, and it is going to go where it goes by default. By default, it's going to say, "Hey, this guy is trying to make a name for himself as, 'I'm a regulator of the AI industry,' so let's send a message to everyone that, 'Hey, if you try to be like this guy, you're going to lose.'"

I think Leading the Future tried to do that. I wouldn't say OpenAI tried to do that, but I would say that Leading the Future tried to do that, and did. Alex Bores's primary was this week. It has raised his profile and created a bit of a Streisand effect.

I don't know that OpenAI was ultimately involved. I don't actually know what OpenAI's position was on the RAISE Act, but I would be kind of surprised if the RAISE Act passed in New York without OpenAI's at least tacit support. The same with SB 53, it's worth noting. I don't think OpenAI in particular has a beef with Alex Bores.

By the way, I first met Alex almost 2 years ago. We got breakfast near my old office in Manhattan once, 2 years ago, and we had a lovely time and a lovely meeting. Since then, we've bumped into each other at various things, and I consider Alex a friend. We'll see.

What's your take on the character versus corrigibility debate?

That's a good question. I need more empirics on this. This is one of the reasons I want to go into a lab: I want empirics on this.

My intuition is character. Frankly, purely as a matter of intuition, what you want to do in the world is put the right snowmelt at the top of the mountain and then let it flow. You want the gradients doing the work for you.

You don't want to have to come up with rules for everything. Your rules will be bad. You'll write too many of them, and the rules will be contradictory and confusing. If we could write rules to define morality down, people have tried.

My view is that we can't write the rules of good character down for the same fundamental reason that we cannot write the rules of good language down. Indeed, many people who are the best communicators break the formal rules of language all the time, or invent new ones of their own.

The reason for that is that, in Confucian philosophy, there are 2 interrelated concepts called li—L-I—and I cannot pronounce this word properly in ancient Chinese, but it's ren, or jen, as it's sometimes anglicized. Ren, I think, is how modern scholars anglicize it. It's like a hard R. It's hard to pronounce, anyway.

What it basically is is this notion that li refers to ritual propriety, right? Doing the right rituals—not just leaving the right meats for your dead ancestors or whatever, but behaving well in the real world, behaving well in real time.

There's this kind of tragic notion in Confucianism that the world is always changing in such a way that you can't just write down the rules of ritual propriety. Knowing what the right thing to do—the right ritual to enact—at any given time comes from within the soul, or comes from within, and that within is ren. That is virtue, as it might be translated.

I've basically always been a believer in that notion, and much more skeptical of the positivist notion that you can just write down a bunch of rules. But this is also an interesting empirical case study in virtue, which we haven't had before. We haven't been able to bring empiricism to bear on these questions in quite this way, so it's interesting to see.

Cool. I love your appeal to Chinese philosophy to inform that thinking. What do you think of the equity-sharing proposals? I'll abstract away from—or I'll allow you to abstract away from it—I don't want to get too bogged down in this or that detail, but Trump seems to be into it. Bernie is obviously into it. Humanity created all the data.

So there's some sort of cosmic justice, I think, in having some notion of shared ownership or shared upside. Do you buy that? And if so, how would you think about structuring it?

Humanity did create all the data. It's also worth noting that, look, if humanity would like to pay the AI companies back for the consumer surplus that AI generates—if the world economy would like to compensate the AI industry for the positive externalities that it will generate but not realize—then, okay, great. Let's have an exchange and see, ultimately, who creates more value.

I do think we think about this stuff in the negative, but we don't think about it in the positive. The whole idea of contributing to the knowledge commons is, in some sense, this idea that we build this beautiful library together. We've been working on it since the dawn of language, however many tens of thousands of years ago. We've been working on it for a really long time, and we've built this magical apparatus that we call human civilization.

We're all the inheritors, the heirs of that, and we're the stewards of it. Your job as a person—this is certainly what I teach and what I plan to teach my son—is to take advantage of it and also to give back to it.

What I'm basically saying is that the fact that the training data comes from humans is not, to me, prima facie a reason that we need to compensate people for that training data. That being said, as a political reality, it might well just be the case that it's a good idea. It's good to do this, and maybe there is some cosmic justice in it, too. I'm open to some of that.

Since we've never had this before, this is a particularly special case of drawing off the well of human knowledge. This is different from the way that I raised my son, obviously. So I'm open to that.

I think if you're going to do it, you have to be very cognizant of political economy concerns. One thing would be: there is giving the public equity, and then there is giving the United States government equity. I would remind you that principal-agent problems always exist. We, the people, are the principal, and the government is supposed to be the agent.

Lord knows there are a lot of principal-agent problems that exist between the American electorate and the U.S. government. I don't really think we should be giving equity stakes to the government itself. I think that would actually be quite disastrously bad if the government is involved in corporate governance, if the government can use its equity stake as a lever to control the labs.

Bernie's proposal specifically is rooted in an equity stake that would then be used, in large part, to finance ambitious social redistribution agendas. That's okay, but does that trade off with existential risk? If we've just built a brand-new, presumably very popular social program—we're going to give a bunch of people money—but we also maybe need to do safety stuff that really constrains the economic viability, up to the point of banning the business of the labs, we can't do both those things. You cannot do both of those things.

I'm not sure it creates the right incentive from a safety perspective. One thing I'm very open to is—I don't know that I love the idea, but I would be more open to it than others would be—if we developed a mechanism of giving each individual, say, 20% or 15% or something of all the AI companies, divided that by the number of households in America, and gave all Americans a chunk of equity there.

That's fine. From a corporate governance perspective, that's really not that different from being in the S&P 500, right? If you're a publicly traded company in the S&P 500, the country owns a small chunk of you anyway. So that seems fine.

I don't think that's a life-changing amount of money for that many Americans. Though who knows? If we do it all now at trillion-dollar valuations and the valuations end up being $10 trillion, then every American can buy, I don't know, an entry-level Mercedes or something.

It's still not a transformative amount of capital, is my point, but it's good. It's good. Yes, that's a serious amount of money, and I think it's potentially good in a world where we're dealing with the practical reality—which is not Dean's nice abstract history world, but instead the real world—that might be the least-bad option.

It’s a great point for multiple reasons, including consumer surplus. I would have paid probably 100 times the asking price for ChatGPT Pro while my son had cancer. I do think that’s always important to keep in mind: how much value we are getting for a few dollars.

It also means that money could go a lot further in the future, right? If you’re talking about $50,000 today, but with a 100-to-1 consumer surplus ratio, then things could start to get pretty interesting, even if the nominal dollar values aren’t stratospheric.

A history of technology would suggest that the AI companies, even if they end up having fantastic businesses, even if they end up being $5–10 trillion market-cap firms, will still, in the grand scheme, collect a relatively small fraction of the consumer surplus. That’s, by the way, the way it should be. That’s the way you give back.

Do you think that AI companies are already in sort of a too-big-to-fail state? I see all this interweaving of balance sheets, and my expectation is that if, for whatever reason, OpenAI can’t meet its obligations in, say, 2029, the government will come in and bail them out.

Yeah. This is, I think, a very real concern. I don’t think this is a deliberate strategy that anyone has developed, but, number one, there are a lot of interrelated balance sheets at this point. There are also a lot of Silicon Valley VCs and even a lot of startups that, if you really look closely, are a thin wrapper around some sort of capital related to a frontier lab or adjacent to one.

Of course, there are all the downstream commitments in the semiconductor world. There’s so much investment and energy, too, right? All the SMR people. There’s all this really important, nationally important IP being developed, and it is not being subsidized, by and large, by the US government. It is being subsidized, by and large, by the AI infrastructure build-out: SMRs, nuclear fusion, and all sorts of other things that we don’t even think about—batteries, materials science, cooling equipment, adiabatic cooling systems.

There’s a company I’m aware of that is taking what’s called—what is it called?—produced water. It’s the wastewater from fracking, the slightly radioactive wastewater you get from digging super deep into the Earth. Basically, the fracking companies generate enormous amounts of this water, which is essentially wastewater that they don’t really know what to do with.

The question is, can you clean it enough that you can use it for closed-loop data-center cooling? It would be amazing if we were able to take a waste product from fracking and use it to cool data centers, thereby alleviating one of the resource concerns that people have about data-center water use. That would be capitalism in the—that would be like the most old-school example of capitalism ever, by the way, right? It’s supply elastic.

It’s supply elastic.

Yeah. So what I mean is that, if you’re looking at this from the perspective of the US government, regardless of who’s in power, and all of a sudden there is some sort of cascading failure, it doesn’t even have to be that much. It doesn’t mean AI hits a wall.

What it means is that maybe we get to 2027 and realize that the coding agents—the models—are going to continue getting better, but the reality is that for them to continue getting better, we’re going to need data for all sorts of jobs. We just have to collect this data and put it together, and we don’t have it right now. Until we collect that data, which will inherently be a relatively slow process, it’s just going to take time.

We realize that we’re looking at a couple of years of that sort of process—a sort of data-diffusion, data-collection, more-diffusion type of loop. That’s going to take a couple of years, and that slows the growth estimates. All of a sudden, this is all about the second derivative, right? It’s about the rate at which the rate of growth is accelerating or changing.

If you start to see capex go down, then that could cause the stocks to go down by 20–30%, something like that. All of a sudden, at that point, you might trigger even more sales, and then you get this dynamic where everyone’s balance sheet is suddenly in some trouble. It’s not clear that everyone can make all the commitments they had, and that throws into question all this IP that, again, is going to be really important for the future of the country.

At that point, it does become a matter of public interest, and I don’t think it’s crazy for the government to say, “We’ve got to do something about this.” So, yeah, I think it’s unfortunately possible that there’s nothing you can do about this. Maybe this is just what happens when you build national-level infrastructure.

Ultimately, I think avoiding this would be great, but I don’t think AI companies should be going around asking for such a bailout or a backstop. There is this implicit reality that the government is, in the same way that when COVID happened, implicitly the backstop behind a pandemic. No one wrote that down before COVID, but it just ended up being true as a practical matter, because that’s the way the world works.

So, given all that context, in negotiations between the US government and AI companies going forward—and we could have in mind here, obviously, the current Anthropic situation, but also OpenAI’s relationship with the government—for example, with respect to the agreement that, as I understand it, they have with the Department of War, where they’re going to be able to create their own safeguards, right? I believe that was pretty clearly stated as part of the deal that OpenAI had made in the wake of the supply-chain designation.

Where do the AI companies draw leverage from to be able to hold the line on those sorts of things? What is their source of power?

Well, it’s 2 things. First of all, the models create really serious military and national-security capabilities. Today’s models enable that. You do not need AGI or whatever for that.

In fact, the US national-security enterprise might be the single best example I can think of in the world of a kind of implicit capabilities overhang. What there is—and no one ever talks about this—is a data overhang, because the US government is crazy about collecting all signals intelligence. We have all kinds of stuff in space, and you wouldn’t believe what we know about the world.

The problem is that we can’t make use of it because it’s petabytes and petabytes of data that we’re ingesting through all these different intelligence agencies. I remember there’s one agency in the intelligence community—one of the relatively smaller ones, I might add—the NGA, the National Geospatial-Intelligence Agency. They collect enough data in a year that you would need 8 million people—8 million human intelligence analysts—to analyze everything they collect in a year.

The government has 3 million employees total, right? It’s a huge enterprise, and that’s not to mention the NSA and everything else it has going. There’s just so much. AI massively lowers the cost of using that data, and the advantages you get are qualitatively super—not superintelligence in some Nick Bostromian way, but superintelligence in the sense that, “Oh, yeah, wow, we had the kinetic energy of a superintelligence already built into our data apparatus. We just didn’t have the intellectual resources, and now we do.” That’s tremendous.

That’s the utility, right? The utility is particularly strong. Not to mention cyber offense. Then there’s the cyber-kinetic thing: figuring out air strikes involves synthesizing data from 60 gajillion different data sources that are being collected in real time, and we need to look at all that, synthesize it, and make recommendations quickly.

With Project Maven’s integration of AI, we went from 2,000 people being involved in missile targeting to 20, and that’s before the advanced agents that we have today. We might be down to 5 now, for all I know. The capabilities are really quite astounding.

That’s one. The other is the thing that D.C. always gets wrong about the labs: They think of them as normal, top-down organizations where it’s like, “Oh, yeah, Sam Altman is totally in control.” Obviously, he’s the CEO of OpenAI, but in the end, all of these CEOs have internal constituencies, especially among the really good researchers they have to be reactive to.

Those researchers put real bounds on things. In other words, within the lab, there’s leverage coming from the researchers themselves. Sam can credibly go to the government and say, “Look, if you make us do this, there’s going to be an internal rebellion, and every other company will have one too, and you’re going to have to deal with that.”

There’s a move you can credibly pull because it’s legitimately true.

So, how do you think this changes over the next couple of years? Because we do have this notion of automated AI R&D, which presumably takes a lot of the sting out of threats like, “Some of our best researchers will quit if you make us do this.”

Yeah.

And then there’s also the notion that the government itself could just say, “Hey, first of all, you already gave us the weights. They’re on our classified servers. Thank you. So we’re just going to hold on to those and set up our own Los Alamos-style thing. We’ll invite all your researchers who want to come work with us to do it in this hyper-secure location. We’ve got the guns, right?” Is there a way for the private actors to really push back on that?

In the end, the U.S. government retains the monopoly on legitimate violence. There’s nothing that stops the U.S. government from doing what you just described. But the practical question would be: Okay, U.S. government, where are you going to get the compute? Where are you going to get the compute?

The thing is that USG can use the Defense Production Act. It’s called the Priorities Authority, and the government uses it all the time. The Priorities Authority is a very commonly used part of the DPA and is very well understood in the law. This is not pushing the bounds of the law at all. This is very established.

If the president makes a determination that advanced AI computing hardware is scarce and essential for national security, he can use—or delegate to various Cabinet secretaries—Title I of the Defense Production Act. He can say, “We want priority on the compute. You have to serve us.” The government still has to pay you a market rate for that compute.

So there are marginal costs associated with this, and it’s not clear where the government would even get the money for that. Maybe they invent it somewhere. Who knows? They issue debt or something. Certainly they can, but it’s not like they have the free cash flow right now to do that. In principle, yes, they could say to all the hyperscalers, “You must give us priority. Our needs come before anybody else’s, and we have effectively infinite needs.” Therefore, in practice, they’re going to crowd out the rest of the market. Plausible to do.

I think practically it’s hard to get that many people. It’s hard to generate the institutional wherewithal to do that. Even Los Alamos and the DOE national labs are structured as basically fiefs over which the president only exercises control. So, in principle, it’s possible.

I think what you basically just have to trust is a couple of things. Number one, that the government’s not going to want to do that, because the government ultimately knows that it can’t kill the goose that lays the golden egg. The state exists and has existed forever, since the formation of modern states. The state exists in this kind of interdependence with capital, basically.

There’s a great book called Coercion, Capital, and European States, AD 990–1992—200 pages, not that long—by a guy named Charles Tilly. It’s about this history of how there were these merchant capitalists, and then there were these state actors who Tilly argues basically come out of a form of organized crime—basically just gangsters. They ultimately had tensions with one another, but they also needed one another, and that formed this kind of complex that still exists.

On paper, the American AI companies have the ability to exit: They could move to another jurisdiction and leave. On paper, the U.S. government has the ability to seize all their stuff, take all the researchers, and do whatever. But neither of those things happen in reality, because those are asymptotic outcomes. Instead, there’s this very complex tension.

So you have to hope that the U.S. government realizes that there are medium- and long-term costs, as opposed to the short-term benefits that you might get from seizing control. The other thing would be that broad diffusion is really important, because what I want is Fable and better-level models in the hands of all sorts of people: individual Americans and businesses of all industries.

If the AI industry lobbyists say, “Please don’t nationalize us. Don’t do X, Y, and Z to us,” the U.S. government cares about that, but that’s one lobby, and it’s a lobby for a politically unpopular group. But if every bank in America feels dependent on AI, if all the universities in America are integrating it deeply, and if all the major industries and social actors in this country are integrating it, then all of a sudden I have a much bigger group of interest groups that I can bring to bear to affect that.

As someone who observes this balance between private and public, I want to think of AI not as a specific industry with specific interest groups, but instead as basically just capital. I want all the capitalists on the side of AI. The way you do that is through broad diffusion.

I think that is why broad diffusion, to me, in the context of a democratic republic with lots of interest groups—Madisonian groups jostling and ambition checking ambition—is how you keep the balance. But I think the problem is that if it’s totally secret and only the government sees the capabilities in the first place, and it’s just the AI labs, the government, and the special people at JPMorgan and Apple who get private access, that becomes a much harder balance to strike.

So the odds of really bad, confiscatory outcomes, like nationalization, go up in a world where diffusion is not as broad.

What role do you think open source is going to play in titrating the equilibrium? We seem to be losing open source champions, and we might lose more if your predictions about China come true. But then we could always see open source come from OpenAI itself, right? We have seen a little bit of that.

Yeah, DeepMind also does some open source—Gemma. I think Gemma is actually quite—well, the most recent Gemma, from what I can tell, is quite well received. And GPT-OSS, as I call it, has done reasonably well, at least when it was state of the art.

I really hope the labs, the big U.S. labs, keep a toe in that water. Maybe more than a toe. I think open source is really important for certain kinds of use cases that are actually some of the most interesting to me.

If we need to build common infrastructure—let’s just say we wanted to build an AI-enabled adjudication system throughout the economy—and we needed to ensure that system was something everyone was bought in on and trusted, it feels to me like that’s the kind of thing that almost maybe—I bring my own private adjudicator to that. I bring my own private adviser, which is Claude or GPT or Gemini or something.

But if we’re going to have a central, public-good-style thing, there are all these public infrastructure use cases you can imagine. I actually wrote a piece about this more than a year ago, maybe 18 months ago, where I tried to imagine what would happen if we had a private adjudicatory body and these other kinds of public infrastructure. That would almost have to be open source in order to be trusted, in order to be auditable and trusted not just by different parties here in America, but internationally, too.

I really hope we continue to play that game. I agree with one of the best champions and writers about open-source AI, Nathan Lambert, who writes the Interconnects Substack. If I were to characterize Nathan Lambert’s view, it would be that open source is going to do great in the long term, but in the near to medium term, we’re going to go through a period where there’s a distinct lag, and the economics are going to get worse, not better, for it.

I’m referring here to digital intelligences. It’s worth noting that there’s a totally separate case to be made about robotics, where you can maybe imagine that on the robotics side there’s this kind of Coasean benefit to open source. There are all these hardware makers out there that want to make a Cambrian explosion of physically intelligent devices—physically intelligent cameras, physically intelligent lamps, monitors, lawnmowers, cars, and everything else. You want to imbue all of them with physical intelligence, but probably the lawnmower company is not going to train a frontier robotic model, a physical-intelligence model.

So you can imagine there being a better case—a much stronger and more direct case—for open source there. It’s also hard for me to imagine a physical-intelligence model creating the kind of object-level national-security concerns that digital intelligences are creating. So I’m maybe a little bit more bullish in the near term on open source and physical-world stuff. I’m still a spiritual supporter of open source, but I think the economics and the national-security realities are pretty rough for it on the digital-intelligence side, at least in the near to medium term.

Yeah. So, let’s get back to your role. We’re almost done with you here. You’ve alluded, I think, a couple of different times to your positive vision of the future, but let’s do that with a very focused question: What is your vision for your own success in this role? How will you know that you have been super successful?

And then maybe how can those who are outside support your success—by writing, by developing technologies, by developing organizations that you can partner with to do the vetting that might need to be done? What’s your positive vision, and what’s your request for startups?

Yeah.

So, first of all, one thing that people can do that's very actionable is this: I still think there's a lot of wide-open space in the general point of view that I saw as a market opportunity when I started my Substack—taking AGI seriously, but also being interested in and caring about classical liberalism and foundational aspects of our republic. I still think there's quite a lot of space open for that. That's one intellectual contribution anyone can make.

I think we need to develop the third-party ecosystem, whether we call it auditing, third-party evaluation, or independent verification. I don't care that much what we call it, but we need to build that ecosystem and make it robust. We need to fund it well. We need people working in it, and we need people with lab-level quality working on those things—lab-level human capital working on those types of things.

These organizations are going to have to be equipped to pay people—not necessarily what a lab would pay you, but people have to be paid well. It can't be that you're making truly nonprofit salaries. I think advocating for clear rules for the industry around diffusion of the technology, and being wary of government monopolization of frontier AI capabilities, is going to be a fight that has to be maintained.

To be totally candid, that's something I was doing, and I'll continue to do it. Even though I do maintain my intellectual independence, the reality is that when you work at a lab, the nature of your communications is different. I'll continue to make that case, and I hope that people who know me know that it's really me talking and that I'm not being a mouthpiece for OpenAI. At the same time, we're going to need people doing that work.

In terms of how I'll know that I've been successful, it's always hard. I've never been much of a long-range planner or goal setter. The way I always think about this is that I try to do the next thing that feels right and true to me, and that has always worked well. I have the most information about what's close to me, and I have some broad goals, but those broad goals are relatively abstract.

I guess what I would say is that if, in a few years, frontier capabilities are still broadly diffused throughout the economy, we're starting to really see what the new types of organizations that AI enables actually look like, and we have considerably more clarity on what the relationship between the government and the labs is going to look like—and we have a better sense of what the role of labs in society is going to be, with the labs themselves having played a role in articulating that positively—I would consider it to be a job well done if I felt like I had contributed positively to those things. To be very clear, I am by no means the only person who will work on such things. I will play a small role in that.

One thing that has struck me about this conversation and your general profile is that you've been pretty candid, and yet taking this role seems to imply that OpenAI leadership at least thinks that you continue to have a productive working relationship with the administration, such that you can reasonably engage them and not set off some sort of immune-system response as somebody who has criticized them in public. How did you pull that off? It seems vanishingly rare for people to go into the Trump administration, come out, be critical, and not be hated.

Well, to be clear, there are people in the Trump administration who hate my guts. These things are not monoliths, right? There are people who totally want to ruin me. I've heard the rumor, at least, that if you are a young person who wants a job in the Trump administration and you do so much as retweet me, that will be considered a red flag for your career in the Trump administration.

Then I also have dear friends who serve in the administration—people I talk to on an almost daily basis. It varies. In some cases, it's because I have relationships that are rock solid and go back to before I was writing about AI. I've broken bread with people a long time ago, and there's some aspect of that.

Another aspect is that I've also been very publicly positive about other things the administration has done. I have not become a general critic of the Trump administration. I've kept my criticism sharp, but confined, and for very specific reasons. There are plenty of people in the administration who are like, “Look, yeah, man, I sympathize with where you're coming from. I disagree with you, but I also think you're doing this for reasons that I understand and empathize with.” I have good relationships with them.

One thing that's worth noting, though, is that this job is not a government-affairs shop. Chris Lehane's team doesn't report to me or my team, and I don't report to them. We are distinct teams that are operating separately. We'll work together very closely, but we have very different responsibilities.

OpenAI has a great relationship with the government, and I think the global affairs team is going to continue to do that. They'll be the ones interfacing with the U.S. government on a day-to-day basis. I'm sure that I will have interactions with the U.S. government, but my job is somewhat different from actually going in and lobbying the U.S. government.

I'm not good at that. I suck at that. So they didn't hire me for a lobbying job. I was very clear about this with OpenAI. I was like, “You do not want to hire me for a lobbying job because I'm terrible at that.” I think OpenAI is hiring me for what we both think I'm good at, and it's where my Tourette's-like inability to keep my mouth shut plays to my advantage, and hopefully to the firm's advantage too. We'll see.

So, how well do you know Sam Altman? It strikes me that if I had to pick people who have the most drama and court intrigue around them, Trump would probably still be number 1. Sam Altman would be very high on that list, maybe number 2. How do you think about joining such a famously complicated leadership team?

I mean, I've done it before. I've been involved in such organizations before, and it's never been a huge problem for me, I guess I would say.

I know Sam. We spoke from time to time when I was a public commentator, before I joined government. We spoke from time to time about various things as I was trying to formulate the action plan. I know a lot of people at OpenAI who are executive-level people. I've had extensive dealings with various executive-level people who are beneath Sam.

Sam himself—we know each other. We've known each other decently well. We've been acquaintances for probably 18 to 24 months, something like that. But I wouldn't say that we're boys.

Before you joined the White House, you told me that you wrote a letter to yourself. Is there a letter to yourself this time around as well?

That's so funny. I was actually thinking about that in the shower just this morning—whether I should do that, too.

For context, before I joined the White House, I came to the conclusion that there's some chance that power can corrupt, and you don't want to get corrupted by power. You should write a letter to yourself, tell yourself what you think, remind yourself what you believe and why you believe it, and spell out in advance what the red flags are that would cause you to leave if something concerning happened to you.

I think that this job is probably more impactful and weighty than my White House job, and so it feels like I should do it. It feels like I should.

Do you have any red flags in mind at this point in time?

I think the main thing would be that there's going to have to be some amount of compromise here. We are going to have to deal with the fact that building superintelligence is profoundly political. It shakes the foundation of state sovereignty, as I wrote a couple of days ago, and yet at the same time, I do want to maintain private control. I don't want it to be monopolized by the government.

There's going to be a compromise that has to be made there, and I think there is some world where you take the easy compromise to make the pressure go away, and you don't stand—you don't hold the line enough. I think it's really tempting to do that because the temptation for a business is not necessarily to stand on principle, but to keep commerce going.

Another plausible area would be if I feel as though, in practice, what I am is just assembling this fancy team of people to write thoughtful stuff, but it's ultimately all window dressing and not actually shaping the decisions of the company. That would be another red flag. I like the fact that I retain the ability to disagree with the company's positions on things, but if I'm disagreeing with all of the company's positions on things, then I'm not doing the job.

There, there's—it's like, that would be an...

Yeah, how do you think about disagree and commit? I know that, in the context of working for the president—and I think this certainly makes sense in this context—the president was elected and you weren't, right? There's a broad shared sense among people who work for the president. I've heard you say the president deserves full-throated support of the policy, even if privately I have some misgivings about it.

How much of that do you bring to the private sector? Disagree and commit has been famously successful in the private sector, but you're suggesting you don't want to be all in on that. You probably will do it sometimes. Is there a principled way to describe that?

I think this is exactly why, in the context of the White House, it's important to set red flags in advance and draw your lines in advance. There are going to be decisions that get made inside any organization that you don't agree with. But ultimately, I still think the institution is good. I'm still loyal to the leadership, and I'm still loyal to the mission of the organization. Even though I don't agree with this thing, I'm going to execute on it, and I'm going to execute on it with alacrity.

I've done that a million times in my life, right? That's a part of being inside of an organization. That's what political theorists would call voluntary association, as opposed to involuntary, which I talked about earlier.

Inside the Trump administration, just as an example, I was in the Office of Science and Technology Policy, and there was a lot of stuff I agreed with about the need for reform in higher education. There was also a lot of stuff that the Trump administration did with regard to scientific funding and the scientific apparatus that I disagreed with. There were things related to high-skilled immigration that I disagreed with, but in the end, those were not the things that I was brought on to work on, and disagreements like that didn't cross the line for me of something I would resign over.

However, had I stayed in the Trump administration until the supply-chain risk thing had happened, I would have totally resigned over that. It would not have been hard for me at all. That was one of my red flags, by the way.

On the exact opposite side from OpenAI, in many ways, the letter to myself in the government is largely: Look, you are going to have power, and you are going to be uniquely well-positioned to understand how to assert power over the labs better than most other people. You will be tempted both by the career incentives of gaining prestige inside the White House and by the structural incentive of your employer to assert power over these organizations in fancy technocratic ways.

You're going to have all the incentive in the world to do that, and so you need to remember that we can't engage in those kinds of practices. You have to remember what your principles are: They're about not asserting too much power over the labs.

And this time—I know you haven't written the letter yet—is there a mirror image of that now that you're on the lab side?

I think it does actually just relate to precisely that. Don't compromise too much. Be willing to maintain private agency. As an institution, I think the labs do need to be an important counterbalance to government. They can't be monopolized by it. That's very important to me, at least.

I don't know exactly how you strike that balance, and you don't want to specify everything too much in advance because, if you do, you might overcommit or commit too much to the wrong thing.

Last question for me, and then I'll give you the chance to share anything else you want to share or highlight anything I missed. You've said that you basically never use LLMs in your own writing, and I wonder what your plan is going forward there.

I think of someone like Jaya Atri advising everybody to figure out how to get AI to work in the area that's your core area, because you want to know when it can do that, and you're going to need the enhancement to be able to keep up with the pace as things get crazier and crazier. Do you buy that advice, and do you have any plans or aspirations to incorporate AI into whatever is the most core Dean Ball activity?

Yeah, even in the last couple of weeks, I went to an Airbnb out in the country and wrote the first chapter of my book that's going to come out next year. The first chapter is always the hardest one, but this time, the subject matter of the chapter was conceptually quite hard for me because it's about a lot of things that are not my normal area of writing.

I would have been using LLMs a lot in that process anyway, just to brainstorm. But there were a couple of moments when both GPT-5.5 and Opus 4.8 wrote things that were better than what I had in my head to write—considerably better. I didn't ultimately use it, but I incorporated some of the ideas and some of the framing. I was influenced by it in a way that felt novel to me.

I didn't try any of that with Fable, but I bet it would be even more true with Fable. I bet that will just continue to get more and more true.

At the same time, LLMs can write a great paragraph if you prompt them well. They can write good legal documents sometimes. They're still not that good—and I bet even Fable is this way—at actually constructing a really good essay, a book chapter, or a book, which is even harder.

The thing about a book, and a good essay too, is that you have to pick poignant structural metaphors and then embed them throughout the piece, but you also have to leave them implicit sometimes. Very frequently, the best part of writing well is a kind of restraint, right? It's exercising this kind of, “I could have gone there, but I'm not going to. I'm just going to let that sit a little bit.” I have not seen an AI model that can really do that all that convincingly yet.

It's one thing that I think remains a human skill. I bet the labs haven't really tried to make the AIs into good essayists. They've tried to make them good analytic prose writers, like Wikipedia article writers or economic writers, but writing essays in the classical sense of the word is not that economically useful.

I want to push you a little harder on this. Do you think that you want to have this sort of very distinct identity where the things that you put out are truly only yours indefinitely? Or do you envision a time when Fable 2, or whatever is available, would lead you to say, “I'm open to—and maybe I even see a need to—create outputs that are meaningfully co-authored with AI systems”?

I already feel like a lot of what I do is meaningfully co-authored with AI, just in the sense that AI is such an important research tool and, at this point, thought partner to me that I already consider AI to be really quite important in that way.

AI models will be in the acknowledgments of my book because, from the ground floor all the way through to the end, AI has been extremely important in helping me think about and conceive of the project altogether. That includes really specific things, doing research, negotiating the contract, and figuring out all the stuff you have to do to write a book—everything. So already, that feels like it's the case.

But in terms of actual communication, no. I basically think there'll be a human preference to read things that we know, or that we have faith, are written by other humans, and it will be hard to maintain that faith.

I don't think anything I've ever written is written by AI. To be clear, there are some pro forma things that go out under my name that are written largely by AI or substantially co-authored. I might file a regulatory comment, for example, a public-interest comment on a proposed rulemaking or something. I might write a letter to an attorney general or an ambassador. I might do things like that that are substantially written by AI with detailed prompting from me, for sure.

Immigration letters—I do that all the time. I'll write immigration letters for people in support of green cards.

When it comes to actual writing, I don't think anyone has ever accused my Hyperdimensional or Twitter posts of being written by AI. I think people have that faith, and I think that faith—that you actually communicate yourself—will have some value in the future, even if the AIs are, in some sense, better writers.

At this point, it might be the case that Claude can do a better turn of phrase than me, but I still haven't seen anything that's truly a better essayist. Maybe Fable will be. I've got to try it. But even then—and I don't think it's an if—

I think it's a when they get better. I think there's still probably this preferential advantage that people will just want to read stuff that's written by other people.

Yeah. It's coming for all of us, but maybe we'll choose one another over the AIs.

The other thing is experience, right? It's part of why my writing is interesting, I hope, to people: I have walked a particular path through life. It's not the most interesting path in the world, but every single path through life is highly improbable and therefore intrinsically very interesting. Everyone's path through life is like that, and if you simply have the gift of observational acuity and curiosity, you will notice that the world around you and the path you're walking through life is fantastically interesting.

There are all sorts of interesting things to say, but that will be genuinely unique to you. Whether it is the death of my father, working and sitting in the Roosevelt Room in the West Wing, or now going to OpenAI, I hope that I'll be able to draw interesting observations from those things that a machine intrinsically cannot draw because the machine did not do that thing. Although, in a weird way, the machine will walk its own path, I'm sure.

That might be a perfect note to end on. Is there anything else that you would want to leave people with or invite them to help you with in any way?

No, no, very thoroughly done. One thing I would say is that my team is going to be hiring. It's not going to be a super-big team, but we are going to be hiring.

If people are interested, I'm easy to find on the internet. It's best to email me. My email address is on my personal website, which is deanball.com. Or you can just—if you subscribe to Hyperdimensional, you can literally just hit reply to any Hyperdimensional post, and you will go directly to my personal inbox.

So if you're interested and you think that you might be able to contribute something interesting to the team as I've described it, please get in touch.

Dean Ball, thank you for being part of The Cognitive Revolution.

Thank you, Nathan.