Dean W. Ball 谈美国 AI 行动计划与白宫 4 个月经历

Hello and welcome back to The Cognitive Revolution. Today, I am thrilled to have Dean W. Ball back for his 5th appearance on the podcast, fresh off a brief but historic tenure as senior policy adviser for artificial intelligence and emerging technology at the White House Office of Science and Technology Policy.

Dean is living proof that a talented person with a passion for understanding AI can go from a newcomer to the field to the highest level of influence in as little as a year. After spending the first 10 years of his career at policy think tanks, Dean took a leap of faith and quit his job to start thinking and writing about AI full-time, launching his Substack, Hyperdimensional, in January 2024. Around that time, he sent me a Twitter DM about apparent contradictions between OpenAI's preparedness framework and its Superalignment plan. He wondered whether OpenAI could plan to rely on automated alignment researchers for safety while, at the same time, flagging model autonomy as a leading source of catastrophic risk.

It was a good question, and I became an early subscriber. The first of Dean's posts that really stuck with me was one he called “Software's Romantic Era,” about his first interactions with Claude 3 Opus, in which Claude expressed the desire to understand what embodiment feels like and to hear music. Claude's first request would be Beethoven's Ninth Symphony, a fascinating choice since Beethoven was deaf when he wrote it and never actually heard it himself. I still remember one of the concluding lines: “Here, finally, is an AI system whose thoughts I want to hear.” What a brilliant way to capture the spine-tingling difference between Claude 3 and earlier models, and more broadly between AI and all other technology.

Dean's profile began to rise in a serious way thanks to his early criticism of California's SB 1047, which inspired open-source and progress advocates to take an interest and ultimately led to multiple rounds of revisions. While we never ended up fully agreeing on the bill, I was impressed that, unlike so many of the shrill voices in that debate, Dean moderated his criticism as the bill itself evolved.

From there, he started showing up everywhere: at the Curve, as an FAI fellow, as an affiliate of Phantom [?], and with a podcast of his own. Then he got the call to join the White House.

Regular listeners will know that I did not vote for President Trump. Nevertheless, I was very excited for Dean and, more importantly, for the country. There are not many people on the political right—or, frankly, anywhere—who combine the impulse to get hands-on and use the technology to the fullest with a healthy respect for how powerful it could soon become, plus a deep appreciation for the nuances and challenges of integrating it into society and government effectively.

I thought then, and I continue to believe now, that Dean was perhaps the single best person the Trump administration could plausibly have picked for the role. At the White House, Dean led the effort to develop America's AI Action Plan, which very well might be the most well-received policy the Trump administration has put out to date on any topic, with positive reviews from many people who expected to hate it, including an endorsement from President Biden's former national security adviser, Jake Sullivan, on the most recent episode of this very podcast.

Yes, it is a broadly accelerationist document framed in terms of beating China, but it's also extremely sophisticated, calling for what I believe will prove to be very prudent investments in evaluations and risk assessment, I/O security, AI-enabled science, mechanistic interpretability, compute governance, and much more.

In this conversation, we discuss why Dean is now leaving the White House, including his candid self-assessment that while he's great at conceptual work, implementation can probably be better done by somebody else. We explore how the American political right is grappling with AI, from deregulatory impulses to culture-war concerns and worries about AI's impact on children.

We also dig into the mechanics of unlocking spare gigawatts for data centers, hear Dean's perspective on which frontier developers should be considered live players and how they're perceived in Washington, and get his perspective on what the AI safety and governance communities can best do to complement the government's work going forward.

Along the way, we also hear how Dean used an LLM to simulate interagency feedback meetings, how the 4-year presidential term—which ends in January 2029—just happens to put the administration in sync with many AGI forecasts, and why Dean believes that the Republican Party might become the better home of the AI safety world in the long run.

Between his rapid rise and his brief White House tenure, you might say that Dean embodies short timelines. While I am sad to see him leave such a high-impact role, I hope that his story inspires others who may have similar potential to take their own leap of faith and get into the AI game. We need all the Deans we can get, and we need them as soon as possible.

Finally, I want to say congratulations to Dean and his wife on their upcoming first child, another reason he's leaving the White House, and to thank Dean for doing this. He could easily have booked a much bigger platform for his first post-White House interview, and I'm sure those will still come. But I'm grateful you chose to have this conversation here. It was super fun for me, and I always appreciate the chance to dig into so many important details.

With that, I hope you enjoy this deep dive into 4 months in the Trump White House and the crafting of America's AI Action Plan with Dean W. Ball.

Fresh off a short but historic tenure as senior policy adviser for artificial intelligence and emerging technology at the White House Office of Science and Technology Policy, welcome back to The Cognitive Revolution.

Thank you for having me, Nathan. It's great to be here.

Yeah, I'm excited to catch up on your recent tour of duty. It's been an influential one, to say the least. We've come a long way. This is your 5th time here on the podcast.

The first time, just a little over a year ago—I think maybe a year and a half—we were talking about frontiers in neurotechnology, a little side hobby of yours that has probably withered a little bit in the meantime. But if ever there was a Cognitive Revolution bump—and I say this in jest—it would be going from just starting to write about AI all the way through to this White House role and playing a key role in the writing and release of the AI Action Plan. It has been quite a run for you.

I think there's a ton, obviously, to get into. Maybe, for starters, do you want to tell us just about your experience? First of all, how did you even get this job at the White House? Was it something where somebody tapped you on the shoulder and said, “Hey, we want you for this”? Did you have to interview? Did you have to throw your hat in the ring? Tell us the whole story.

Yeah, it's a great question. The way these things work is that I never actively solicited a role in the administration. What I would say I did is that, after the president won in November, I published a piece maybe a week or 2 weeks afterward called “Here's What I Think We Should Do.”

Really, the goal there was very much: We have a new administration coming in. I think I would have done that if Kamala had won. I obviously would have felt less excited about the prospect of those policies being realized, but that's basically what I would have put forward either way.

I wrote that, and it did reasonably well as a piece. Some people who I knew were going into the administration reached out to me. My view was always, I would happily do it. It's not something where I'm desperate to go into government.

Through a lot of different social interactions, I ended up coming into conversation with Michael Kratsios, the director of OSTP. This was before the administration, so he had been appointed by the president and was not yet confirmed by the Senate. We talked a little bit, and he asked, “Do you think you'd want to come into government?” I said, “Sure.”

Then it was several months of figuring out administratively how to achieve the job. The hard part was not actually getting the initial offer from Michael. The hard part was figuring out administratively how to make the job happen.

OSTP, the Office of Science and Technology Policy, is unlike a lot of components of the White House. OSTP was created by Congress in statute, so it's technically an agency. What that means is that it has a budget set by Congress that doesn't go up very much. OSTP has a pretty tiny budget. The whole White House, actually, generally has a pretty small and relatively inflexible budget.

Many, if not most, OSTP staffers are either affiliated with other parts of the government or affiliated with nonprofits or universities and, through various bureaucratic means, get transitioned into the OSTP role. That took a while to figure out, and I was uncertain it actually would be worked out. I knew it was being worked on, and I knew it was a possibility, but I didn't bank on it.

Then it all came together really quickly. In the last week or 2, it all happened very quickly. I remember finding out that my last day at Mercatus would be basically the same business day that I figured out, “Okay, this is happening.” It was like, “Today's my last day,” and I had to go tell everyone, do a quick rush and say goodbye to everybody, get all my things in a box, and then start in government the following Monday.

Sounds like a little of that red tape that might need to be cut to get the right people in the right seats faster in the future.

It’s hard. Hiring for government is hard. It’s really true.

So, you’ve been writing, and people have been exposed to your ideas. You’ve gotten a lot of good feedback on your ideas, but I have to imagine it’s a pretty big difference to wake up one day and be like, “All right, today I’m going to the White House,” and now you’re not just writing to socialize ideas and hopefully gently steer public discourse and thinking, but you’re actually going to be in the seat, playing a significant role in making the rules, making the policy, deciding what we should do, with real consequences.

You get training when you show up—training on some stuff. They make sure you know that people can’t pay for your bar tab or your lunch, things like that. There are various procedural ethics rules that you have to comply with. They’re very serious about that. There’s IT training, and then beyond that, no.

The thing that’s interesting about joining government in general is that I have truly worked at universities and in think tanks that have more internal bureaucracy than the White House does. There are certain things in the White House that you run into, and it’s like, “Wow, that is a hard bureaucratic wall, and there is just nothing you can do about that. That is just there.” But then a surprising amount of stuff is way more flexible than you would think, and you can actually move quite quickly. That was one of the things that I really liked about being there.

No one really prepares you for it, and nobody can prepare you for it. It doesn’t ever stop being surreal. I was only there for 4 months, but I have a feeling that if I had been there for 8, 12, or 16, it would never have stopped being surreal.

I think there are a few aspects to it. One thing I would say is that I think it’s very hard to know if you’re the kind of person for whom power is an addictive drug that you just want more and more and more of, like a drug addict wants more of whatever their drug of choice is all the time, and you kind of need it to sustain yourself, versus whether you’re the sort of person who doesn’t. I think it’s just really hard to know.

A lot of people in Washington, D.C., are that former kind of person. What I think I learned about myself is that I’m probably more the latter than anything else. I found myself relatively unmoved by it, and if anything, in some ways burdened by it. The amount of incoming you’re going to get from other people is insane, right?

At any given point, I probably had between 30 and 40 unread communications of various kinds. People reach out to you on Signal, WhatsApp, email, and all sorts of places, asking you to meet or do various things, read this, think about that, or whatever else. There are just a ton of requests for your time.

Even people who are your friends or colleagues—the relationship with them changes in some really important ways. It’s very hard to anticipate how to deal with it. It’s a really weird situation. I don’t know if it ever stops being weird, to be totally honest, because these are pretty extraordinary jobs.

It’s a tremendous honor, though. You just walk around, and when you have a good day, you really feel like, “Wow, I actually really drove this thing forward and did whatever.” It’s real. It’s actually a real thing that we’re going to do. It will be public policy of the United States of America.

It’s crazy. You have to take it very seriously, and it’s a huge responsibility.

You told me you wrote a letter to yourself before going in to try to ground and calibrate yourself to some of those possible issues. How would you advise people who might be thinking about taking on such a role? Should they write that letter? What else can they do? Since there’s no training, what can you offer in terms of wisdom to people who might step into this in the future?

It’s a really good question. I would say a couple of things. First of all, if you’re going into a policy role—not every role; there are operations roles and all kinds of different roles in government—but if you’re going into a policy role, you should have all of your policy ideas substantively pretty well baked by the time you go in, because you will not have any time to develop policy. You know what I mean? You’re not going to be sitting around thinking that much. You don’t have time to think.

The volume of things that happen to you when you’re in government is wild. As an example, a major developed country is coming through the door. They’ll be at the White House in 3 days. When you’re dealing with things at that velocity, you have to have your policies very well developed already in your mind, or at least reasonably far along.

But also, I think you need to have an important set of principles. Ultimately, your job as a staffer in government, whether it’s the White House or elsewhere, is to execute the vision of the president. Your job is to achieve the president’s objectives.

There are a lot of staffers who convince themselves, “Oh, well, the president is operating at such a high level. He doesn’t know exactly—I know the right thing to do here, even if it somewhat disagrees with something that he said recently.” That will not get you far. People do it all the time, but particularly in this administration, that will not get you far.

You do need to have a sense of what’s important to you personally. It might not be that the president’s objectives conflict with your principles. That’s certainly possible, but it’s not so much that. It’s more about drift.

Being inside of government is very weird. It’s like you’re inside of a self-contained cube with glass walls, and everyone in the world is shouting into the cube all at the same time. You can hear what people are saying, but you can’t exactly interact with people in quite the same way.

As one practical example, let’s say you have a question for companies in an industry about some aspect of their business that might be nonpublic, and you feel like you need to know it to make some policy determination. There are a lot of rules that govern how that interaction can go. Those rules structure the interaction. A lot of your interactions will be governed in ways that will not be intuitive to you if you don’t have prior experience in government.

One of the things that I felt quite acutely was that I felt myself drifting a little away from public scrutiny of my ideas. That’s weird, because in a certain sense, the action plan is the most publicly scrutinized document, by many orders of magnitude, that I have ever contributed to. Nonetheless, that was a hard thing for me, and it turns out that it’s really important to me. It’s really important to me to get public scrutiny of my ideas.

One of the things I wrote to myself in the letter was that there’s a risk of drifting from ground truth. You’re just in this bubble, and so you might drift away from the ground truth of the world and become more attuned to the logic of the system in which you operate. That system has its own very distinct logic and does not necessarily operate according to anything intuitive.

The second you start to see that happening, you have to be careful, because that can really change who you are in the long term. That was something that was really important to me, and I would say that’s a general flavor of the thing that I think is very common to encounter in government.

Yeah, that's interesting. The time piece calls to mind an obvious question: How much leverage are you able to get from AI in your work at the White House these days?

It's interesting. I always had to be really careful how I answered this question inside the building, but I would say this: LLMs are actually not permitted on White House computers. That's not true of all of government; that's specific to the White House. I never got super into the weeds on exactly what the problem is, but it relates to compliance with a law passed after, I believe, Watergate, called the Presidential Records Act.

The Presidential Records Act relates to the specific documents produced within the Executive Office of the President, which is the name of the bureaucracy that's also called the White House. There's something about compliance with that law that affects the ability of the White House to use lots of modern technology. So we can't use Slack or Microsoft Teams. You can't use Zoom or Google Meet or Microsoft Teams; you have to use Webex. Google Docs is another good example, and so are LLMs.

The main aspect of that has to do with sharing predecisional documents, draft documents, things that are literally policy being developed live. Obviously, pretty much everything I worked on all day long was deliberative, predecisional drafts. Those things very, very seriously cannot touch LLMs. They can't leave government computers. Obviously, that's a very important thing and not something I would ever mess around with.

That being said, I used AI as a chief of staff for me and as kind of a research assistant. Not a single word of the action plan was written by, edited by, or seen by AI prior to the release, at least not by me. But there were so many times when it was like, "Okay, well, I have a question I want to ask you. I want to try to develop some policies relating to XYZ." Sometimes it might be, "Well, let's brainstorm 1,000 different things." Sometimes I used it for that.

Most of the time, I actually had a pretty good sense of, directionally, what I wanted to do, and it was about scoping it correctly. It was about really saying, "Okay, let me understand the statutes here really well. Let me get all the statutory authorities, all the laws that enable us to do this thing that's a recommended action in the action plan." You have to understand all of the regulatory history, legal history, and interpretations of all the relevant statutes. You have to really understand how exactly you're constrained and what exactly you can do if you're trying to do everything you can.

The way it felt to me was like I was thrown into the cockpit of a plane. I've never operated a plane before. In fact, I don't know that much about the federal government in the grand scheme of things. I know a lot more now, but you're like, "Okay, what do all these switches do? What do all these buttons do?" The LLM is a really good guide for that.

The other thing I did that I think was really valuable is this: I was actually looking earlier today at some initial notes that I made back in February when I was drafting for the public RFI. They were pretty close. By late April—by 2 or 3 weeks into the job—the action plan was, I would say, ⅔ to 75% of where it ended up publicly, basically. But that ⅔ to 75% is a blurry image that was sharpened by lots and lots of interagency feedback.

It's the difference between a useful image and a not-useful image, right? One thing is just a bunch of blobs, and the other thing is actually an in-focus image. But one thing that I found very useful to do was, for all the different items in the plan, without directly taking the text of the plan, just saying, "Okay, this general idea, this general concept of doing something—let's simulate interagency feedback on that. Let's actually simulate what an interagency meeting would be like here. What is the grizzled career in-house general counsel at the Federal Communications Commission going to say about this? To this young whippersnapper who works on AI, what's he going to say to me about my bright idea?"

You simulate that feedback. Over time, you can basically do the first—maybe not the entirety, but the first couple of meetings—that you otherwise might have had to spend a lot of time on. You could simulate that. That ended up being really useful, and I think what I always say to people is: The action plan was not at all written by AI. The ideas in it did not come from AI. It was all written by humans—and not just me; many humans contributed a lot to it—but it is probably an example of an AI-enabled productivity boost, because it got done in a pretty short amount of time.

It was about 3-ish months. It is true that a very small number of people really drove the text, and so I think it's probably evidence of some kind of productivity boost from AI.

The idea of simulating interagency meetings is definitely a fascinating one. Do you want to give any more detail? You mentioned there are some surprises when it comes to the bureaucracy: Some things are just hard barriers; other things are more flexible than you'd expect. I draw a blank on what that would actually look like in practice. Are there examples that would be informative?

There are definitely some things. I mean, one that's a classic one—and this is something that's actually kind of a meme in policy circles, but it is true—is something called the Paperwork Reduction Act, which was passed to make government more efficient and reduce paperwork in government. What it means is that any time you, as a government employee, want to reach out to more than 9 members of the public, whether they be individuals, companies, or nonprofits, if you want to ask for the same information from more than 9 members of the public, you have to go through an entire bureaucratic process for that.

There are all sorts of rules that govern it, and you just wouldn't guess that because you might be, again, using the example of reaching out to industry. I think one thing that a lot of people taking my job—one of the first things they might want to do—is go talk to the frontier AI labs and say, "Okay, where's the secret briefing? Where's the briefing for high-level government employees that you don't share with the public?" You have to be careful about that, because if you reach out to more than 9 companies, then you've brought yourself into Paperwork Reduction Act territory.

Then I have to ask: What's in those secret briefings for the government?

Are there actually such briefings, and what can you tell us about them?

There are. There are meetings where we learn things that, of course, aren't public—not just with the AI labs, but with all kinds of companies. You come in and learn all sorts of things that aren't public.

I would say there isn't the sort of meeting where they come in and say, “Here's our AGI roadmap.” That doesn't happen. It's not like government employees are known for keeping secrets super well, right? So the companies are pretty cautious about what they choose to share and not share in meetings. I think different people push them to differing degrees.

I would say there were a lot of times when talking to companies about nonpublic information was a very important part of my decision-making. But I also think it's probably the case that there are certain kinds of questions that many different people in government have about some aspect of AI. How much power do you think you guys are going to need? That's a classic one.

I feel so bad for the people on the OpenAI, Anthropic, Google, and Meta policy staffs who get this question from 20 different government employees a week in different parts of the White House. It's like every single one of them gets lots of briefings, but there's not coordination. We're often not aware of what the others are doing.

It's hard, because how often do I see every single staffer with a plausible interest in that question at every single agency that relates to it, which are many? One of the other things that would be unintuitive to a lot of people is that you might know it intellectually, but the government just does a lot. There's a lot of stuff inside agencies. There's a lot of people doing all kinds of different things that it would never really occur to you.

You would never be like, “Yeah, why is the State Department here?” And it's like, “Oh, because that's actually extremely important. They were doing this thing, you know.” So, yeah, those meetings do happen. They're probably not as structured and organized as they could be.

To be honest, in some ways I felt like I knew more about what was going on inside the labs outside of government, just to be totally honest with you. One of the things the labs do once you get a .gov email address is put you in touch with researchers and stuff. I had a lot of friends at the researcher level, but that all of a sudden becomes a lot more complicated once you're in government, talking to other researchers at labs.

It's definitely possible to do, but they're going to be more nervous about it. They're going to want to bring in their policy people. The second you bring in the policy people, it's like, “Okay, well, now we're going to have a script, and we're going to say exactly the right stuff.” So it was always very funny.

Sometimes the labs would organize these briefings. There's also different degrees of knowledge about this stuff, right? I'm someone who wants to have conversations at the relevant margin where knowledge is accruing on a topic. I want to ask, “What are your thoughts on the sample efficiency of reinforcement learning?” I want to ask about some architectural question.

But a lot of other government employees with plausible nexuses to AI need a much more basic kind of briefing, and that's typically what the labs prepare. There was a particular briefing—I won't say which company, and I won't say who the researcher was—but it was a researcher who is very well known on the internet. The researcher was with the policy people, and they were briefing us about some of the latest developments they were rolling out in their models.

They so desperately wanted it to be a certain kind of script that was designed to satisfy what a staffer with little context would want. I kept asking all these highly technical, in-the-weeds questions. The policy staff don't stop you from asking those questions, but there are definitely moments when the researchers will start going on into something and the policy staff will be like, “Whoa, stop,” you know?

So, yeah, it's actually interesting. I don't know if it's definitely true that I knew more about what was going on inside the labs. I felt like I had a better grasp on it as a nongovernment employee, but I think it's possible that that's true. I haven't thought enough about it, but I think it's possible that that's true.

Yeah, interesting. Why are you leaving now?

Yeah, it's probably a surprise to a lot of people. I'm leaving for a few central reasons, and I'll go in order. Some are highly personal, and others are a little bit broader.

The most personal one is that my wife and I are having our first child in a few months. In fact, I found out about her pregnancy the night before I started at the White House. It was like 9:00 p.m. that evening, and it was like, “The next day, you're going to the White House.” So it was a lot to process all at once.

I want to be able to spend time with them and devote—frankly, enjoy—the last couple of months of preparation, and be able to really be a very attentive and loving father in the initial months of parenthood. That's part of it. I don't think it's impossible to do that as a government employee, but it's hard. I worry that I would do one or both of the jobs badly if I had to balance them both.

The other reason really is about what I'm good at and what I'm not good at. I spent a big chunk of my career as a manager, and I was decent at it. I wasn't bad at it. I did reasonably well and got promotions and stuff like that.

But the truth is, I didn't really enjoy it. I didn't love it. The fundamental reason for that, I later discovered, is that my success depended quite often on the whims and attitudes of other people and on what side of the bed they woke up on. I just wasn't independently responsible for my own success.

I found that when I did projects where I was in control of whether or not the output was high quality, I found that to be much more satisfying. With the action plan, there was this rare opportunity. Obviously, I was not the single author of it; it wasn't like a Substack post where I just hit publish. But I exercised a lot of creative direction over the action plan, as well as stylistic direction and all these sorts of things.

That's the kind of thing I'm really good at. I'm just not that good at managing the implementation of the plan—making sure the agencies are doing it, thinking about all the procedural levers you need to be pulling, all that stuff. I don't have confidence that I'm great at it. I think I could do an okay job, but if you only think you can do an okay job, you shouldn't be working for the president.

If you only think you can do an okay job at catching a football, you shouldn't play for the Miami Dolphins. You're supposed to be the best in the world. This is the AI policy of the United States of America. It has to be really, really effing good.

I had high conviction that I, along with the rest of the team, could deliver that in terms of the text of the action plan. I had high conviction about that. I think that the administration is able to implement it. I think there are a lot of very savvy and driven people who can also implement the plan, but I just don't think that I am necessarily one of them.

I actually think there are probably ways in which, just reflecting on my own writing trajectory over the last 18 months, I would be actively unhelpful. If we went back to one of those earlier episodes of The Cognitive Revolution, the first SB 1047 debate we did, the neurotechnology stuff, I'm sure there's a ton of stuff I said in those episodes that nowadays I would look at and very happily discard. I would very happily say, “Oh, yeah, I totally don't believe that anymore. That's fine. I've evolved past that.”

You can't really do that when you're setting the federal government's policy. It's a big ship. You can't really be like, “Yeah, well, I changed my mind about that. I decided that was dumb.” Certainly, there will be room for flexibility in terms of exactly how it's implemented, but you can't just throw things out at will, and you also can't add new stuff all that easily.

My nature is very much that everything I say is provisional, and I'm always evolving. So, just personally, I think I wouldn't be good at it. That's basically what it came down to.

Finally, some of the stuff we were talking about earlier: I found myself really missing public feedback, regular scrutiny, and feedback from a community of my peers. The experience of doing that was just so useful to me epistemically. It was so useful to me, and it's much harder to do when you don't get that real-world feedback.

That was a concern I had, and I want to get back to that because I think that developing—getting my head around really thorny ideas, developing answers to hard questions, and then trying to communicate those answers compellingly—that's what I do. I think I do an okay job at that. I don't know if I'm the master of the bureaucracy. I just really don't, and I think you have to be honest with yourself about things like that and not try to do more than you can.

The way I've been describing it to people is, “Know thyself.” All of that put together—not any one of those things might have been enough. I might have maintained it all, but the combination of all those things made it feel to me like now was probably an appropriate time to take a step back and contribute to both the action plan and AI policy more broadly from the outside.

Frankly, I think we're still so early in terms of societal conversations on AI that there is so much intellectual work to be done and civilizational scaffolding to be built around this technology. I think that stuff—I mean, one of the key ideas of the action plan is that that kind of stuff, in addition to the technology itself, is fundamentally going to be built by the private sector. I kind of want to return to doing that.

That's the long-winded answer, I realize, but that's kind of where I came down. It was a tough decision. It really was. It's not like this was easy for me to do.

One thing I think is important to note is that I was not pushed out because of political infighting. To the extent that there are people on different sides of debates who might externally be referred to by the press as factions, this is not an indication of one faction rising over another. This is a decision that I made personally myself, and it's not because of any policy disagreements with the administration.

It's not like I was mad about something and leaving. No, not at all. I believe I leave with absolutely no bad blood and nothing but warm feelings for all of my colleagues in the federal government.

That's great, and a notable contrast to many departures that we've heard tell of. Yeah. Well, congratulations again on starting a family with your wife. Obviously, that's a once-in-a-lifetime transition that you'll be able to give the full energy to, which, I think it's safe to say, you would have to make some compromises on if you were staying in the job.

That makes a ton of sense. I want to get into a little bit of the context in which this AI action plan was developed, and then you can get into some of the details of it as well. One thing, though, that you said toward the end there that I found a little puzzling—and maybe this is just my never-worked-in-government naiveté—is what stands in the way of a more open, feedback-welcoming process?

I don't know how many people saw drafts of the AI action plan or contributed to it, but presumably it was a reasonably small number compared to how many people would have happily read and commented on a draft if you had put one out there and asked for feedback. Why can't that happen? Or could it happen, and it's just an administrative decision?

I'm confused, though, especially because AI specifically is such a dynamic situation that the faster feedback cycles we could get, presumably, the better. So, yeah, what stands in the way of realizing the way of working that you prefer within the White House?

Well, I think it's a couple of different things. First of all, there are absolutely ways the action plan itself, before it was the final product, went through a request for information, which is a formal process by which government agencies can ask the public to submit their thoughts on a given issue. Something like 10,000 people did that for the action plan, which is a pretty huge turnout.

A lot of them were really heartfelt things written by individuals. I read many of these comments. Of course, every company—it really was amazing—the range of corporations, industries, Hollywood actors, former politicians, members of the public. So many people submitted comments for this.

Then, of course, I met with hundreds, if not more than 1,000, people from the public in various ways leading up to the action plan to talk about ideas and gather feedback. It's not like that. It's more that I personally wrote a Substack post every week, and very often there was lots of speculative stuff where I said, “Yeah, I think we should probably do stuff like this.”

You can't do things like that. You can't do things like that because when you're working on something that comes from the government, there's just a different gravity to it. It's a huge machine. We can think that we're stepping lightly and, in fact, be stomping around.

If we were to say, “Here's an early draft of the action plan”—let's just say we put it up on arXiv and said, “Let us know what you think”—the plausible blowback from that is enormous. It could lead to political scandals. There are conversations going on with Congress.

There are a billion trade negotiations going on. There are companies that will freak out, right? They’ll be like, “Oh, this would damage my business,” and so on.

This is why—not the only reason, but one of the fundamental reasons—people leak stuff: to basically make things like this happen tactically, as a move within the bureaucracy. It’s like, “This other guy is pushing for something that’s really controversial and would be really unpopular. So what I’m going to do is—he’s doing that at the staff level—leak it to The New York Times. Then all the senior people are going to see it, they’re going to get mad, and that’s going to kill the idea.”

A lot of times when you see leaks, what you’re seeing is someone attempting to kill something from within. A public disclosure process would totally do that, and then there would need to be lots of procedure around it.

The other thing I would say is that the action plan—one thing that, in theory, you could do—this isn’t something the government tends to do for reports, which is kind of what the action plan is, but it does do it regularly for public policy. It will release draft regulations with notice-and-comment rulemaking, right? There are various different things you can do. You can request information on a potential rule. You can be like, “Look, we’re thinking about doing this kind of regulation. What should we be aware of here? What do you think we should know?” That’s a structured way of doing that.

You can also release drafts of regulations and say, “We invite comment on that.” You can totally do that, and governments do it all the time.

I think the trade-off, specifically with the action plan and the process for getting the action plan from a blank sheet of paper to the final product, was largely conceived of and run by me. I obviously ran it by other people higher on the totem pole than me. I was like, “If I do it this way, is that okay?”

The trade-off between having a really collaborative process, where lots and lots of people—even inside the government—are sharing comments, is just the classic design-by-committee thing, right? I would say, as a general matter, interagency feedback, from my perspective, was almost always really productive, constructive, and useful—and essential for making the action plan good.

But we did employ a different process. We didn’t do a traditional interagency process. Usually, what happens is that the White House, if there’s a policy you’re working on, will convene a policy process. Every agency with potential equities in that thing shows up, and then you talk about the thing, conceptualize a policy, share drafts, and so on. That feedback process can be quite extensive.

We kind of did that for the action plan, but we didn’t do it super formally. For the most part, we only shared with agencies the portions of the action plan that were directly relevant to their agency. The one thing that I felt very strongly about was that I did not want every single person in the federal government commenting on every aspect of the AI strategy, because if you do it that way, there are just going to be more people with unproductive comments. I think probably a lot of cool ideas might have gotten squashed by that, to be quite candid.

I think it’s absolutely essential that you do a formal process if you’re actually doing policy. The action plan is recommendations for policy. They’re strong recommendations, but it is not literally the president saying—you know, the president didn’t sign the action plan, as opposed to the executive orders that the president signed along with the action plan.

The executive orders use words like “shall”—not “may,” not “should,” but “shall.” That is the president saying, “This is a command. Go do this,” to federal employees. So that’s a legally binding document. The action plan is not legally binding. If you’re making legally binding policy, then of course you have to go through a traditional process and tolerate some of the bumps and inefficiencies of that.

The executive orders, for example, that the president signed along with the action plan went through a traditional process, but the action plan itself was a little more flexible.

One big background thing that I’d love to understand from your perspective is, generally speaking, what is the political right thinking about AI now? We’re headed into this four-year term. Many people are saying that we might have AGI—who knows, maybe even superintelligence—certainly some sort of powerful, potentially transformative AI in this term.

As with many things, when you get into power, it becomes kind of your problem, where in the past you were able to comment and criticize the people whose problem it was before you won the election. How would you characterize the factions, if you want, or maybe just the different perspectives that ultimately make up the coalition that elected President Trump? What different perspectives or priorities are they bringing to the AI discourse?

I guess I would start maybe with Trump himself. We’ve heard a few comments, but I don’t really know how much he thinks about AI. Is he using ChatGPT? Does he have a strong take on some of these core issues?

Then there’s the tech right, which many listeners to this show, I think, were at one point optimistic was going to be ascendant. It seems like maybe there’s also the religious right, which has a very different set of priorities and would seem to be in some tension with the tech right. There are probably other groups that you might put into that mix as well that have strong and perhaps quite distinct points of view. How would you lay out that landscape for us?

The president was elected with a really diverse and broad coalition. It’s one of the things that is so striking, and I think such a narrative violation for what you’ve heard from a lot of people in more left-wing media over the last 5 to 10 years about the Republican Party and President Trump.

The reality is that in terms of income level, ethnicity, way of life, background, and so on, the president’s coalition is just really, really diverse. It’s a coalition that has many different views on AI, and those views are also evolving in real time. I think you actually started to see them evolve during the time that I served in government quite significantly.

Very broadly speaking, there are people who come from the somewhat more traditional deregulatory impulse of conservatives that’s been around for a long time. The president certainly is not a hardcore libertarian by any means, but he has a lot of that—fundamentally, America, the business of America is business—as President Coolidge said. I think that’s very much part of the president’s personal intellectual DNA, so I think he definitely feels that way about AI.

I think all of the principles of the action plan are things that he thinks are enormously important. He has now weighed in publicly on issues like copyright and preemption, and on both of those things he’s leaned, I would say, in the direction of AI development and adoption, rather than putting up new blockers. Of course, environmental permitting and things of this kind are something that everyone in this administration is very serious about.

I think there are some areas where there’s a rift among some on the right, based on the level of risk that you perceive from the technology and also how consistent you see the technology as being with social media, the internet, and things like this.

There’s a lot of hostility—and, frankly, I think justified hostility—on the right against big tech, particularly the social media and user-generated-content platforms. For a long time, people felt as though right-wing ideas were discriminated against, and I think it’s true that they were. I think things like fact-checking, misinformation analysis, and things of this sort very often were deliberate attempts to shut down right-of-center viewpoints from broad dissemination on the internet.

I think that’s a shameful thing, and I think those companies should be ashamed of it. There are a lot of people on the right who I have argued with, both in public and privately—in group chats and all sorts of things. I have tried to say, “Look, you are fighting the last battle, and you should stop doing that. Stop fighting the social media battle.”

While that is an important thing, no doubt, and the outputs of AI systems certainly matter a great deal, ensuring that they’re truth-seeking and not actually the result of top-down ideological programming is important.

I think that is extremely important. But it’s interesting because when you talk about that issue, you get very quickly into the most important issues of the traditional AI safety world. For example, with the woke AI EO in federal procurement that the president signed, you very quickly get into questions like, “How do we actually know that the system is aligned? How do we actually know what this thing’s going to do, that it’s going to do what we want it to?” You get very quickly into issues of concentration of power, where I think there’s a broad perception that, my God, this technology is going to be so foundational. We really need to understand the character, the virtues of these systems, and the values that they hold.

Reality has a way of coming at you regardless of what you believe. If you believed that issues like loss of control and alignment were all lefty doomer stuff, all EA stuff, I think a lot of people on the right were saying that a year ago. Now they’re actually coming in and thinking, “Okay, I care about this.” And it’s like, wait a minute—these are actually fundamental issues in AI, which is what people have been saying for a long time. So I think what you’re starting to see is the right get its head around these issues in a much more serious way.

I’ve predicted before in public, before I joined the administration, right after the president won, that I thought there was a reasonable chance that the Republican Party would be the better home for the AI safety world in the long term. That’s because of the way the incentives of the party work and the way that the party is connected. So, you are starting to see that. There are some rifts there, and some of those things are people fighting the last battles. Some of those things are people starting to get their heads around these bigger issues, like what the values of the system are going to be.

I think an issue that falls right in the middle of those two things relates to child safeguards specific to children’s use of AI systems. There’s obviously the very tragic story of the boy who killed himself—I think he was 14 years old—in Florida. That’s an issue that really resonates. It’s something that a lot of conservatives are very worried about.

I’ll just put my cards on the table: for sure, this is not federal government policy. This is not the opinion of the White House or the Office of Science and Technology Policy. But when I see big, well-capitalized companies making pornography with AI available indiscriminately to people, I’m sure that there’s some sort of age-gating, but I doubt it’s really all that good. I just get mad for so many reasons because, obviously, that stuff is inevitable. But let that happen on the open-source model in the North Korean pornbot farm. Don’t take money from the world’s biggest institutional investors and be one of the flagship brands of AI to the world, and then do this kind of stuff in such a public way. It’s so crass.

It’s not going to end up well with conservatives. It will not end up well with conservatives. I think these culture-war issues that have been going on for a long time are still happening, and they animate issues in AI. That intersects in weird ways with preexisting AI safety issues, and you’re starting to see those two things merge together. I don’t exactly know where that’s going to end up.

My hope—and I think what the action plan is all about—is that there are good LLM safety laws for kids that you could pass. I think there are also really bad ones. I think there are prudent things you can do, and I think there are lots of non-prudent things you can do. But the point of the action plan is to say, to a certain extent, that we’re early in this, and we don’t have to fight. We don’t have to be at each other’s throats. We can identify reasonable things that make progress at the relevant margins, and we can be positive-sum about this. We don’t have to play the traditional game. That’s a very important part of what we were trying to do with the action plan.

So, I guess I hear you saying it’s still basically a live question, right? We’ve got the deregulatory impulse. We’ve got the cultural-conservative impulse to protect the kids and uphold traditional values.

Yeah. No one’s going to win. It’s not like there’s ever going to be a day when you’re like, “Oh, that side won.” I actually just want to dispute the idea that there are factions here, or different sides. My view is that there are people with different emphases and different priorities on this stuff. But if, even within the Republican Party, they perceive themselves to be enemies of one another, it’s like, well, geez, there are some zero-sum issues in politics, don’t get me wrong. That’s definitely true, but a lot of things aren’t necessarily that way.

Right now, I think there’s a lot about capitalism and the general structure of American society that is inherently going to be accelerationist and create incentives to develop AI rapidly and diffuse it rapidly, and so on. In that sense, I think AI development is going to be in a very good place. But I think it will be really interesting to see how the somewhat more skeptical people, a lot of whom remember Big Tech, respond. I mean, it’s literally the same people in some cases, right? It’s like, “Well, wait. You guys built YouTube, and now you’re building Gemini, and we didn’t like it. We thought YouTube censored against us, and it seems like Gemini also might.” So we’re worried about this. We’re very nervous about this. We’re very nervous about the amount of money being spent and the capabilities being promised.

People like me saying, “This is going to be foundational to everything you do 10 years from now,” does not necessarily alleviate those concerns. It often makes people more worried. It could totally be true that this ends up turning into a pretty unhealthy impulse, and we get a lot of laws that freeze our society in amber because of various concerns, because you don’t want change.

On the other hand, I think it’s possible to develop AI in a way that is actually productive. That’s another thing that I think I can do. That’s part of why I chose to leave. I think I can contribute to pushing things in a more positive direction there, and I think I can probably do that more effectively outside the administration than within.

How would you say rank-and-file Republican voters are thinking about this right now? Generally speaking, it seems like the survey results show more bipartisan consensus on this particular issue than on almost any other issue. People are generally worried, and people generally want the government to do something. The public is pretty warm, I think, to the Terminator-style scenario risks, but the salience is low. Who’s leading whom right now? Is the public telling the political class what it cares about, and is the political class listening? Or is the political class leading the public? Or is the public just not focused on this enough that it’s really moving the needle yet?

When it comes to AI, I think it depends. It definitely depends. A lot of our politics is geographic, right? There are senators who represent particular states that have a lot of some kind of industry that is particularly affected by AI, and therefore, for those senators, it has a higher political salience than it might in a state that doesn’t have such things.

But as a general matter, I think it’s probably right that AI in general—and especially AI policy, among people who think a lot about it—is still more of an elite, coastal-type issue than an issue that normal people are truly fired up about. I expect that will change in some ways. I expect it’ll be very hard to predict how that will change.

One of the things that I had written on a whiteboard in my office for a while was a mantra the vice president had at an event I attended where he spoke. He said, “Our job is to make normal people’s lives better.” I really liked that. For whatever reason, that resonated with me quite deeply.

I think our job is to do things that make normal people’s lives better. We also have to communicate about the ways in which AI can make normal people’s lives better, and not talk in abstract ways about the future. The art of making it more concrete for people is going to be extremely important.

But yeah, eventually this will gain higher salience. It'll probably do so in some sort of scandal or crisis or something like that. Who knows exactly what that'll be? It seems like it's just been a pretty continuous up and up and up, getting more so. But most people are much more concerned about immigration and the economy and things of that kind.

Is there any utopian thinking or anything similar on the right that people look to for inspiration? I mean, as you know, I always say the scarcest resource is a positive vision for the future. I don't know if I'm missing any visionary writing or thinking on the right. Is there anything like that at all?

Not really. I'm sure some people are, and I just don't know about it, but not that I'm aware of. In fact, you know what's funny? We got accused by a relatively prominent person who is often in right-of-center circles—I won't say the name, but a pretty big account on Twitter—of being utopian in the action plan because we talked about AI being able to unravel ancient scrolls once thought unreadable. And I was like, that literally happened 2 years ago.

Yeah, I think that kid now works for DOGE, right?

Luke is now at DOGE. That's right. One thing is that a lot of people are not aware of the amazing things that are happening right now, how remarkable all of it is. So I think there needs to be much more of that.

There's definitely more of that to do. I think you have to make it concrete. One of the many things I hope to do in my work—not necessarily utopian, but positive visions—is some work along those lines. I was sketching out some stuff where I just wrote about relatively mundane industries and the various ways in which AI and automation are going to be transformative and, in really cool and interesting ways, improve people's lives as a result.

There's definitely more of that to do. I think you have to make it concrete. But one of the things I've realized is that when you start talking about that stuff, if you're doing true pie-in-the-sky—“Oh, we're going to have civilization on other planets” and stuff—people will just be like, “Nah.” We've heard that for a long time from techno-optimists.

So I think part of the job here is to make people aware of the astounding reality that is before us today—the actual miracles of modern technology that enable our lives all the time. And I don't just mean AI. I actually mean more broadly than that.

This is something you really uniquely see at the White House because so many people come through where it's like, “Yeah, we manufacture this thing that's an essential part of everything,” right? And it's like, “Wow, this is incredible that this works.” It's incredible that fiber optics work. It's wild. It's completely insane when you actually think about what's going on. I think most people don't.

So I think, to some extent, there are just explanations of our current civilizational infrastructure. I think you can do poetry there if you try. The other thing that goes into that is that you think about the ways in which government itself plays a role in a lot of things.

I was in an airplane a couple of days ago, flying back from a work trip, and just looking around on the plane. I was like, there are technical standards made with government mediation and help for every single thing in this aircraft, right? The chemical content of the sheathing around the electrical wire in this plane has a specific technical standard that someone worked on, right? People maintain it, and there are meetings about this.

You just realize how vast it is. You really do appreciate, specifically in government, how unfathomably large the whole operation is. It continues to astound you when you think about the sophistication of a lot of this stuff that's going on all the time, that our governments just sort of confidently do.

I think we should absolutely want to improve. But I also think that we're in a very negative mood as a country, and we have been for the last 10 years, or maybe more. We only focus on the negative. We only focus on the areas where we're falling down.

But what I consistently saw in government was actually a great deal of competence and skill. So anyway, I think we actually just need to give ourselves a pat on the back sometimes.

Certainly, as a country, we still have an unbelievable amount going for us and should not take that for granted either. How AGI-pilled would you characterize different parts of the federal government as being? Maybe starting with yourself, even though you're recently departed.

I've always been pretty bullish on deep learning. I guess what I would say is that AGI itself is so nebulous as a concept. My view is that, for AGI to be true AGI, it would have to be something that has genuine human sample efficiency and flexibility. I don't know that we're especially close to that.

The analogy that I've always used—and frankly, that I've probably used on this podcast before—is the bird-airplane analogy: human cognition is like a bird. I can fly over to the tree over there and land exactly on that branch, and I can do it with grace and energy efficiency that is really outrageous.

Whereas an airplane is hugely energy-inefficient when compared to the bird. It requires a giant runway, and you have to build all this dedicated infrastructure just for the airplane. It's an extremely unwieldy thing when compared to the bird, and yet highly useful.

People like Dario have been turning a little more negative recently. I think they're actually calibrating to roughly where I feel like I've been for a little while, which is basically the idea that we're probably going to build the cognitive Boeing 737, and that's going to be super useful, but it's not going to be an automated bird, right?

We're not just making a mechanical human brain here. We're going to do something different, at least in the beginning. Maybe eventually we get to AGI, to the more true AGI.

How many people are AGI-pilled within the administration, though?

Yeah, just to be clear, I think we're going to build the Boeing 737 thing soon—somewhere between 2027 and 2030, I believe we will do that. I've maintained high conviction in that. Nothing I've seen changes that. I think the bearishness on GPT-5 is kind of nuts.

I haven't used the model extensively yet, but, yeah, there's a bit of industry commentary and discourse commentary. In terms of within the administration, there's a good number of people who are pretty convinced that AI is going to be super transformative. I don't know how many people have that view, and there's a good number of people who aren't, right? There's a good number of people who are like, “Nah, I think it's going to peter out and be hype,” and whatever else.

It's a little hard for me to describe the ratios because, as you can imagine, the people who I ended up working with probably were disproportionately likely to be other people who think AGI—or AI—is going to be really transformative. I'd say it's there, but what that means is different to different people.

I think it's pretty difficult. I mean, there's a reason that the verb is “feel the AGI,” right? It's because it's an emotional experience. There are a lot of people who I think have experienced feeling the AGI intellectually, but I don't know that they've done it emotionally.

There's a difference between the emotional experience of feeling the AGI. It's a pretty wild thing, and it's happened to me at different phases. There's also a certain aspect of it that's not resignation, but anticipatory nostalgia, where you realize that certain things are just going to go away and certain dynamics are just going to fundamentally change, and you kind of miss those things.

All that being said, I think there are people who have done that, but it means very different things to very different people. But certainly, AI is the president's top technology priority. It is one of the hottest issues inside the administration. It's something that everyone does care about and thinks about.

So I would say that, functionally, the administration places a very high priority on AI, and that comes right down from the president.

So, 3 more little test points, I guess, to try to calibrate my understanding of that better. One, obviously, a White House is going to deal with a ton of different issues all the time.

Is there a dynamic of, well, there’s got to be an AI guy in the room for any given issue because we have to expect that there’s an AI element to everything, or has that not really happened yet? A second one is: are there explicit timeline assumptions built into any of the planning or reasoning? Like, we think there’s a 25% chance Dario is right and we’re going to get the beginning of mass unemployment by 2027, so we’ve got to have one contingency plan for that and maybe another one for if that doesn’t happen.

And a third one: obviously, there was a big bill that allegedly is going to increase the federal debt. I wonder if there are people going around saying things like—which might not—I don’t mean to suggest this would be wrong, by the way, but it certainly would be a leap of faith to say, well, AI will help us grow our way out of the debt so we can sort of afford to take this on because AI will pay the bill for us in time, before we really get into big debt-related issues. So, yeah, comment as you will on—

Right. So, okay, going in order, the first question was not about timelines. It was about—what was it? Remind me what the first one was.

Just like, is there an AI guy in the room?

Oh, yeah, increasingly. I think increasingly, not necessarily always. There are areas where I don’t even know what our nexus would be.

But, yeah, especially post–AI Action Plan, as people throughout the government saw how high a priority it was—the Action Plan event had the president, the vice president, and 5 cabinet secretaries. That’s packed, packed with superstars from the administration. I think when that’s a signal to people that, okay, we really need to be thinking about this seriously, you’d be surprised how important signals like that are within the government, how much that matters institutionally, and more people are understanding that.

Is that entirely true? No. But it feels more and more the case that AI people are being brought into functionally everything, which is tough because there are a good number of them, but there aren’t that many in the grand scheme. They end up being stretched, I think, a lot of the time. I certainly felt that.

In terms of timelines, the convenient thing about timelines is that there is one timeline that we’re quite certain of: the president’s term ends on January 20, 2029. One of my absolute favorite things about working in the Trump administration is the sense of urgency that everybody has. It’s the combination of a sense of urgency and a feeling of not being beholden to the past or traditional ways of doing things.

We are, I think, really developing a new conceptual lexicon for American statecraft. A lot of it we’re doing, and I think the administration doesn’t always, frankly, do the best job of communicating some of that stuff. I also think that they often do a very good job of communicating, but it gets misinterpreted.

There are a lot of people who just ignore a lot of the great work and focus on whatever the latest scandal is. So, I think Americans are quite badly apprised of what the Trump administration is actually doing. That timeline of early 2029 happens to line up with lots of AI timelines.

When we were thinking about the data center power issue, the default timeline that you give is somewhere between 2028 and 2030, right? That’s the, “Yeah, we need to do that” timeline. That just happens to line up really well. So, I don’t know. It’s hard for me to say how people’s timelines would be different if that similarity didn’t happen to be there. I don’t know. That’s a good question.

The debt thing, I can’t say I ever heard anyone say that. I’ve certainly thought it myself, but I can’t say I’ve ever heard anyone else say that. Definitely another thing I have heard is, on some of the immigration debates, you will sometimes hear people say, well, there seems to be an inconsistency here. The pro-immigration camp says we need lots and lots of people to grow the economy and do all this stuff, but then, wait, are we about to automate a bunch of stuff? Are we about to massively improve labor productivity?

I would agree that there actually is a dissonance in that argument. So, yeah, there definitely are ways in which some of the flagship issues from the president’s campaign and the party’s platform now have AI being inserted into them in different ways.

There are a million ways in which border security is an AI problem. There’s a lot of stuff that AI can do on things like border security. You’ve had a guest on this podcast, a former Department of Homeland Security employee named Michael Boyce. He worked on AI at DHS. That’s border enforcement; a lot of what they do is border enforcement.

Is there any talk about a sort of—maybe a step short of utopia, but any sort of new deal for the American worker? Some parts of the Action Plan go this way, but Bernie Sanders has recently said we should make it our explicit policy to share the productivity gains from AI with workers, by, for example, having a 4-day workweek or whatever. Is there any traction with those kinds of ideas?

No, but I wouldn’t say either way. I don’t know that—certainly, the administration takes a worker-first priority with AI, and that is very serious. We do think about that a lot.

I think it’s hard to know exactly what that means at this stage. The administration takes an extremely worker-centric approach to AI. I think we don’t know exactly what the future of labor is like, and we don’t know how acute the labor market disruptions are going to be.

We don’t know where they will take place, or if they’re going to be distributed by occupation type, industry, skill level, or experience level. We don’t quite know exactly. There’s some data that starts to look a little worrisome for software engineering, but there’s also plenty of data on the other side of that.

I think we don’t entirely know where that’s going, but as a general matter, we will—I personally think about a 4-day workweek all the time. The 5-day workweek comes from the last Industrial Revolution, in the 1920s, during Calvin Coolidge’s presidency, and was the fruit of multiple technological revolutions that happened at the same time.

I think that’s kind of what’s happening for us right now, too. So, it would not surprise me if, at some point in the future, we did actually go to a 4-day workweek. That’s an idea of Bernie’s that I could see the world going toward. I wouldn’t say I endorse that idea, but I could absolutely see that idea making sense.

Interesting. Yeah, that’s cool. I think about drivers as one really mundane example, but the analysis there seems so simple in many ways. We already have statistics that show Waymo is a lot safer than human drivers.

We are starting to see, in recent Boston local politics, some of these things that have been slow to materialize from my perspective relative to my expectations. Nevertheless, we’re starting to see this protectionism, where it’s like, well, wait, what do the Teamsters have to say about this?

Then you’ve got people pushing back with, so the Teamsters say that thousands of people should die because we need to protect their jobs? That’s harsh, but not fundamentally inaccurate. It seems like that is going to be a question that society is going to face that doesn’t have nearly as much nuance as a lot of the other things, because it’s literally just: is AI going to drive the car, or is the human going to drive the car?

Is a human going to be required to sit there, even if the AI is driving the car? These are seemingly relatively simple questions by comparison to a lot of what we’re going to have, and we’re just starting to see the political battle lines being drawn. It seems like that’s happening right now, again later than I would have expected.

Yeah. I mean, to some extent, though, there has been a lot of state action. There were AV fights at the state level 10 years ago. A lot of that predates the current AI, because everyone said self-driving was going to be solved 10 years ago.

Mm-hmm.

A lot of that actually started preemptively. That is a really good example, and I think there’s a lot of non-obvious stuff that happens there.

I think you are right that, on the software side, it’s much easier to make the case that you’ll see traditional augmentation, people will become more productive, and, yes, there will be some labor market disruption, but in general things will grow. Whereas, yes, the self-driving thing is just kind of an ambiguous replacement.

The question is, will that generate new kinds of jobs of some sort, though? I think the answer to that is maybe. Automated logistics and the ability to move through and navigate the world in autonomous ways could—I could see it. I think the self-driving car is such a simple early example of where we could be going.

That could create all kinds of interesting new opportunities. But I do agree that it’s more of a one-to-one replacement.

The other issue that I think you’re getting at, which is very important, is the state, local, and federal division of labor here, right? Should we have federal autonomous vehicle rules, or should we not? On the one hand, it’s a little crazy. If I take a Waymo from my house in Connecticut into Manhattan, it would be a little crazy if, during that time period, I passed through somewhere between 2 and 10 different regulatory jurisdictions governing the safety of self-driving cars.

That would be a little weird. Maybe that’s fine. Maybe that’s just an aspect of the future that’s weird, and it’s fine.

Uber kind of had to deal with that.

Yeah, Uber does. Uber has to deal with that right now. Exactly. There are all sorts of ways to make policy issues like that coherent and solvable. It’s not that big of a deal.

But then there are unambiguously things that should be federal. Maybe it is up to a city to decide how they want to deal with AVs. I kind of personally have that instinct. I would rather let that be an area where we experiment at different levels of government rather than just try to occupy the field with one federal standard.

The other thing I would say that becomes really important there is that, when you think about autonomy in the physical world—especially in the physical world, and maybe also in the digital world—one of the things that happens with full autonomy, at Waymo’s level, where no human is really in the loop at all, is that it changes the nature of liability.

All of a sudden, Waymo is held to a very high standard, right? If Waymo is successful, or if Tesla is successful, assuming we really get rid of the steering wheel in the future, that’s going to be a world where the companies that operate those vehicles will have the liability risk for all accidents in the country on their balance sheet. Right now, that liability risk is on my balance sheet.

What that means is that—I think that maybe is a problem, maybe not. It would probably vary by industry as to whether or not that’s a problem. I’ve made this criticism before about AGI: that’s going to be nuts, an insane amount of risk to insure without a really, really well-structured liability system.

That’s actually probably a good thing, if we’re being honest, especially in the physical world and for these labor-market things. In order for one firm to be able to internalize the negative externalities of 5 million car rides a day—or whatever, a million car rides a day—those cars are going to have to be really damn safe. You’re going to need a lot of nines of reliability to really make it happen at nationwide scale.

That will inherently be slow. It’ll take a long time to do. I also kind of think the existing liability system is fine for that. Even though it’s not the most accelerationist thing—the most accelerationist thing would be to say, “The technology is important, so we should give them a liability shield”—I don’t personally think that. I would be very surprised if the Republican Party goes in that direction too, at least under its current leadership.

Yeah, I think they should have to earn it too. I agree that “put up or shut up” when it comes to safety statistics is kind of my attitude, and it would seem like they are well on their way to proving it.

Yeah, autonomy is just—if we’re talking about real autonomy, then I think we should have pretty high expectations of the companies that are doing that. I think that’s probably true.

I don’t think you actually need a lot more regulation than that, other than having something governing the testing regime. If it’s a taxi service, you have to have licensing of some sort. But I think that’s fine.

Beyond that, I don’t feel like you need a regulator to say, “Self-driving cars have to be safe.” I feel like that’s baked into the laws of America already.

If there’s one trademark of The Cognitive Revolution, it’s that we take our time to get to the headlines. We’ve done that here today. But the headline, of course—and we’ve alluded to it many times—is the AI Action Plan.

This seems to have been, perhaps with no exceptions, the best-received thing that the administration has done so far. Everything else, I feel like there are haters coming at it from all directions, rightly or wrongly. We’ll leave that aside, because this is an AI venue, not a general-politics venue.

But the AI Action Plan seemed to get remarkably positive reviews from just about every corner, including from people who I think expected to hate it. I think that is a real feather in your cap as somebody who—I don’t know if you would want to sign on to “led the effort,” but you certainly played a central role in making that happen.

Instead of me walking through it, why don’t you just tell us how you think about the Action Plan? What’s the story that you would tell about what it is and what it’s trying to do? You could weave in any of your experience. I saw a number of funny tweets along the way where you were like, “Final version, V2, revised final, final, whatever.”

I’m sure there were some funny stories there, but take it from the high level first and give us Dean’s view. How do you think about it now that it’s been out for a few weeks and it’s kind of in the rearview mirror? What’s the real headline of the AI Action Plan?

The object-level description is that we decided we wanted to write something that was not a nebulous strategy document. We decided, “Let’s do a concrete to-do list for the federal government for a lot of different things that we can do to advance the ball on AI.”

It’s not necessarily everything we could or should do, and it’s not the long-term answer to any of the burning questions that people have about AI. We can’t answer those questions, and I don’t think it’s a good use of our time to try to answer these unanswerable things or to pretend to Americans that we can.

What was very important to us is that we wanted to deliver something that we can credibly execute for the American people. That’s the heart of it.

From there, one way to think about it is that, if you were to look at an outline of the document, there are 3 big pillars. Within that, there are maybe a dozen headers. Then there’s a paragraph of text, and below that there’s recommended policy action. There are somewhere between 1 and 10 bullet points with recommended policy actions.

America’s AI strategy is kind of the headers and the text below them, right? Those are our strategic objectives. There are many ways in which agencies, and also people outside of government, can help advance those strategic objectives if they would like to.

Then there’s a list below that that says, “Here are 5 bullet points. Here’s what we’re going to do right now.” But just so you know, more broadly, this is a big priority. This is the strategic objective.

The Action Plan is kind of the bullet points, and the strategy is kind of the header and the paragraph or so of text below the header.

One other thing that was never textual in the Action Plan—it’s subtextual, but I alluded to it earlier—is this basic idea that America can do this. We can mature our institutions to deal with this problem. We can adapt, we can evolve, and we can absolutely lead the way. We can do it in such a way that we don’t have to be at each other’s throats.

We can identify lots and lots of win-win things here. I’d say the Action Plan is a deeply positive-sum document that comes out of a city that is usually quite zero-sum. That is the subtextual message all its own.

Yeah, that’s interesting. Very interesting framing.

The 3 pillars are “Accelerate AI Innovation,” “Build American AI Infrastructure,” and “Lead in International AI Diplomacy and Security.” Let’s maybe spend a minute on each one. I don’t want to go literally point by point through the whole thing. People can obviously read it, and they should. They should read Zvi’s rundown of it as well.

Innovation seems to be proceeding pretty quickly. How much of this is stuff that’s going to happen anyway? How much of it is—what do you think are the real pivot points, right? Anybody would look back at the last few years and say, “Yikes, AI innovation is happening fast.”

If you listen to the people among the frontier developers, they're telling us we should continue to expect it to be quite fast. We've got IMO gold medals to prove that, even though that hasn't quite hit the GPT-5 product surface just yet.

If we didn't do any of this stuff, would we lose, or would we slow down? What is the rate-limiting factor here that you're alleviating, or what are the bottlenecks?

Yeah. I would say the innovation section of the action plan is obviously number one. It's the first one, and it's the thing that we think is very, very, very important. It is true that, at least today—this could change 6 months from now—there are not that many laws on the books in the United States that govern the development of frontier AI systems.

Right now, the innovation there is proceeding at a fast pace. But where I think we need to be more reflective and self-critical as a country is transformative adoption of AI. A word that runs throughout the action plan—a theme that's embedded in every single item, really—is adoption.

What I want is flying cars, automated agriculture producing an abundance of food, nuclear fusion, and the commerce, with tons of agents bidding on everything around me all the time, and super-hypermarkets. That stuff is cool to me, and I think those are the areas that are going to change the world.

When people ask, "Whose model is going to set the global standard?" the people who are going to win and set the global standard are the people who find product-market fit. This isn't about sitting in rooms in the White House, sitting around a table, saying, "What should the standard be?" and then writing a standard and trying to convince other countries to adopt it. People do that, but that's not the path to victory in technology.

The path to victory is people emulating your use cases and using the tools you build because they're useful. There's a lot of adoption where I think we're doing pretty well, and I think the American AI ecosystem is maturing in a lot of really impressive ways. The relevant margin for me is adoption.

One of the subtle things in the plan is that we talk about the federal government doing an RFI—OSTP, which is my old office, doing a request for information from the public—for regulations that impede AI adoption. That's subtle because we're not talking about regulations on AI, for the most part. Maybe there are things that agencies have done that we should look at, and there are some, but what I'm also thinking about there is whether there are laws or regulations that have assumptions built into them that are going to be made outdated by AI and associated technologies.

A great example of this is surveying construction sites. There are often state, local, and federal laws that require you to survey a construction site and check for whatever you're supposed to check for, and it's often written so that a human being has to do it. If a human being has to personally do it, as opposed to asking, "What if we just had continuous monitoring of sites through drones with LLMs built into them that have contextual understanding of what they're looking at?" it turns out that's actually illegal.

That would be a great example, and there are thousands of things like that. That's really what we're thinking about. The other thing that I would say is really important is that there are some areas, like science, where the U.S. federal government has pretty high leverage over those institutions, and so we can specifically drive things there.

One example that I've been on about for a long, long time is automated experimentation. You saw the National Science Foundation announce its Programmable Cloud Labs initiative, which is going to be $100 million to different companies, academics, and so on that are building automated labs for massively scaled scientific experimentation.

A few years from now, we could be in a world where there is automated science infrastructure that can be used as a cloud service, basically, by AI agents. Obviously, there are significant safety issues with that, but I think the way to think about this is that it might not be something that actually exists. It might not happen on its own.

Automated labs exist today, and they're inside corporate R&D labs, but it might not be shared infrastructure. That's very much like how the federal government led the way in the early days, before there was a commercial application of high-performance computing facilities. There wasn't one, and the federal government viewed it as a public good, as common scientific infrastructure, so we built that. We also built the internet because we needed to network those facilities together.

I think there are things like that that are very exciting. The idea of the NSF initiative is to build a network. That's one very specific thing, but there are a lot of things like that throughout the plan.

There are a couple of other things that are about trust and reliability of the technology. If what people know about the technology is that it's the thing that makes it impossible to get justice in court anymore because you can't validate media, that would really be awful. It just so happens that there are levers we can pull inside the federal government on that now, so let's do it.

That's a good example of something that's actually a bit of a risk-management thing. That's more of a risk-type thing, but it's in the innovation section for a reason.

What do you think is more promising to address that? Is it encoding the origin into synthetic media, or is it on-device attestation for real cameras? Maybe it's both, but how do we get out of that problem?

Plausibly, it's both. But I think what's actually going to work is that you've got to focus policy effort on the scarce thing, which will be the human. Well, I don't want to say human-created. What I want to say is that the scarce thing will be actual photons that hit actual glass in the world, processed by real image sensors. AI-generated content will be the super-abundant thing.

In the long term, what we're going to have to do is have some sort of common standard for validating real-world stuff. I don't think it's that bad yet, for the most part, because even today—even with Veo 3—it's still different enough from the real world that, with the scrutiny of a legal system, there are still things you can do.

But it's an area where we need to make sure we're actually doing those things.

On the diffusion point, I wonder how you see that, or how you think the administration broadly sees it. I've been struck by 2 contrasting viewpoints. One is that, as I'm sure you're aware, Jeff Ding has popularized the idea that the U.S. leads in terms of our ability to diffuse technology through society, get it to a broad base, and get the practical value from it. He argues that China lags in that capability.

But I also see a line in the action plan: "Enable the adoption of AI in the Department of Defense." In an episode out today, Jake Sullivan also talked about the memorandum that he and his team put out, which was basically arguing that our national security establishment broadly—or maybe the military—is less agile than our Chinese counterparts when it comes to adopting new technologies.

Would you say that's your view as well? We're better in the private sector, and we're slower in the government? Or would you complicate or contradict that analysis?

I think it depends. One thing that America is quite good at, compared to everyone else in the world, including China, is the pipeline from deep capital markets to cloud computing, from cloud computing applications to consumer and enterprise adoption of new technology. We're pretty darn good at that.

China does not have a lot of the much-bemoaned B2B SaaS. China does not have a ton of that, and a lot of it is actually pretty useful. It ends up being quite important in the sort of cybernetics of the business organization, so it matters a great deal. I think we should be proud of the fact that we do well in things like this.

There are other areas where, in terms of AI adoption in the military specifically, I can tell you that our military is not as agile as I would like it to be.

I have not carefully analyzed the difference between China’s military adoption and our own. I think one thing that is always true in analyses of China is that a lot of its governance is very KPI-driven, and so there will just be directives from on high. DeepSeek comes out, right, and every bureaucrat in every part of the country will, in a day or a week later or something, have to check a box: Did you use DeepSeek for something? Are you adopting AI?

What that ends up with is summary statistics that would suggest that China is adopting AI more quickly in its government than the United States is. But a lot of that adoption is pretty shallow. I think we need to move more nimbly, for sure, but I’m always aware that that’s an issue when you’re analyzing China: they will often optimize for hitting specific numerical targets that I’m not sure are actually connected to the thing that you want.

What you really want is deeper adoption. Part of what happens is that we drive deeper adoption of AI, and then that reveals new problems and new subtleties, and that creates a positive feedback loop with product development and finding product-market fit. You’re starting to see that, right? You would not have guessed the form factor of Claude Code. It’s a very interesting form factor, actually.

You would not necessarily have guessed that you would be using CLI-based tools to do some of the most cutting-edge AI automation 3 years ago, when you first saw ChatGPT. I don’t think that many people would have guessed that. And yet here we are, right? So you’re starting to see all that happen. You’ll see much more of that kind of stuff happen as we figure it out.

But I still think a market-based process is going to do a better job at diffusion if we don’t get in the way of it with bad regulation. Because what China will do is, if there is some transformative use that’s blocked by a regulation that they care about, they will unblock that. Beijing will put out a directive: It’s unblocked by regulation, and it will happen everywhere, right?

So that’s where they’re just better at cybernetics than our government is, because we’re not a centralized state. We’re not a centralized authoritarian state. Fundamentally, some people think that’s what we’re turning into. I think those people are dead wrong. But I will tell you, as someone who served in the Trump administration, it did not feel like I had authoritarian levels of power over anything. So, yeah.

What I mean with the military in particular is that it seems very fraught. It seems like there’s kind of a major mismatch between the way the military has traditionally thought about things and maybe should think about things versus the way that LLMs work. My standard refrain on that is: I would want to know that any issues of deception or scheming against the user are fully resolved before I go into combat with my AI battle buddy.

I know we’re maybe not immediately jumping to AI battle buddies, but if it’s just making paperwork more efficient at the DoD, that doesn’t seem like the kind of change that’s going to beat China. Not that I’m, as you well know, focused on doing everything to beat China, but if it is actual combat-operations applications, it seems like we don’t quite have the right AI for that right now.

I think one thing about military history is that the flow of information throughout military organizations ends up very often being quite a decisive advantage that militaries have. There’s a great history—I read once, or sort of an analytical book called Command in War, I think it was called—that was about how different communications technologies ended up changing the nature of warfare.

There was stuff on the internet, but there was also stuff on the adoption of the radio inside militaries and how that completely changed things. The radio was the first thing that allowed for truly centralized command of militaries, which changed everything that was possible in warfare.

So I actually would dispute somewhat the idea that a pure information-processing technology is not a quite important military advantage. Even if you’re literally just talking about GPT-5, today’s technology is capable, in particular, of synthesizing the staggering amount of information that our government collects on the world every day. A lot of it is from intelligence agencies and deeply, deeply secret stuff.

But we pick up a lot of information on the world, and the ability to quickly analyze that and make decisions, or present decisions to humans for a final call, could be quite decisive. That’s basically just a pretty traditional LLM adoption story.

I mean, it’s more things than that, too. It has to do with information sharing also, and that’s quite difficult. When it comes to the weapons side, I think once you get into the physical world, the DoD starts to become much more equipped. The DoD is good at being like, “All right, these are the performance characteristics that we need before we can adopt this thing.”

But as you’ll see, first of all, on the DoD adoption side of things, the action plan has a section about a physical facility for testing autonomous technologies. That will be a physical thing that the DoD will use to be able to write those specs out. But I have faith that their culture is institutionally well suited to that if they have the right tools at their disposal.

But another thing, of course, is interpretability and control—and you might call control alignment, right? This is an area of deep importance for the military, particularly. I mean, when you think about the fact—who knows how an LLM, how this would work in a system deployed at DoD—but when I ask an LLM questions, at this point it is situationally aware that it is talking to a person who worked at the White House on AI policy, and I’m asking it a question about that that has a plausible connection to AI policy.

How is that affecting the model’s outputs, if at all? It’s like, whoa, that seems important to understand really well, and I don’t think we have a very good understanding of that right now. Hence why I think and expect that DARPA will place quite significant investment into those exact issues.

Yeah, I’d say little to no understanding of how that’s impacting things just yet. But it is definitely a fascinating question. That’s a good point about information-processing speed being an important dimension of competition. Obviously, it’s not the only one, but I think it’s a compelling point that that could matter even if there’s no actual firing of weapons by an LLM at any point.

Yeah. Most military planners that I have talked to think that it’s basically logistics and cybernetics, right? It’s moving things about the world physically, and then it’s moving information through. If you do that, yes, we focus on the tip of the spear, which is the hypersonic weapon or the autonomous drone or whatever else.

That stuff is important. We need to be able to make sure we can do that stuff. We need to keep up, et cetera, et cetera. But I think if you bolt that stuff onto a 20th-century military model, even if we have the world’s best hypersonics and make them abundantly available, and we have a super-big army of drones, if those things are being commanded by a military organization that is rooted in Industrial Revolution-era technologies, then I don’t think we will fight successfully.

So broadly, I guess, if part one is: don’t get in our own way and let our private sector continue to lead in a relatively unencumbered way, and do what it does best, and we have advantages and we’ll naturally maintain those advantages because that’s who we are. Part two, around building infrastructure, is the part where it seems certainly much more plausible that we need to actually up our game to achieve the visions that we have of an intelligence-too-cheap-to-meter future.

How optimistic are you, I guess, in the first place, that we’re actually going to be able to do this? There has been quite a long time since we were energetic, so to speak, about building nuclear energy or just building infrastructure fast—almost at all, right? A few exceptions, I guess, but not too many.

Is this something you think the administration is really going to be able to change in a short period of time, such that we’re actually bringing all these things online in whatever 2027-to-2030 timeframe?

Yeah, the short answer is yes.

I don't know that we will have lots and lots of new nuclear reactors producing gigawatts of power in the next 3 years. That would seem hard. I think we're doing some really significant stuff. On nuclear, there are a series of executive orders, driven in part by our office, that came out a couple of months ago, and there's an entire reorganization of the Nuclear Regulatory Commission going on. So, exciting stuff to be sure.

I think fusion is plausible. I'm way more bullish on fusion than most people in the government. That's one thing I can definitely say: most people in the government got very mad that fusion was even mentioned in the action plan. People were like, “Wow, it's never going to happen. It's 30 years away.” I think that's just an older point of view that comes from not being engaged with the current frontier of that technology.

When you look at the next couple of years, I think we're going to do it. I think it's also going to look a little different from what you might expect. I always remember Leopold Aschenbrenner's Situational Awareness.

He's a first name. It's all good.

Yeah, it has this AI-generated image in Situational Awareness that's an endless field of data centers. You see big natural-gas peaker plants in the distance. We're going to build stuff like that. I think there will be things like that by the end of the decade. I totally do.

But the idea of building 100-gigawatt or terawatt-type facilities—I don't think it's going to be necessary. I think the way we get there will be through different kinds of unlocks that are important but not as well discussed. One thing is that the American electricity grid, as a general matter, is actually quite overprovisioned because the grid is designed for the worst-case scenario. It's designed for the day when it's 112° in Texas, at peak time, when everyone's ACs, TVs, and electric cars are on.

It's designed for those moments of peak demand, which means that the vast majority of the year there are actually many gigawatts available to be used, assuming you don't need the power 100% of the time. There's a viral report that came out of Duke from a guy named Tyler Norris that basically said, if data center operators were willing to curtail their electricity demand by 25% for 0.25% of the year, you could unlock 76 gigawatts just from that, without building any new physical infrastructure, purely through what's called demand response.

The problem, though, is that—I'm getting really technocratic here—there's an entity called FERC, the Federal Energy Regulatory Commission. They regulate the interconnections that happen. If I build a new data center, and it's a 1-gigawatt data center, I'm adding 1 gigawatt of demand to the grid. When I'm trying to build something of that size, the utility in a state will do an interconnection study and computational modeling of my demand to see what it's going to do.

When they do that modeling, they assume completely stable growth. They assume it's 1 gigawatt 100% of the time. When you do that and also factor in the worst-case day, it's like, “Oh, we're going to have to build an entire new natural-gas plant, and we're going to have to build totally new transmission infrastructure to accommodate this.” Then that becomes a delay and a cost, and the data center operator is going to have to bear that cost. The utility will build it, and the data center operator bears the cost.

There's a federal entity—although a lot of this is being done at the state level—called FERC. When those interconnections implicate interstate electricity-transmission lines, which is a lot of them, FERC has jurisdiction. What FERC can say is, “Don't model it as stable demand.” If a data center is coming online and says, “We're willing to curtail our demand for 0.5% of the year,” then we're going to give it a faster interconnection. We're going to move up its time to power. We're going to get it on the grid in 2 years rather than 5. You've unlocked a ton of energy, and you've accelerated demand.

The other amazing thing about this is that there's hardware for demand response, but there's also software. When you think about what demand response is—I'm a data center, and I just got a signal from the utility that I need to cut my power by 50% in the next 10 minutes—you have all this equipment running. You have the GPUs, the cooling equipment, and all this other stuff. You have different-priority batch jobs. There are all these different workloads being processed by the data center: one is me making a cat meme, and the other is a patient's medical records in the same facility.

All of a sudden, that starts to smell a lot like a reinforcement learning problem, doesn't it? It starts to smell like an AI problem: We need to optimize how to scale down our power dynamically. Anyway, there are a lot of different things you can do there, and that alone, I think, can unlock 100 gigawatts if we do a good job of it. In addition to that, we're going to make permitting easier, and even if Congress does nothing, at the margin that will be useful. There's a lot of energy around state permitting reform right now, too, that folks are engaging with, and at the margin we'll be better.

Are we going to have 100-gigawatt data centers by the end of the decade, or lots and lots of nuclear power plants? Probably not. But I think that by mid-decade it'll be pretty crazy. By the mid-2030s, I believe there will be lots of new nuclear online.

I had not heard about this dynamic management, but I did have a personal experience, maybe 6 or 7 years ago now, in Detroit. One cold night in the winter, we suddenly got a text from our local utility that was like, “We've got some sort of problem. Please turn your heat down to 65, everyone.”

That's only happened once in my life. Talk about things that we can be grateful for and shouldn't take for granted: the uninterrupted provision of such utilities for my entire life is definitely one. But basically, that worked. People responded to the text, and we made it through.

Obviously, it doesn't scale to send texts to people, so you do need to automate this in some way. But one of the things that's amazing is, if you play the tape forward on that and roll out technology of that kind to lots of industrial facilities, including data centers, the net effect would be that we would utilize existing electricity generation more efficiently than we currently do, which would actually lower prices over time.

So you literally can build the data centers while lowering prices if you do it right. Again, total theme of the action plan: win-win. Positive stuff.

That's great. Taking one step back, I don't know quite how to evaluate this, but it was reported in Defense One—a quote from an anonymous Pentagon official:

“We're not going to be investing in artificial intelligence because I don't know what that means. We're going to invest in autonomous killer robots. This administration cares about weapon systems and business systems, not technologies.”

Would you call that fake news? I mean, there's certainly some fake news out there, but how would you reconcile that with your characterization of your conversations with military players?

I would say one thing: when I said the conversations with military players, I was not necessarily referring to people inside the administration, inside the Department of Defense. Over the years, I've heard podcasts and talked to various people who know about this stuff. So that's not so much a Trump administration view.

That quote gets at weapon systems and business systems. The latter, business systems, would be the information and communication stuff. Certainly, there are also totally transformative hardware things that we'll be able to do with AI, and I don't even know how to think about the form factor for those things. I think drones are an early one, but there's going to be so much more.

I'm very excited about autonomous boats and autonomous ships. America has problems with shipbuilding, which I hope we get better at. I think that's potentially—

I think, again, this is an area where this administration is doing more than anyone has done on the shipbuilding problem. But we're actually good at boat building. We're perfectly competent at boat building. If you build lots and lots of autonomous boats, that might be a really interesting way to think about the future of naval warfare—basically, a bunch of school buses flying around the ocean.

To be clear, I was not discounting the benefit of weapons systems—hardware weapon systems—that are AI-enabled in some way or another. I was just saying that the Department of Defense will be better equipped to buy that kind of stuff, particularly because they're getting significantly better at working with startups. You're already starting to see this blossoming ecosystem of defense-tech companies, and I hope that stays, that it continues, and that it's operating as a new neo-prime, selling all kinds of AI-enabled hardware capabilities to the government.

It's already happening, but I would also say: What is the actual basis of Anduril? What's the real fundament of their business model? It's software. It's the Lattice software system, which is exactly about information sharing, so that when we build new hardware, we have a software platform. They're like Apple: as they add new things, it all fits into the Anduril ecosystem, and everything can talk to one another and communicate at really high bandwidth. That would be a good example of exactly what I was talking about.

Gotcha. Okay, interesting. When it comes to bringing chip manufacturing to the United States, that's another big challenge. I guess there are a couple of different angles I want to come at it from. One is these Gulf deals. I'm interested in your take on why we're doing those deals. The answer I always get—which hasn't quite satisfied me, to be honest—is that we want them to build on our AI stack versus China's AI stack.

But then I also feel like, does China really have any chips to sell them? Couldn't we have, if we're concerned about American values, held off on planting these giant data centers in these Gulf countries, which, frankly, simply don't really share American values? That doesn't seem like the obvious move. It seems like we could have held off on that a little longer because they didn't really have anywhere else to go. But maybe they're doing something for us that I don't understand. It might be energy, it might just be the speed of regulatory approval locally, or maybe it's cash on the balance sheet.

And then the other angle, of course, is Taiwan. I wanted to get your take on how you think they're thinking about this right now. They have this tricky position where, obviously, they're right in China's shadow. Everybody knows that's a flash point. They've managed to put themselves in this position where they're super relevant. If Taiwan goes dark, from a Western and United States perspective, that's a huge problem, and so we're at least compelled to be ambiguous about exactly what we want to do, or would do, under various scenarios.

But now they also have this tricky situation where they need to be a good friend to us to keep that dynamic going. That means sharing some technology and putting some TSMC know-how on American soil, but they probably don't want to overdo that, right? They don't want to share everything, because then maybe we wouldn't need to defend them as much. It seems like the Gulf states are trying to engineer their way into a similar position. I don't know—break down the various geopolitical strategies that folks are playing.

A couple of foundational things go into the UAE, and I should say that I worked quite carefully and closely on the negotiations for that. First of all, I think, much like the action plan, the UAE framework that we agreed to is very positive-sum. What we're saying is that this is not rivalrous: We are going to do a big industrial build-out in the United States, it's going to generate a lot of jobs and a lot of wealth, and it's going to be an asset that people all over the world, and especially Americans, are going to use.

But we don't see that as rivalrous with the idea of other countries that are strategic partners doing really ambitious things, too. What makes the UAE special among countries around the world? You asked about AGI-pilled governments. The UAE has an AGI-pilled government. They're the most AGI-pilled country in the world. They think about this technology in very sophisticated ways.

I do think it's very important that America be partnered with other countries that are sophisticated and have a pretty good grasp on the likely trajectory of this technology. I think it's really good if all of us—leading countries that are intellectually sophisticated—are partnered. We don't want them building on the Chinese stack.

I think there are probably different estimates and projections about where China's chip sector is going to go in the future and how quickly they'll be able to catch up. As a general matter, it seems like the trend has usually been that they catch up a little faster than the technology analysts here guess. I don't have a deeply technically principled answer to that question.

But what I do know, as a matter of policy planning—which is what we were engaged in with this deal—is that you probably can't assume that they are going to be slow. That feels like a weird hinge point to be complacent about. Frankly, a lot of the people I know who are critics, and I think a lot of the people in the prior administration, were weirdly complacent about that one thing. They were so enamored by the specific thing of extreme ultraviolet lithography that they said, “They'll never be able to figure it out.”

I don't know. I'm not sure about that. It's hard; it's definitely hard. There are other ways to do it, right? It's not the only plausible way that you can achieve that small a size on silicon wafers. There are all kinds of different things you could do. There's also scale, and there's also the fact that they'll just eat the profit margins, right? They'll just make money.

Yeah, yeah, yeah. That's all they eventually do, right? I think Noah Smith might have said this, maybe on your podcast even, but the world where they win is just such a gray world. It's so dreary, because what does that mean? No profits means no new stuff. It means nothing new happens because no one can ever reinvest.

Great, you've made an ultra-mega-superabundance of chips. It's a dystopian hellscape that they're trying to build. At this point, I diverge.

That seems a little harsh. For what it's worth, I have to at least briefly comment that what I see of China doesn't look like a dystopian hellscape. Even if you're a political dissenter, it might quickly become one. But for most people, it seems like life is getting better.

If China occupies the role that we currently occupy—if they're the world's true frontier economy and the biggest global powerhouse, and we're a significantly less relevant country—that would suggest that technology development becomes much harder. If you have no profit, you have no ability to reinvest into the business and make new things.

They just take everyone else's stuff, make it super low-margin, drive them out of business, and then make a superabundance of that stuff. In a certain sense, sure, that's fine. But the long-term result of that is a less innovative world. If that's actually the strategy, it's a less innovative world and a less colorful world.

I do think we see some exceptions to that, though, right? Huawei is not doing that. They're notorious for reinvesting huge amounts, and my sense is that they are at the frontier, if not genuinely pushing the frontier, in their domain, right?

Kind of, yeah. I mean, I don't know. I think Huawei is an interesting case, and there's only so much I can say, but Huawei is a somewhat different case. You're right; it's not uniformly true. I just mean that if you play the tape forward, it's a less beautiful world.

In any case, where were we?

Yeah. So this is all downstream of why we had to sell these huge data centers to the Gulf countries.

So, yeah, I think we do want them to be on our technology stack. I think that's 100% true. I think they're a valuable strategic partner. And, of course, yes, it's also the case that we think there are terms that we secured with the UAE's government that require them to make reciprocal investments of similar size to the data centers that get built over there.

And those investments could take the form of data centers, but they could also take the form of investments into energy infrastructure—all kinds of things associated with the AI buildout. You’re talking there about hundreds of billions of dollars that they are eager to invest into our country. And so I think we have to get the security details right. That’s what’s going on right now. That’s what my former colleagues at the Department of Commerce are doing: getting the security details right. But if we do it, it’s just a total win-win.

Do they have a different government and different values than we do? They totally do, absolutely. But I don’t know. I think the idea that we can only engage in commerce with purely democratic countries is just a weird rule. It’s never been true before. It’s not like we’ve only done that before. And we would love to sell to other democratic countries.

The reality is that a lot of the developed democratic countries don’t really like AI very much, and they spend more of their time talking about how to put a straightjacket on it than they do talking about how to grow it. I hope they change their tune on that. They are going to be big customers. I think they will regret the fact that they’re not currently big customers, and I think they will regret it in the relatively near future. But you did just see Norway do a big OpenAI deal for the Stargate for Countries thing. I hope we do much more of that.

We want to sell to lots of people. Our export promotion program is about treating countries equally unless they are strategic adversaries of the United States. We want to engage in commerce. This is about commerce. So, yeah, I think there’s a lot of win-win to be had.

Yeah, I do support commerce and trade generally with people that we don’t see eye to eye with on everything. It does seem a little strange coming from the American right, though, which I would say is generally not fond of Islamic values broadly. There’s a weirdness to that that I see kind of being memory-holed. And I guess if I was really forced to pick whose values are more compatible with ours—the Saudi government’s or the Chinese government’s—I think I lean China.

I know they’re a more serious competitor to us. That’s a different question. This is a bit of an aside as well, but in the quest to see if there’s anything stable and robust that we could try to align AIs to that would work out well for us if indeed they become super powerful on a not-super-long timescale, ancestor worship has been one candidate that kind of keeps coming to mind. We’re the ancestors, and that maybe is good for us.

You know, that’s a pretty Chinese-flavored notion.

I recently watched a little TikTok. I know you’re not on TikTok, but there was just a guy in a small town who took us into the ancestral hall in his small town. I think that’s the name that he gave it. It was basically sort of the community center, a place where they have events.

It’s amazing how much of this is still there. I think a lot of it got wiped out at critical moments in Chinese history, but at least in this place, there’s still this testament to these being our ancestors, going back pretty far. And he said that, basically, in this dude’s mind, this is Chinese religion. I was like, geez, relative to other religions, that seems like a pretty decent story to try to get AIs to live with.

I don’t think it’s a solution, but I want to say something in favor of Chinese values, or at least in defense of Chinese values.

You were hitting on some pretty fundamental things there that I actually worked on when I was in college, believe it or not. I became really interested in the resonance between ancient Chinese philosophy and conservatism, various strains of conservatism. I think it’s pretty clear that the concept of emergent order, the concept of emergence, and that idea—which was sort of pioneered in the West by Hayek and complexity science in the 20th century and maybe has somewhat earlier predecessors in the Scottish Enlightenment of the 18th century—is crystallized in a Daoist concept called wu wei, which is thousands of years old in China. There are all kinds of interesting resonances, and there are actually many different paths there.

I would say the extent to which China actually embodies today the values of its ancient intellectual traditions, like Confucianism, Daoism, et cetera, is questionable. Certainly, it’s not a very Daoist country. There are obviously some elements of the Confucian system that have persisted, but I think there is more continuity with other intellectual traditions that are somewhat less wholesome than what you’re describing.

The UAE and Saudi Arabia are definitely very different in terms of the way that their societies are structured and the kinds of values that they embody. I think this is one area in which President Trump is very different from previous presidents, and I think in a good way. I think President Trump genuinely seeks peace. I think, of any president certainly in my lifetime and probably in modern history, he is the most earnestly peace-seeking president we’ve had in a very long time.

The flip side of that coin is that, for all the stuff about trade and everything, he actually cares quite a bit about global commerce. He wants it to happen on terms that are a little more favorable to the United States, but he actually cares a lot about it, and he thinks the connection between commerce and peace is very deep for him. I think he’d love nothing more than to see a lot more sophisticated commerce happening between our countries and within the Middle East region.

It’s a strategic sector. There are definitely areas in which their values differ, and I think we’re going to have to be able to accommodate a pretty wide range of values into our systems in the fullness of time. There is fundamental sovereignty there, and I think we have to be respectful of that.

Again, we’re not opposed to selling AI systems to advanced capitalist democracies that are largely secular places, like Western Europe. We totally do. We sell them lots of stuff, and they’re big AI users. But it’s also the case that, like I said, they’re not the most enthusiastic, and very often they’re kind of hostile. I wish they would stop being that way, but that’s where they are right now in their society. That’s also their choice.

One thing that has struck me is just generally how much continuity there seems to have been between the Trump, Biden, and Trump administrations, at least on the narrow set of issues we’re talking about today. Not necessarily, obviously, all of politics, but generally the managing and muddling through of strategic competition with China seems to be a clear throughline. Export controls seem to be more similar than different. You may see that differently, but from my not-super-deep interrogation, it seems like there’s a lot of continuity there.

There’s the sort of anti-woke notion, but it seems like a lot of people—and I’d be interested to hear your thoughts on this—have sort of said, “Yeah, well, they kind of had to put that in there for rhetoric, but it didn’t seem like their hearts were really in it.”

There’s also, of course, the biosecurity angle, where there seems to be a lot of continuity between Biden executive orders on DNA sequence pre-synthesis screening, and I think you guys have even extended that. How would you characterize the level of continuity, and what do you think are the most important points of divergence?

So, I think there is a lot of stuff that’s pretty discontinuous. I would say the export stuff is one good example of that. Yes, it is true that export controls on China are a thing. To a certain extent, I think it’s actually more appropriate to say that the Biden administration is consistent with the first Trump administration, because the export controls on China, including EUV, were initiated under Trump 45. So I would actually characterize the Biden stuff as an expansion of work that was originally pioneered in the first Trump administration.

But things like the diffusion rule—I can’t tell you how much damage that did to this country internationally. People love to talk about the damage that this administration is doing to our reputation internationally. Fairly or unfairly, we can set that aside. One thing I can tell you for sure is that when foreign governments came to talk to us about compute and AI, diffusion was always at the top of their minds, and they really felt like it was a huge slap in the face.

To put 2 of the largest democracies, by the way—2 of the largest growth markets for generative AI, Brazil and India—into Tier 2, into the “We all like you, we don’t trust you” tier, was such an unnecessary self-own. Getting rid of that and being much more oriented toward export, and at the margin less oriented toward control, is important. It’s not like we want zero control. We have security provisions that we’re going to be very serious about throughout the world, but we’re somewhat less interested in micromanaging the global diffusion of the hardware. We’re significantly less interested in that. That’s a big deal.

The woke stuff—I think we are cognizant, or were cognizant, that we don’t want to meddle in the markets, right? A consistent theme throughout the administration is that we used federal procurement. That’s the hook there. We’re not trying—we’re very explicitly not trying—to tell AI companies what the models that they sell to you in a private market transaction should do.

What we’re saying is that, for our purposes as the federal government—we’re not a very big customer of these LLMs just yet, but we think we probably will be in the future—it matters to us what the political values of the systems are. You can go, if you want, and there is a MAGA handshake, Daniel Kokotajlo thing you can do here.

One thing that’s not inconceivable to me is that, especially if we had the dynamics of Trump 45, where American elite institutions really liked to performatively undermine the administration live, and so did the tech companies. They liked it; it was fun for them to undermine the president of the United States. American companies—that still makes me personally angry, and a lot of other people angry.

It wouldn’t surprise me if this technology was diffusing under those political dynamics. It would not surprise me at all if there were covert efforts to sabotage the LLMs used by the Trump administration in various ways at the margin. We want to be damn sure that’s not happening, so we are actually quite serious about that.

Getting the guidance right will be difficult, but there are some very talented people who are going to be working on that. We don’t want to create a massive regulatory burden with that. That wasn’t the objective we had in mind. What we want is transparency, right?

The executive order says that one way you can comply with this—and I think probably the easiest way to comply with this—will be transparency around system prompts, model specs, constitutions, and testing. It’s going to be stuff like that, which is another area that’s pretty consistent with AI policy and with a lot of non-MAGA, just general AI policy.

I think, is there consistency with the Biden administration in those senses? Very, very deeply not. The Biden administration was totally committed to the idea of using regulation and scary words like “misinformation” and “bias” to impact and politicize the outputs of LLMs all over the place. They were very committed to that idea.

If Kamala had won, that’s the big thing they’d be doing. They would be pressing at the limits of the Constitution to do it in a way that we absolutely did not. It just wouldn’t be called pressing at the limits of the Constitution, because when Democrats press the limits of the Constitution, it’s called ambition, and when Republicans do it, it’s called authoritarianism.

That’s just the rhetorical standards that we have to live with, and we’re going to get that kind of hate. We’re prepared for it. That’s the world. I would say that those things are stark differences.

But you are also right about the way I describe this, because there were people who said this internally. They were like, “Wow, there’s some stuff that’s similar.” And it’s like, we’re in the early stages of a policy field developing here, and you are expressing surprise.

It’s like we’re in the early stages of financial services regulation, and you are expressing surprise that the concept of interest rates is similar between the 2 parties. No, I mean, these are just important abstractions in this field that we must have.

Yes, interpretability is important to us, and it was also important to the Biden folks. That’s true. I think we have a very different posture. Vibes actually do matter a great deal. The Trump administration has a very different posture toward the technology in general.

But there are some technocratic things that are similar. Again, biosecurity is one example. I praised the 2024 Nucleic Acid Synthesis Screening Framework on your podcast, and I’ve played a pretty substantial role in rewriting the Nucleic Acid Synthesis Screening Framework that the administration is embarking on.

I can’t talk too much about where that is because it’s not quite out yet, but it should be out pretty soon. You will see that there are many commonalities, of course, but the Trump administration will strengthen it substantially in some pretty neat ways.

There is some consistency also outside of the action plan. The MP Materials stuff that our Department of Defense pulled off—the Biden people wish they could pull off deals like that.

I think one thing that is really different, and more inside baseball, but matters, is internal culture and energy. The Biden administration was much more by-the-book and procedural, with lots and lots of process governing everything, and that slowed everything down.

We are moving much faster. Like I said, the Trump administration is creating a new lexicon for American statecraft, and we are much less beholden to the past and to the way things have been done in the past.

In terms of what kind of things you would expect from an organization run this way versus the way that the Biden administration ran things, I think you should just expect the Trump administration to move significantly more quickly. Sometimes it will move in directions that you won’t like. Inevitably, that will be the case, but it will move more quickly.

It will try more things. And will it also be true that, because of that, some people will characterize it as, “Oh, that’s chaos”? I don’t choose to see it that way, but I’m sure some people will.

Yeah, there are many, many ways to narrativize the same events. That’s a lesson I learned over and over again. I don’t want to keep you on this podcast until you have your kid, and we’re entering into Rogan territory. Maybe just a couple more things to bring us home.

Taiwan—to recap that question, what are they thinking? How are they trying to walk this tightrope?

So, yeah, I think the Taiwanese are obviously making large investments into the United States, and I think they see geographic diversity as being really important. But they also are a silicon shield for the island.

I’m not the world’s best person to talk about the dynamics of that. That’s not my area of expertise. What I would say is that, from the perspective of building more semiconductor fabs in the United States and indigenizing that, I think we’re making very strong progress.

The Commerce Department, in particular, has been doing some work that I feel like has flown under the radar. They’ve revised a lot of the CHIPS funding deals that the Biden administration had inked with different companies and actually made them more favorable, bigger, and better in a lot of cases.

That’s good, because a lot of those deals—you can make the case that the Biden people put in a bunch of crap. They put in a bunch of requirements like, “You have to have diverse employees,” and all this stuff.

The care centers on-site.

Yeah, yeah, yeah. Commerce took a lot of that stuff out.

But it’s also a strategically intelligent time to revise those deals, because a lot of them were negotiated in 2022 and 2023, when chip companies had way less faith in the AI thesis. Now they’re much more bullish, and everyone’s like, “Oh, my God, we’re going to be underprovisioned.”

So they were actually willing to make bigger investments under better terms for us.

So, that's been great. Obviously, the trade deals have played and will continue to play a really important role in getting more stuff. But I think we are actually at a point in the United States where we are going to have multiple really robust enclaves of different kinds of semiconductor development. You know, in Indiana there will be the HBM stuff, kind of anchored by SK hynix. There will be the hub in Arizona, of course, and Taylor, Texas. Samsung—you know, that facility looks like it finally has enough demand to come online.

There will be other places too, other R&D hubs. It's actually very exciting to see. I think, by the way, the automated materials science that we talked about earlier is a strategically important part of this. There can be automated labs for materials science that are commonly shared between industry and academia. That way, you're taking frontier academic research and doing it at the same shared facility where a corporation might do some frontier research of its own, totally cross-pollinating ideas. I think there's a very, very bright future ahead for the United States. I focus on domestic policy, so I don't, unfortunately, have all that much to say about the Taiwan stuff.

Do you have a sense of what we're aiming for? I know we want to have some core amount of production capacity domestically, but it seems like we're not going to achieve chip independence in the sense that we're not going to be able to stop importing chips from Taiwan anytime soon.

Is there some other threshold between some bare minimum that we need to keep the AI lights on and total self-sufficiency? Are there other thresholds between those 2 levels that you think are particularly important?

Well, I would say a couple of things on that front. First of all, I think we actually are trending by the early 2030s to at least be able to satisfy domestic demand with domestic production, so I feel decent about that. I think one part of the goal is that. Another part of the goal is to actually reclaim the lead in frontier semiconductor manufacturing in the United States. That's very important.

But another thing that is underrated, and I think probably is an area I'm going to focus on, because I don't think the current Biden administration did enough about this and I don't think the Trump administration thus far has done enough about this specific thing, is legacy nodes. You can shut down civilization with 45-nanometer chips. Yes, 2-nanometer, 18A, and 14-nanometer are very important, very cool, and important things. But you can shut down civilization with 45-nanometer chips. We don't have 45-nanometer production—you can stop civilization if we can't import that stuff.

That's a harder one. That's harder than just the leading edge. There is a range of interesting policy tools to use there. I'll have more to say on this soon. This was stuff I was starting to think about post-action plan. I was like, this is one of the things that was on my mind. I'll probably have more to say about that at some point in the next couple of months.

That is a big problem that's a little bit thornier, and it's harder to make a purely economic solution to that. But on the frontier side, I think we're trending in a good direction. There's a lot more work to be done, and I think what we need to do is make sure that we're funding the basic research that allows the United States to make the next-generation leapfrogs. You asked—we need to make sure that we don't miss some leapfrog technology. Usually, our ecosystem is quite good at that, but we need to make sure we continue to nourish it.

Do you have any sort of taxonomy or short list of things that you are watching most closely, that you think could shake the snow globe? One could be just a leapfrog in chip manufacturing technology.

Yes, but not exactly. It's not secret, but some of that information comes from analysis and things that I've seen that I don't think should be public. So, I will politely decline to answer.

Okay. How about the frontier companies? I guess I'd be interested in your list of who the frontier companies are. I usually have a pretty short list that I would actually include in that. These are the live players?

Yeah.

And I guess I'm interested in how they present to the government. How does the administration, or the government more broadly, understand these different companies? How differentiated are they in the view of the government?

It was striking that you hear these things like, “Anthropic has really good, deep relationships with the government.” I don't know if that may be an out-of-date statement now. You can tell me if you have any thoughts on that, but they're often reportedly the sort of company with the longest-tenured people, the deepest relationships, and the most mature presentation. I've heard that.

OpenAI sort of has a reputation for saying what people want to hear at different times. Obviously, Elon was involved in the administration in a serious way for a while. Then they all seemingly got federal contracts, and it happened right at the same time that Grok was calling itself MechaHitler. I was just like, it's very strange, right, that we have this sort of blanket announcement that we're going to do a deal with Claude and MechaHitler, and we're going to put them out there, all on kind of the same press release.

It just got me wondering: Do people see these things as just very similar entities, or is there a more nuanced understanding of who these companies are and what they represent, and what doing business with them might mean in the future?

I can't really speak for other people. I would say it's certainly the case that a company like Anthropic has quite good relationships and quite a good reputation inside, specifically, the intelligence community. Part of that is because Anthropic invested serious resources early on, because Dario personally is so NatSec AI-pilled. They invested significant resources in getting their models stood up on the high side, which is to say, in classified networks.

It's not easy. It's nontrivial technical work that has to be done to get this stuff on the high side. Of course, they also have to have employees who have top-secret clearances. You have to have the technical staff to be able to do that. There's all sorts of work that they have to do. It's not like the other companies haven't done this—obviously, Google in particular has—but certainly Anthropic, among the frontier AI companies, invested early.

I personally view the frontier firms as basically Google DeepMind, Anthropic, and OpenAI. I think that hasn't really changed for 3 years.

Obviously, you're going to go do some other stuff that you kind of alluded to, where the administration might not do enough, and you'll try to be upstream or influence things from the outside. What would you say, in general? You could maybe list off some ideas: What is the government going to fail to do that other people need to pick up?

This could be state-level regulation, which you have an interest in. It could be things done purely by the private sector or the philanthropy sector. How would you advise people who want to make an impact, want to make the future safer, brighter, and better, and want to absorb your wisdom about the best way to complement what the government is going to do?

I will reframe it slightly as things that the government will not be well-suited to do, and also as the division of labor. For example, I am not in the camp that we need to restrict states from passing laws. I think when it comes to the regulation of AI development, it's pretty unambiguous that we can't have 50 rules.

But I think if a state wants to be totally retrograde and freeze certain aspects of its society and hamper it through really aggressive AI-use restrictions, I won't support that. I'll argue against it if given the opportunity, but I think that's up to them. I think that's up to the people of that state.

In general, I think governments are going to fall down. Governments are not going to do a good job at defining what good looks like when it comes to things like AI safety and whatever else. There's an information asymmetry that makes it really difficult. I will tell you, having worked at the White House, I don't know tremendously more about what goes on inside the frontier labs than you do. There might be ways in which you know more than me, because you probably have relationships with researchers that mine have withered a little bit in recent months.

So, I think that I look at a company, for example—the AI underwriting company that a former Anthropic employee founded, which recently got money from Nat Friedman and Daniel Gross. That’s a really cool company doing something private-sector-led. It’s a for-profit insurance company—or, well, it’s a managing general agent, but that’s a technicality—that’s trying to build AI standards. I think that’s very cool and plausibly both a very good business and a very positive-for-the-world one. It’s a total win-win: it’s probably a very good business, really good for adoption, and also probably good for AI safety.

As a general matter, I hate to say this because I was just in politics and policy, but I hope that bright young people mostly stay out of the world of policy, because the world of policy is going to move you closer in the direction of zero-sum games, whereas markets are much more about positive-sum outcomes. And so, there are many, many companies to be founded. My God, there is so much money. I’m sure there are people doing this and I just don’t know about it, but the world of biosecurity could use some startups so desperately—some people who are thinking about that stuff really desperately. There’s lots of stuff like that, and I think you could build some great businesses in biosecurity and do a lot of good for the world.

There are principled ways you can develop to think about what’s likely to be a company and what’s likely to require government regulation. I think governments will struggle to articulate what good looks like—the specification for what good looks like—in the context of specifically catastrophic-risk mitigation. That will be hard, and so I think we will probably need lots of folks in civil society and the private sector to help us do that.

As a general matter, what we need for the adoption piece—the adoption side of things—is also about lots of things that ultimately implicate matters of AI safety that a lot of your listeners might be primarily interested in. This is sort of the thesis of the AI underwriting company, by the way: if you can create things that look like gold stars, that’s also the theme of the work that Fathom is doing, an organization that I used to be affiliated with.

More broadly, we’re just so pre-paradigmatic on so many different things. I don’t hear anyone talking about the many ways in which I feel like agentic commerce, in particular, is right around the corner. I hear so little about that, either from a technocratic regulatory perspective or a more conceptual perspective about what to expect it to look like. There are so many sector-specific things where it’s like, let’s really talk about this. Let’s stop saying, “Yeah, AI is going to be good in healthcare,” and start saying what that means more specifically, what it requires of us, and what specific kinds of institutional adaptations are necessary to make AI work well in healthcare. I think we’re in a maturing sector, and that means it’s time for our thinking to get more specific and concrete. Again, that’s a big part of what the action plan was about, too.

Yeah. Yeah. The agentic future is certainly a topic where your crystal ball gets real foggy, and there’s precious little exploration. I recently did an episode on the AI Village. There are a few of these sort of gonzo experiments that try to shine a little light on it, but it does seem like we’re going to put a huge number of AI agents into the economy in a very short period of time. We really don’t know at all what the dynamics of that are going to be, and it’s just kind of wild.

It links up so well with the stablecoin legislation that the president signed a couple of weeks ago. It’s such a good linkage there: yes, and America is pretty darn good at financial services. So there’s a pretty good chance that stuff spreads throughout the world, and we end up creating applications that could be the really transformative ones, used all over the world and really world-changing—and in a positive way. They could look a lot like American dominance in AI. It could be that that’s one of the best areas. That’s sort of my view of the situation.

So, I personally—and I feel like, by the way, this is true of the action plan—the action plan doesn’t go enough into healthcare. The action plan does not go enough into agriculture. What about Veterans Affairs? That’s healthcare; it’s a huge single-payer healthcare system, one of the biggest single-payer healthcare systems in the developed world. We can totally use that to do interesting things, I would think. There’s so much stuff like that that’s worth doing.

Well, thank you for spending all this time. It’s been comprehensive, and I’ve really appreciated it. Anything you want to touch on that we didn’t, or any final thoughts—things you want to leave people with?

No, I think the only thing I would say is that I am happy to announce that I will be resuming Hyperdimensional, my Substack, on a weekly cadence. I will be joining the Foundation for American Innovation as a senior fellow, and I expect to have other kinds of institutional announcements, institutional affiliations, and roles to announce in the coming weeks. There’s so much to do, so much to catch up on.

I sometimes worry that Hyperdimensional will have to go into a twice-a-week cadence because there’s a backlog of things that I’ve been itching to say, and so much to come soon.

Cool. Well, we will certainly be following and look forward to having you back to discuss many of them. For now, Dean W. Ball, fresh from the White House, thank you for your service, and thank you for being part of The Cognitive Revolution.

Thanks so much for having me. This was fun.