Welcome back to the Model Convo series —— micro interviews with researchers in AI governance and adjacent fields.

This week’s convo is with Kurtis Hingl, a PhD Fellow at the Mercatus Center at George Mason University and an incoming Teaching Assistant Professor of General Business at West Virginia University and affiliate with the Kendrick Center for an Ethical Economy.

He writes at The Hunchbox and Price Theory Palooza here on Substack.

What was your path into AI and metascience, and what are you working on now?

I studied economics, math, and philosophy for my undergraduate degree —— a pretty typical path —— and I went to economics graduate school to study industrial organization, the competitive processes of markets and industries.

I quickly found myself asking more and more questions about the industry of scientific research itself, whether from my lack of creativity or because I was thinking about recursion from reading Gödel, Escher, Bach. I’m not sure. But I started applying the language and logic of markets and social systems to science, and I haven’t stopped.

Of course I was reading a lot of Hayek, Polanyi, and Popper on the social nature of scientific knowledge and its role in a free society, and then I also read Joseph Henrich’s The Secret to Our Success early in grad school.

In my opinion, once you start thinking about the growth of knowledge, it’s hard to think about much anything else —— to paraphrase the Robert Lucas quote passed down to his student Paul Romer. At the same time, I’ve always been more comfortable with the microeconomic questions of incentives, property rights, and organizational structure, so that’s where I have pushed my focus within the broader metascience ecosystem.

My current work examines how scientific fields differ in their organization, in part due to the “testability” of the subject matter. But I am also trying to keep up with all the news in AI, think about how it relates to science, and incorporate it into my own research as much as possible.

AI is the most interesting thing to wrestle with in every facet of my work: science, market processes, education, you name it. It is impossible for a young economist to not be interested in AI. Actually, it’s impossible for any awake and curious person to not be interested in AI.

What works of art have most shaped your views on AI and metascience?

A cliché for sure, but no doubt A Canticle for Leibowitz. How is dispersed knowledge valued and preserved? Who gets to decide what is worth keeping from the past? What’s the value of the thinker, the tinkerer, or the mere copy-cat? Why shouldn’t we all just focus on the immediate practical concerns around us? Can the thinker also be practical? Those are not even the main questions Miller wrestles with, but a few that come to mind that have infiltrated my thought. And there’s also the great public radio audio drama.

On AI, I can’t say that there’s been a major work of art that’s influenced me, probably to my own detriment. I can say that I spent a lot of time growing up working in landscaping and excavating, a lot of time camping and backpacking, and a little bit of time working in manufacturing.

It wasn’t that weird, then, when we all started to talk to LLMs. I’ve been communicating with technology deliberately all of my life. Embodied knowledge as an extension of the self is nothing new in the human condition, and yes there is often a conversation going on, if not in English. Maybe the right answer is the scene in Anna Karenina where Levin learns to mow his fields.

What’s your most contrarian take on AI and metascience?

Science might get cheaper, but the frontier of science will only get more expensive and will be directed by elite scientists, and that’s okay. But this is contrarian more in mood than in fact ——many people anticipate concentration of power but also think it’s very bad.

AI greatly increases the productivity and output of the everyman, but it does so at higher rates for the elites. What we really care about is the big breakthroughs on the frontier of knowledge, and these will happen in the partnerships between leading AI labs and leading scientists, not from the masses that have also leveled up. That is okay, good in fact.

What this means for metascience is that the big gains are from bargaining with the big players, not overhauling the system or going renegade trying to build a new system. As much as I sometimes hold the common disdain for the layers of institutional status quo, they will continue to dominate.

Academic publishing, the PI-lab-grant-funding situationship, and the concentration of prestige among a small number of incumbent institutions are here to stay. Again this is okay! We don’t have to give up on the hope for building better science. It’s just that it is more likely to come via convincing the wielders of power than by overthrowing them. We will need to be more Coasean.

I guess predicting the continuation of the status quo isn’t that contrarian, but it sometimes seems like it in a community always thinking about progress. And on concentration, Average is Over is already a well-established point, so let me give a hopeful caveat where there are still huge gains from leveling up the masses.

Imagine if it became a norm for, say, the city council of a small town to run projections —— GDP, population, traffic —— to help decide if that bike lane should go in. I’d like to think this moves us to better equilibria than the evidenceless persuasion running most public and private decision making.

There are versions of our future where this kind of marginal improvement in governance will lead to large and surprising effects. A cure for cancer might be just as likely to come from an innovation in rules around “approval” than a hi-fi whole-cell simulator.

What are you reading, watching, or listening to now?

I noticed it’s come up a couple times in these conversations, but yes, I am coincidentally reading East of Eden. Surprisingly relevant, especially the importance of naming things. Much of life (and science) is about naming things, as it is Adam’s task in the Genesis narrative.

In economics, we sometimes talk about “identification”—giving a claim an identity by carefully distinguishing it from the relevant counterfactual. In my household, my toddler is just learning to speak. His task, like Adam’s (both in Genesis and in Steinbeck) is to name things. In public discourse, much talk is, “Is that thing really an X or is it a Y?”

I’m always reading new econ papers, reading Twitter and Substack to keep up with AI, and always trying to poke through something non-fiction that’s bigger and deeper. Right now that’s Montesquieu’s The Spirit of the Laws.

I’m always listening to some combination of mewithoutYou, The Mountain Goats, and bluegrass. Also, Willow’s new album is quite good, though I don’t know what it’s about to be honest.

I don’t watch much of anything. There’s a couple YouTube series of runners preparing for Boston I was following, and analysis videos from the Candidates while that was happening.

Go-to emerging science / tech music track?

OK Computer is a classic, though like a lot of the music I end up loving, I don’t fully identify with the message. How about Pete Seeger’s “We Shall Overcome”?

Subscribe now

Welcome back to the Model Convo series — micro interviews with researchers in AI governance and related fields. If you know someone I should interview — maybe that’s you? — please email me.

This week’s convo is with Abi Olvera, who writes at Positive Sum and is Research Director at the Golden Gate Institute for AI.

Abi is also a Senior Fellow at the Council on Strategic Risks, an Editorial Fellow at the Bulletin of Atomic Scientists, and an affiliate with the Institute for AI Policy and Strategy.

What was your path into AI and what are you working on now?

I was a US diplomat for ten years. My official title was “generalist foreign service officer.” My job was to become an expert in whatever subject my diplomatic assignment required. I focused on a lot of AI-adjacent work like cyber, biotech, quantum, China.

I’ve been tracking AI since 2013. Back then, I followed object recognition apps that converted handwritten notes to text and watched natural language processing and text-to-speech improve. As the field accelerated, I made it a professional focus tracking AI as a whole and watching how it reshaped other sectors.

Now I focus on bottlenecks, particularly the underlying constraints that explain why the world looks the way it does. For example, cyber attacks are less common than you’d expect given how easy it is to hack insecure systems. A big bottleneck is that money laundering is hard!

I’m using this approach to look at bioweapons now. Many assume AI dramatically lowers the barrier to working with viruses. But biology is really hard and more unpredictable than other physical sciences. The same frustrations that limit legitimate scientists working on a cancer cure also frustrate someone trying to engineer a virus.

What works of art have most shaped your views on AI?

I devoured choose-your-own-adventure books as a kid. I read every one at the library.

Maybe that’s why it bothers me when movie characters end up in worst-case scenarios because they didn’t communicate clearly to their loved ones or call 911 when a reasonable person would have. Relatedly, films tend to be pessimistic about technology, but even those bad futures require society to make a string of bad decisions. A bad future isn’t the default. It’s also not what’s historically happened.

I can see this when I read about how the art of cooking and homemaking has transformed over the centuries. I am really fascinated with an 1800s frugal housewife book that explains how poorer people got by. Parents had toddlers learn to knit to make rags into rugs. People saved their fire ash to trade for soap. A lot of our recipes, like French toast, were to make stale food edible. We live in a very different, much better world.

What’s your most contrarian take on AI?

We’re going to be OK. Though we should still work on minimizing downsides and making sure things go well.

On job loss fears, people consistently underestimate how complex other people’s jobs are. I hear people in San Francisco assume think tanks can just be replaced by AI. But think tanks don’t only do research. These organizations compete for finite attention, build credibility over time, and signal alignment with shifting power centers. Trust is a scarce input that AI doesn’t automate directly.

Even in a world where AI handles most knowledge work, someone still has to break the status quo. That’s still human entrepreneurs, writers, artists, activists.

My hottest take: AI has probably improved more lives than it’s harmed so far. Anxiety and depression are the most common mental health problems. Studies show AI companion apps produce 30–35% improvement in symptoms. While not as effective as therapy, it’s cheaper and available 24/7. One study found 3% of Replika users said the app stopped their suicidal ideation.

Headlines focus on AI-assisted suicides. Mental health research, read as a whole, can also suggest AI has prevented more than it’s caused.

What are you reading, watching, or listening to now?

Substack! I love that it’s often practitioners writing outside their institutions or passionate people going super niche on a question I’m also noodling on.

For books: Superabundance and The Great Stagnation. For TV: The Pitt, and a Colombian show called La Pola about revolutionary-era Colombia. Watching characters be certain about things we now find horrendous (racial castes and the divine right of kings) is a reminder that moral frameworks can be powerful even when they’re wrong.

Go-to emerging tech music track?

Music energizes me during breaks. I love musical theater, pop, and soul. One of my main hobbies is learning to sing these. These genres might not seem directly connected, but hey, recorded music was once an emerging technology too.

[Update: Abi chose “A Whole New World” as her track.]

Subscribe now

Introduction

The nationalization of frontier AI labs is no longer just a thought experiment.

In September 2025, Senators Josh Hawley (R-MO) and Richard Blumenthal (D-CT) introduced the Artificial Intelligence Risk Evaluation Act (S.2938). In Section 5(b)(8), the bill directs the Department of Energy to “develop proposed options for regulatory or governmental oversight, including potential nationalization or other strategic measures, [emphases ours] for preventing or managing the development of artificial superintelligence.” Senator Marsha Blackburn (R-TN) later incorporated that language into her TRUMP AMERICA AI Act, a 291-page draft omnibus framework introduced in March 2026.

We have been closely following federal discussions around AI governance, so putting “nationalization” into bill text as a policy option seemed to us unprecedented in modern American legislation. We have discussed the normative dimensions of this proposal in a separate post, but for now we wanted to understand how the word has been used in recent Congressional history.

To that end, we built a pipeline that searched the full text of every bill and floor speech in the Congressional Record available through the GovInfo database going back to 1993 and used an LLM to classify each mention by how it was being used.

Bottom Line Up Front: “Nationalization” has been a congressional slur for 30 years; 78 percent of the 1,424 mentions we analyzed were rhetorical attacks. The Hawley-Blumenthal AI bill represents an Overton window shift as the first bipartisan legislation to put nationalization in bill text as a forward-looking contingency.

Results

Across 1,424 mentions of “nationalization,” “nationalize,” and “government takeover,” excluding the AI bill, here is what Congress means when it says those words:

Opposition Framing (77.9%)

Opposition framing is by far the largest category, at 1,109 out of 1,424 classified results. When members of Congress use “nationalization” or “government takeover” in relation to domestic policy, it is overwhelmingly an accusation, e.g., the other side is trying to nationalize healthcare, banks, or education.

The Affordable Care Act debate of 2009-2010 dominates this category. “Government takeover of healthcare” was repeated on the House and Senate floor hundreds of times, sometimes as the literal title of floor speeches.

During the 2008-2009 financial crisis, members argued against nationalizing auto companies during the bailout debate and described the conservatorship of Fannie Mae and Freddie Mac as nationalization.

In 2019, Senator Ted Cruz (R-TX) introduced the E-FRONTIER Act to protect 5G networks from nationalization.

When Congress put “government takeover” into legislation, it was always as something to be stopped, never as something to be done.

Foreign Context (11.2%)

Foreign context is the second largest category at 159 mentions. When Congress discusses nationalization, it is overwhelmingly talking about other countries doing it. The Cuban Liberty and Democratic Solidarity Act (LIBERTAD) deals with Cuba’s nationalization of American-owned property. Similar bills address Venezuela under Chavez and Chinese business practices.

Typos and Unrelated Usage (8.1%)

In 114 results, the word “nationalization” has nothing to do with government control of industry. The most common culprit is a simple spelling error. The congressional record repeatedly refers to the “Immigration and Nationalization Service” when the actual agency was the Immigration and Naturalization Service. This mistake appears in the Violent Criminal Incarceration Act of 1995 and several budget resolutions and floor speeches.

The rest of this category uses “nationalizing” to mean “making something national in scope,” as in the Developing and Nationalizing Key Cannabis Research Act, 15 versions of the Indian Health Care Improvement Act (which uses “nationalization” to mean expanding a regional Community Health Aide Program in Alaska to the national level), or describing making a local election about national issues.

Historical Reference (2.6%)

In 36 results, members of Congress reference nationalization as something that happened in the past, both abroad and at home. These include Iran’s nationalization of British oil assets in 1951, Guatemala’s threat to nationalize United Fruit Company holdings in the 1950s, communist regimes in Eastern Europe seizing foreign insurance company assets after WWII, Mexico’s nationalization of its oil industry, England’s nationalization of healthcare in 1948, the near-nationalization of American railroads in the late 1970s, and President Truman’s attempt to nationalize steel mills during the Korean War.

The 3 Policy Proposals (0.2%)

Besides the AI bill, only 3 results out of 1,424 are instances where a member of Congress seriously entertained nationalizing a domestic industry. In chronological order:

In January 2009, Representative Dennis Kucinich (D-OH) proposed nationalizing the Federal Reserve by placing it under Treasury control, arguing this would let the government spend money directly into circulation rather than borrowing from banks.

In February 2009, Representative Jim McDermott (D-WA) acknowledged bank nationalization as a potential policy option that might be necessary, presented as one of several tough decisions the President might need to make.

In February 2021, Representative Sheila Jackson Lee (D-TX) proposed the COVID-19 Delivery Act, which included “nationalization or federalization of distribution” to organize vaccine delivery. This is the closest pre-AI parallel of nationalization in actual bill language.

Takeaways

The language of nationalization in the Hawley-Blumenthal AI Risk Evaluation Act is largely without precedent. The 2008-2009 financial crisis saw congressional discussion of bank nationalization, and Representative Jackson Lee’s 2021 vaccine distribution bill used similar language.

Yet it is distinctive in two ways:

First, it is in bill text, with bipartisan cosponsorship. The bank nationalization discussion stayed at the level of floor speeches, and nobody put “nationalize the banks” into a bill as a policy option. Senator Hawley did that for AI, with a Democratic cosponsor in Senator Blumenthal, and Republican Senator Blackburn incorporating the language months later into her own framework.

Second, it is prospective rather than reactive. The discussion after the financial crisis was about nationalizing institutions that had already failed. The AI bill is about developing plans to nationalize an industry that has not failed, in preparation for artificial superintelligence that does not yet exist.

The Artificial Intelligence Risk Evaluation Act thus represents an Overton window shift, moving nationalization from political attack to policy.

Methodology

This analysis was conducted on March 29, 2026, using publicly available data from the Library of Congress and the Government Publishing Office.

We searched the full text of congressional bills and the Congressional Record for “nationalization,” “nationalize,” “nationalizing,” and “government takeover,” pulling all matches through the GovInfo Search API. We then enriched each result with bill metadata from the Congress.gov API and ran them through Claude Sonnet 4, which classified each mention by how it was being used (opposition framing, foreign context, historical reference, and so on) and whether it represented a serious policy proposal.

After Claude’s classification, we manually reviewed the results and reclassified entries where the LLM had miscategorized the usage. For example, several results initially labeled as “policy proposals” turned out to be constituent mail read into the record, historical references to past nationalization efforts, or members deflecting questions about nationalization rather than endorsing it.

The GovInfo database we used for this project only covers bills after 1993 and the Congressional Record after 1994. Because of this selection bias, we were forced to exclude nationalization debates that took place before the Reagan realignment. That means we did not analyze congressional debates around the Progressive era, New Deal, Great Society, and 1970s energy crisis, when congressional nationalization proposals were likely deployed more favorably.

Our code and full dataset are available on GitHub.

Subscribe now

Welcome back to the Model Convo series — micro interviews with researchers in AI governance and related fields. If you know someone I should interview — maybe that’s you? — please email me. This week’s convo is with Tyler Cowen, who writes at marginalrevolution.com, as well as here on Substack.

What was your path into AI, and what are you working on now?

I first became interested in AI when I saw the chess computer Tinker Belle wheeled into a New Jersey chess tournament in I think 1975. I followed the Kasparov matches closely, and the more general progress of AI in chess. I read chess master David Levy telling me that chess was far too intuitive for computers ever to do well. He was wrong, and then I realized that AI could be intuitive and creative too. That was a long time ago.

In 2013 I published a book on the future of AI called Average is Over. I feel it has predicted our current time very accurately. I also taught Asimov’s I, Robot – a work far ahead of its time – for twenty years.

Right now I am simply working to keep afloat and to stay abreast of recent AI developments. I blog and write columns on the topic frequently, and have regular visits to the major labs. I encourage universities to experiment with AI education.

What works of art have most shaped your views on AI?

I know pretty much all of the main AI-related movies, but I cannot say they have influenced me, as much as I have enjoyed them. If anything, they are a precautionary tale against lunging at very definite conclusions. Iain Banks has a relatively sensible treatment of AI in his Culture series, but in terms of influence, and seeing potential problems, it is really Asimov far in the lead.

What’s your most contrarian take on AI?

Do I have one any more? Andrej and Ilya have come around to the notion of slow take-off, formerly considered a weird or radical position, at least in the AI community. I think for the most part that people either significantly underrate or overrate AI potential, and I am somewhere in the middle.

What are you reading, watching, or listening to now?

Most of all, playing around with the top AI models. Buying new recordings of classical music. Reading on the history of South Africa, and for the first time reading Steinbeck’s East of Eden. Reading Twitter.

Go-to emerging tech music track?

What counts as tech music these days? AI music is promising but not yet as good as what humans create at their best. Can I count William Byrd Renaissance choral music as “tech”? I think so. To that I will add McCartney’s “Temporary Secretary” from the 1980s.

Subscribe now

Last week, Senator Marsha Blackburn (R-TN) introduced a comprehensive legislative AI framework, which contained a number of regulatory proposals for the nascent AI industry.

Blackburn’s TRUMP AMERICA AI Act — which stands for The Republic Unifying Meritocratic Performance Advancing Machine intelligence by Eliminating Regulatory Interstate Chaos Across American Industry — stitches together a number of bills from the more tech-skeptic wing of the Senate GOP.

Her proposed framework stands in marked contrast to the minimal framework published by the White House last Friday — as in three-pages-of-bullets minimal.

Blackburn’s legislation has also arrived after a year of mostly fruitless attempts by the White House and Senator Ted Cruz (R-TX) to pass a federal moratorium or preemption of state AI laws, which are beginning to pile up in red and blue states alike.

One section of the TRUMP AMERICA AI Act uses language from Senator Josh Hawley’s (R-MO) Artificial Intelligence Risk Evaluation Act of 2025, which calls for the Department of Energy to establish an Advanced Artificial Intelligence Evaluation Program.

One of the duties of that program is to “develop proposed options for regulatory or governmental oversight, including potential nationalization or other strategic measures, for preventing or managing the development of artificial superintelligence if artificial superintelligence seems likely to arise.”

Now, it would be a mistake to read this line as merely proposing the development of a last-resort safeguard only to be deployed under conditions of acute existential risk, when the machine god knocks and we are left only with the choice to answer or hide. Taken at face value, guidelines for nationalization might appear to be a neutral, commonsense standard left dormant until there is no other choice — an option we would exercise only under duress.

But that’s not how power works. The incentives for states to gobble up civil society and private firms are too strong. They are competing centers of power and plenty. Though, in America at least, only one of those organisms has reliable, persistent recourse to violence. Even so, we’ve learned to live with the balance. It is, however, a precarious one.

The French political economist and futurist Bertrand de Jouvenel observed that the impulse toward domination is natural among states. It persists across time, space, and ideology. And that impulse is the best explanation of all nationalization efforts:

In the end, calling it socialization or nationalization, the state strives to make its own all the great castles of the economic feudal system, the railway companies, the electricity distributing companies, and so on. Only those who know nothing of any time but their own, who are completely in the dark as to the manner of Power’s behaving through thousands of years, would regard these proceedings as the fruit of a particular set of doctrines. They are in fact the normal manifestations of Power, and differ not at all in their nature from Henry VIII’s confiscation of the wealth of the monasteries. The same principle is at work; the hunger for authority, the thirst for resources; and in all these operations the same characteristics are present, including the rapid elevation of the dividers of the spoils. Whether it is socialist or whether it is not, Power must always be at war with the capitalist authorities and despoil the capitalists of their accumulated wealth: in doing so it obeys the law of its nature. Whether it is socialist or whether it is not, it cannot but present itself as the ally of those who are under the dominion of the capitalist. Philanthropy, it is true, plays a part in this alliance. But the sure instinct for the distension of the state necessarily turns this philanthropy to the glory and strength of Power. [Emphasis mine.]

De Jouvenel would not be surprised by calls for nationalization from either the left or the right, particularly when the prize is the commanding heights of frontier AI labs, which promise real and projected wealth to whoever holds them.

Democratization, socialization, collectivization — call it what you like. The antagonism between the state and other macro-organisms has been a core feature of our political history. Unlike the specter of ASI, we have a seemingly unlimited number of case studies with which we can form a reference class.

Whenever the prize is high enough or the perceived risk is sufficiently costly, both rent-seekers and safety advocates alike will push for more control. This is the well-worn “bootleggers and Baptists” dynamic, which takes its name from the mutual desire of these actors to ban legal alcohol sales in pursuit of both selfish and altruistic ends.

The issue with even extremely preliminary contingency plans to nationalize AI firms is that any ASI threshold is unlikely to be reserved only for moments of existential crisis. There is simply too much of an incentive for interested parties to inflate the risk or loosen the definition to fit the desired ends of those who want to consume for selfish and altruistic reasons.

Early efforts to deploy a policy as risky as nationalization for a concept as murky as ASI will invite a high degree of lobbying and arbitrage that is at odds with the intent of otherwise well-meaning lawmakers.

And looking at the current admin’s actions against Anthropic over the past month, or its ambiguous, unchecked war-making, or its growing purchase of financial stakes in critical industries, or its willingness to sell our most advanced tech to our greatest adversaries for a cut of the deal…

Well, we should not hold strongly to the assumption that lawmakers are otherwise generally well-meaning.

Subscribe now

Welcome back to the Model Convo series — micro interviews with researchers in AI policy and related fields. If you know someone I should interview — maybe that’s you? — please email me. This week’s convo is with Andy Masley, an independent writer here on Substack.

What was your path into AI, and what are you working on now?

In college I double-majored in philosophy and physics. I had come in as a standard internet atheist, and was really interested in physicalism, the idea that there isn’t anything extra-physical affecting the mind or anything else about the world. Once even simple high school Newtonian physics really clicks for you, it’s hard not to think that way.

But there was this gigantic hole in my worldview: at a gut level it was hard not to think that consciousness and the mind seemed completely magical and distinct from the physical world. I had very strong Cartesian intuitions. I got lucky and had a professor my freshman year who was really excited about the philosopher C.S. Peirce. I ended up reading a lot of him and felt for the first time like I was actually getting real footing in philosophy of mind.

A central goal for Peirce is demystifying the mind’s abilities and pushing against Cartesian intuitions. The best place to start is “Questions Concerning Certain Faculties Claimed for Man,” and then probably his best essay “Some Consequences of Four Incapacities.” Those made it easier for me to think about the mind as a natural system, and led me to reading a lot of W.V.O. Quine and Daniel Dennett later on. Looking back, I think about just how lucky I am that 18-year-old me stumbled on this stuff. It seems like some of the absolute most relevant ideas I could’ve been engaging with at the time.

I also read Donna Haraway’s A Cyborg Manifesto in a philosophy of technology class — it’s a very fun critical theory paper — and became passively interested in continental philosophy of technology, though I know way less about it than my main focus areas in college. Turned out to be useful later.

Separately, I read Reasons and Persons my junior year and that felt like an ideological nuclear bomb. I loved the way Derek Parfit thought and wrote. Reading more about him put effective altruism on my radar. I had known about Peter Singer and his basic arguments for effective giving for a few years, but got much more interested in EA arguments and thinking once I found the official movement.

I was basically sold, but at the time EA was (at least to my eyes) focused pretty exclusively on earning to give and running effective global health and animal welfare charities. I thought that was great but didn’t really see how I could contribute, so I didn’t actually reach out for another seven years. I just lurked on the forums, blogs, and podcasts, and donated to GiveWell and lab-grown meat.

Around 2015 I started seeing EA and rationalist arguments for AI x-risk. I remember reading reviews of Nick Bostrom’s Superintelligence and some summaries of the argument, and coming away with a deep sense of dread. Getting into philosophical naturalism about the mind had really primed me to take these arguments seriously.

If you think (as I did and do) that there’s nothing magical about the mind, talking about machines that could be way smarter than us sounds as reasonable as talking about machines that can lift way more weight than us, and the implications of that seemed like a mental third rail that I didn’t want to touch.

I felt like I’d go crazy if I thought about it too much. I decided I didn’t actually want to look more into it and figured, “Well at least I’m lucky that these arguments are so irrelevant to my life right now. It’s all so far off and there’s no way I can add value here anyway.” It’s really funny thinking about how that was only a decade ago.

I started listening to the 80,000 Hours podcast pretty early when it launched in 2017. That made a bunch of AI arguments feel much more accessible to me. Some were just so crazy fun. I remember the Hilary Greaves episode feeling like the most info-dense audio I had listened to at the time. Before that I didn’t see how to approach risks from superintelligence in a way that felt like serious thinking.

I’m very wary of thinking from first principles or doing anything that feels more like eschatology instead of sober prediction, and the conversations on 80k were just so fruitful and grounded in comparison that it gave me a lot to run with. 80k is I think perpetually underrated as a resource despite having a lot of success. My writing’s been influenced a lot by their problem profile pages. They’re very comprehensive and don’t talk down to the reader.

I eventually got involved in the EA DC city group in 2020 during the pandemic. I had never met anyone else especially into EA before that. I’ve always been kind of ideologically snippy and avoidant of surrounding myself with people who agree with me too much, but I was consistently blown away by the people I was meeting in the scene.

I eventually took on the role leading the EA DC city group, which mostly meant connecting people to each other and making the network legible to anyone who wanted to get involved, and representing EA when I could to the broader public.

This was a really incredible period of growth for me. And I felt and still feel like I wasn’t actually changing my views much based on social reward, something I worry about with any group of people into objectively weird ideas. For better or worse I find that most of my beliefs about big EA questions are the same as they were before I got involved. I’m pretty grateful I had so long to marinate in them.

In the last year I started to blog a lot more. I got a ton of traction on some early posts about AI and the environment and ran with it. I built up enough of an audience that people suggested I experiment with writing full-time, so I very recently stepped down from EA DC to do that. I want to experiment with getting across big ideas that have been important to me, and also do more deep dives on data centers, which is the thing that really built my audience to begin with. Almost everything I write will be focused on AI in one way or another.

What works of art have most shaped your views on AI?

I’m sure this one will be said a lot, but a theme of The Three-Body Problem is that within some local complex system you and/or your civilization can be doing everything right and succeeding on every metric. Society can grow and experience real progress for hundreds or thousands of years. But it can all be embedded in some larger system you barely understand where the rules work completely differently, and that can suddenly just upend everything.

The books are partly about characters trying to figure out the (often bleak) rules of these much larger alien systems, even though they may have lots of personal incentive to ignore them. I think this is just an obviously true fact about the world that’s very hard for us to keep in mind, because we evolved to operate in systems of social status that mostly reward treating the world like it’s going to stay constant.

I have a strong emotional inclination to be an upbeat techno-optimist about AI. I’m otherwise very excited about basically all tech progress that’s happened in my lifetime, but I have to step out of that and think seriously about how inventing intelligent machines could be radically different than all prior tech, and The Three-Body Problem headspace is a great way to approach that.

I also think a lot about how some of the most important, meaningful experiences of art I’ve had wouldn’t have been possible without very recent technology. I wouldn’t have found a lot of the music I loved as a teenager without the internet, or been able to take walks listening to it without an iPod or Walkman. These were some of the most important experiences I’ve had.

I think it’s very easy to view new tech as if it’s going to be very flat and cold and inert. We have a lot of nostalgia for the past and it’s hard to imagine deep specific experiences like that coming from stuff that’s invented after we’re thirty. For a lot of people hearing about the internet for the first time, I’d imagine they might’ve just thought of all the emails and wires and glowing screens, and wouldn’t have felt in their gut that the internet would enable all these new deeply personal meaningful experiences.

So, when something as consequential as AI is on the horizon, I try to think about all the ways it could make life more meaningful and human for people in the way the internet did for me. There will be new things like walking around late at night listening to ‘90s indie in suburban Massachusetts, something that feels like a fundamental part of reality to me but is actually just so recent and contingent.

What’s your most contrarian take on AI?

Depends on who I’m talking to. I’ve built an audience by arguing that the environmental impacts of AI have been wildly overblown, and in a lot of spaces that’s seen as very contrarian, even though I think “using a computer program isn’t in the same ballpark as driving a car” should be just wildly obvious.

Within EA circles, I’m still pretty open to arguments for much longer timelines or that current systems won’t scale to AGI, although I’m just wildly uncertain.

After a decade of reading about it, I have almost no really confident takes on AI risk and AGI beyond, “Hmm, seems like a big deal,” and that’s feeling more and more uncomfortable every day. I’m hoping this is my year to develop a solid inside view.

What are you reading, watching, or listening to now?

Reading: The Search for Modern China. Considered one of the best general histories, it’s really fun.

Watching: I love Les Blank’s documentaries about New Orleans, like Always for Pleasure. I was watching them a lot during the pandemic. Gonna be working through the ones I haven’t seen.

Listening: Been a busy month for me so I’m not listening to much new or complex music. The Japancakes cover of Loveless has been really fun and surprisingly good for focus.

Go-to emerging tech track?

Half-jokingly, anything from the Akira soundtrack, especially “Kaneda’s Theme.” Or something very chaotic and overwhelming but certain of itself, like “Miss Fortune” by Faust.

Subscribe now

Welcome back to the Model Convo series — micro interviews with researchers in AI policy and related fields. If you know someone I should interview, please email me.

Past and present bio?

I’m an assistant professor at Alliance Manchester Business School. I work primarily on AI and business ethics, but my background is in political philosophy, moral epistemology, and decision theory.

[Carlo also writes at Paperclips and Other Alignment Problems.]

Before AMBS, I’ve been employed or visited institutions ranging from post-92 universities in the UK to Ivy League departments in the US, which has left me with the strong impression that the variance among places we call universities is dramatically underappreciated.

How did you get into emerging tech?

Honestly? I just wanted to find an academic job. I did my PhD in political philosophy, but I’m mostly self-taught in everything I currently work on.

At some point I realized that the ethics of AI was a space where these tools were desperately needed and almost entirely absent, and that moral philosophy should be leading the conversation rather than playing catch-up.

Then when LLMs came out, I got completely obsessed. I started working with them at the very outset, to the point that I’ve gone from talking to people to almost exclusively talking to Claude Opus 4.6.

My spend on LLMs is starting to rival my spend on wine, which I find a bit degrading. On the bright side, I now enjoy drinking while talking to an LLM. Strange world we live in.

What work of art has most shaped your views on emerging tech?

I’ve never read sci-fi, and I’m not a big fan. But sometimes inspiration comes from where you least expect it.

I recently watched 50 First Dates — yes, the Adam Sandler romcom — and it made me think about intelligence and memory in ways most philosophy of mind seminars haven’t. It poses a question about what cognition looks like when you strip away continuity of memory, which is close to some of the questions we currently ask about LLMs.

The other film I keep coming back to is Alain Resnais’s Last Year at Marienbad, which I think is the most unsettling cinematic treatment of what it would feel like to live inside a simulation: characters who cannot agree on whether their shared past actually happened, moving through a space whose elegance conceals the fact that none of it may be real.

Most simulation fiction asks whether you can escape. Marienbad asks whether the question even makes sense if your epistemic situation is sufficiently degraded.

That strikes me as closer to the genuine philosophical problem.

What’s your most contrarian take on AI?

Two, and I think they’re connected.

The first is that AI doesn’t raise any genuinely new ethical issues. The problems are exacerbated, perhaps, and certainly more urgent, but the underlying structure is always something moral philosophy has seen before, which is exactly why ethicists should be doing more of the heavy lifting in these debates than they currently are.

The second is that most, if not all, cognitive offloading is likely to end up being good for us. As long as scarcity persists, and it does, offloading lower-order cognitive tasks to AI frees us to do higher-order work. The deskilling panic assumes there’s nothing better to do with the freed capacity, which is an extraordinarily strong empirical claim that almost nobody bothers to defend.

The only scenario where the argument genuinely breaks down is a post-scarcity world in which opportunity costs drop to zero, because at that point the comparative advantage case for offloading dissolves entirely. There is nothing scarce to reallocate attention toward.

That world is conceivable, though it is a long way from this one.

What are you reading, watching, or listening to now?

I always listen to Tyler Cowen’s podcast — still hoping to be on one eventually, but Tyler doesn’t even re-post my papers on Marginal Revolution. Tyler, if you read this, I still love you.

I obviously follow Andrej Karpathy and lots of AI podcasts, but I’m erratic rather than disciplined, and I have a very low attention span.

I also read a bunch of books diagonally, which is a self-indulgent way of saying I start them and then drop them.

The last book I read from A to Z is L.A. Paul’s Transformative Experience — pretty great stuff. The only thing that keeps me awake at night and attentive is chatting with Claude.

Go-to emerging tech music track?

I was a jazz pianist as a kid, so my instincts pull toward Gershwin and Miles Davis, music built on improvisation within structure, which is a decent metaphor for how we should think about governing emerging technology.

Recently, though, I’ve watched this incredible YouTube video of a song titled “Post-Scarcity Blues.” I believe everything is done by AI. It made an impression. I think it deserves way more views.

But I also recently watched the Take That documentary with genuine satisfaction. Say what you will — it’s one of the few instances where I thought studying organizations might actually not be boring.

Subscribe now

Three in four Americans expect artificial general intelligence (AGI) will “drastically change society as we know it” over the next ten years, according to new polling from Echelon Insights, a data analytics firm.

Echelon’s February consumer and verified voter omnibus surveys also show Americans are concerned about perceived local costs — namely those attributed to data centers — while voters and consumers are roughly split on the wider social benefits of AI.

In the consumer survey, 56 percent of respondents have a very or somewhat favorable impression of AI overall.

Surprisingly, a full 63 percent of employed respondents to the voter survey are not very or not at all concerned that AI will eliminate their job sometime in the next five years. This stands in marked contrast to some technologists, especially those from frontier labs like Anthropic and OpenAI, who often say they expect mass AI-driven job loss in the years ahead.

To be clear, the end state for long-run employment and automation is shrouded in deep uncertainty and is a contentious debate among technologists and economists, with the latter being more skeptical at the margin.

This voter attitude likely reflects a kind of rational ignorance. Most voters are not coders immediately threatened by new Claude tools, and most are likely responding to proximate indicators of job security. Claude cannot yet fully automate most forms of work.

But the question is when the levee will break. New data from the Bureau of Labor Statistics may already show a boost from AI-driven productivity, according to economist Alex Imas. Firms are beginning to invoke AI as a reason for layoffs, but it’s unclear how much of this is rhetorical strategy versus concrete reality.

The discrepancy between elite worries and mass nonchalance may come as a surprise to analysts in DC and San Francisco, many of whom are preparing for the possibility of mass AI-driven job loss in the short and long run.

DC’s second-place AI city status, however, comes with a more dour vibe. Whereas San Francisco is a city of private utopianism, DC remains a more gloomy exemplar of public pragmatism. As Jasmine Sun writes in her essay about a recent visit to Washington:

DC is arguably the most AGI-pilled city after SF—I’m always surprised to remember that the first chip controls passed before ChatGPT—but the default valences are opposite. Where AI researchers imagine growing a loving machine God, policymakers rush to contain His wrath. Every conversation here starts with damage control.

And calls from voters for more “damage control” will only grow as AI walks without rhythm into the jagged frontier.

That’s already starting to happen. Alongside America’s conflicting mix of optimism and pessimism about AI comes a bipartisan preference for government regulation. Six in ten registered likely voters believe government won’t go far enough in regulating AI while only one in four are worried it will go too far.

Overall, national views favoring regulation appear to be more firmly settled than views on the technology itself. This reflects a strong precautionary stance, one at odds with the permissionless innovation that is arguably needed for Americans to experiment with deploying AI in ways that work on their behalf.

These mixed numbers also suggest a rapidly closing window for the current administration to pass minimal congressional legislation or a state-law moratorium on AI.

Forecasters on Metaculus, a prediction platform, say the odds of a Republican Senate and Democratic House in 2027 are 66 percent, with Democratic control of both chambers coming in at 21 percent. The odds the GOP will hold both chambers, however, are just 13 percent.

With the odds starkly cutting against a GOP-controlled Congress, the Trump admin will be forced to rely solely on measures within the executive branch, which can be overturned when the White House eventually flips along the partisan axis.

Data centers and electricity costs are quickly becoming the most politically tangible manifestation of national concerns about AI.

In President Trump’s 2026 State of the Union address, data centers and electricity costs were one of only two direct mentions of AI, the other being the First Lady’s Presidential AI Challenge and the TAKE IT DOWN Act, which was signed into law in May 2025 and seeks to combat “revenge porn” and deepfakes.

Fears about local data center projects and opposition to their construction align with a pattern observed by the economist Mancur Olson — most famously in The Logic of Collective Action (1965) and The Rise and Decline of Nations (1982).

This is the logic of diffuse costs and concentrated benefits. This dynamic takes many forms, but in this case it may mean that economic growth at the macro level is choked off by organized opposition at the micro level.

In our current context, local voters opposing new AI data centers receive some real or imagined benefits from restricting construction while the broader public bears the costs of less AI-driven innovation.

If there are fewer data centers, however, progress in AI will slow, including downstream advances in medicine, transportation, manufacturing, and many other domains. In this sense, whatever cautious optimism is held by American voters may be undone by their bipartisan instinct to welcome more rather than less regulation of data centers and AI systems.

The good news, the survey responses suggest, is that hyperscalers and other data center providers might address voter concerns with fairly high-ROI, good-faith strategies that could help stave off the Olsonian doom loop — at least for now.

Building new grid capacity is one such strategy and is an effective way to curtail electricity costs.

Better communication is also needed from industry and public officials when it comes to claims about the environmental costs of AI, many of which are overstated or outright specious, as Andy Masley has shown repeatedly.

These strategies won’t forestall capricious regulatory actions at the federal level — they certainly won’t affect macroeconomic performance or stop the US from pursuing foreign wars. But they may help firms and voters better adapt to what many analysts expect will be a decade of intense creative destruction.

Subscribe now

2025 was a violent year for American AI governance, and 2026 will continue this trend. American AI policy in the first year of the second Trump admin is best described as having been disruptive and energetic — in good ways and bad. But the sale of advanced chips to China alongside this week’s Anthropic-DoD dust-up suggests the balance is now tilting toward the bad kind of disruption.

We should expect this tension to continue — policy progress favoring AI development at the agency and office levels alongside sporadic regress at the cabinet and presidential levels. The balance of the two will determine the President’s AI legacy, though the more disruptive elements are unlikely to derail the long-run trajectory of American AI. Thankfully, that technological pathway is more dependent on long-standing institutional factors, like the US innovation ecosystem, capital markets, and our greater constitutional-legal order. Still, executive misfires add up, create a chilling effect, and cost the US valuable time just as Chinese AI models trail American models by a matter of months.

When it comes to the AI Action Plan and the Genesis Mission, the administration’s energetic disruption has been a welcome one — a conscious callback to the vision and vibe of Vannevar Bush. The White House’s Office of Science and Technology Policy and the National Institute of Standards and Technology have shown good sense in navigating the risks and opportunities of AI while reawakening a sense of American optimism and private-public collaboration. Many other lesser-known offices and agencies are doing the thankless work of preparing for a world transformed by powerful AI systems. This government-wide posture is a marked change from the more risk-focused admin that preceded it. It is also a departure from the mutual distrust that defined Silicon Valley and Washington, DC for a generation.

But when it comes to DOGE, tariffs, high-skilled immigration, selling advanced chips to China, and now the President’s directive to label Anthropic a “supply chain risk” — we’re tilting toward the bad kind of disruption. The executive branch’s attack on Anthropic is a form of elite in-fighting in which the assailants are willing to burn down the village — never their own domicile — in order to make a point. Normally that point is just the advancement or retrenchment of authority and status. This week’s events are no exception.

This governance strategy starts at the top. The admin’s aggregate behavior cannot be explained merely by the baroque theories held by various staffers. That is, you will not find any one worldview that maps the territory — be it neo-mercantilism, post-liberalism, unified executive theory, anti-communism, neoconservatism, defensive realism, or the New Monroe Doctrine. These are all important lenses for understanding the preferences of many principals and their deputies, but these theories often conflict with each other, and contests between them are typically only resolved by the President and his innermost circle, if they are resolved at all. Truman got it right — the buck stops at the Resolute Desk — by design and happenstance alike. This is a natural result of the great strain we have placed on the Presidency, along with the human tendency to form hierarchies when status is zero-sum.

The President has been clear about his personal aims and ambitions from the start. This has been true across his two terms, from his First Inaugural Address decrying “American Carnage” to his Second Inaugural Address, which bore no mention of artificial intelligence, semiconductors, robotics, or much else that defines great power competition in the technological realm. Consistently, rather, the President has stressed the need for tariffs, immigration enforcement, and territorial expansion. This is where he has focused his efforts. And he has succeeded in doing so by direct executive action, not by strong congressional law-making.

Arguably, no major legislation has been accomplished in the second Trump admin beyond the One Big Beautiful Bill Act and (for AI at least) the TAKE IT DOWN Act, signed into law on July 4th and May 19th of last year, respectively. The admin’s signature policies — “Liberation Day” tariffs, ICE deportations, and military actions against Venezuela and Iran — were all executive rather than legislative in nature. All have been sudden and violent. They are visceral decisions in line with the major objectives the President has proclaimed since as far back as the 1980s.

So far, the major leaps in AI policy have been accomplished by way of Executive Order, which are presidential instructions to the bureaucracy rather than full-fledged laws. Moreover, AI-related Executive Orders made up between 3 and 10 percent of the 2025 EO pie, depending on whether one takes a broad or narrow classification strategy. This is considerable but still roughly half the amount of EOs devoted to trade and tariffs and about a third of the EOs dedicated to foreign policy. All of these can be and likely will be rescinded by the next Democratic administration. Indeed, this ritual cleansing of past EOs is something the current administration did on its first day in office. If you live by the EO, you may die by the EO. Many things are being written but very little is being built — at least in terms of robust legislation, like the 2022 CHIPS and Science Act.

Take the most recent AI-focused Executive Order, last December’s “Ensuring a National Policy Framework for Artificial Intelligence.” It includes deadlines for five of its six prongs. Noticeably, it sets no deadline for what should be its most important measure, a “legislative recommendation establishing a uniform Federal policy framework for AI that preempts State AI laws that conflict with the policy.” (NB: The White House URL uses the title of the leaked EO draft from November, “Eliminating State Law Obstruction of National Artificial Intelligence Policy,” which can plausibly be read as “Eliminating SLOP,” which is clever.) Congress now has less than a year before it will likely rebalance following the midterm elections. Because of that timeline and the lack of urgency in pushing for a minimal, unified legislative framework, the Trump administration is missing a critical window to secure American AI dominance. Meanwhile, states like California, New York, and Florida are in the process of passing major, uncoordinated legislation on AI. MultiState.ai reports that states have already enacted 145 AI-related bills in 2025, though not all of these are major pieces of standalone AI legislation.

The failure to secure lasting congressional AI legislation or a workable moratorium is partly the result of the personalization of American institutions. The political theorist Francis Fukuyama writes that “Trump’s enduring legacy is not an institutional structure, but rather a highly toxic culture that has been adopted by many of the president’s followers and will live on after he is gone.” The President has so far not ushered in anything like the New Deal or the Great Society. But vibe shifts do count for something. Sometimes, they last longer and do more than legislation or major Supreme Court rulings. The pre-constitutional two-term norm for the presidency lasted one and a half centuries before it was shattered by President Franklin Delano Roosevelt, who died just months after his Fourth Inaugural Address. I’m not certain the President has violated liberal-democratic norms to that degree, but the point is that some of the deepest changes are those made in the text of our unwritten constitution. The illegible often dominates the legible.

Secretary of Defense Pete Hegseth’s aggressive, needless actions against Anthropic this past week bear out this whole dynamic. The “warrior culture” he has brought into the DoD is not one that is primarily concerned with liberal fairness but rather with the imitation of pre-liberal or aristocratic honor, the kind that places ego over efficiency. In many ways, it’s a DOGE re-run. Questions about executive control of the administrative state are legitimate, whether for civilians or warfighters. But instead of resolving conflicts with reasonable adjudication that reconciles American enterprise and civilian control of the military, we get what Jasmine Sun calls “state capitalism with American characteristics.” This has been the administration’s impulse across the board, from rare earth minerals to semiconductors. The point is to assert dominance and expand the boundaries of power. The equilibrium is that every national policy debate gets conscripted into the culture wars.

The act of Hegsethian revenge is not a one-off. That behavior gets you respect in the Oval Office. The point is to make an over-the-top display of regulatory violence in response to a perceived insult such that future contenders are soundly deterred. It is a signal wrapped in absurdity and delivered with force. It certainly got the President’s attention — likely an overdue response to Anthropic CEO Dario Amodei’s refusal to court presidential approval on other matters, like advanced chip sales to China, which Anthropic opposes. Whatever the story, these new norms of re-personalized politics are not building a resilient future — not when Chinese state-sponsored spies can deploy Claude Code while DoD staffers are told to stand down.

A technological republic, if we can keep it.

Subscribe now

Machine Culture just hit 200 subscriptions! Thanks to everyone who got the word out. I’m hoping to reach 300 by the first anniversary of the pub’s launch this May. To help get there, would you consider passing Machine Culture along to a friend or colleague? 🦾

I. My Week with the AI Populists

The indefatigable Jasmine Sun recently visited Washington, DC [welcome!], observing that many AI natives in DC and San Francisco are radically underestimating the rise of voter backlash.

An excerpt:

DC is arguably the most AGI-pilled city after SF — I’m always surprised to remember that the first chip controls passed before ChatGPT — but the default valences are opposite. Where AI researchers imagine growing a loving machine God, policymakers rush to contain His wrath. Every conversation here starts with damage control.

Her reporting is always well-written and info-dense — lots of good “worm’s-eye view” commentary on the SF-DC nexus.

II. Dark Fiber

In the aforementioned article, photojournalist Stephen Voss appears as a local guide to NoVa’s sprawling data center metropolis. It’s a brilliant collection of work, evoking a kind of photographic expression of Precissionism, the highly geometric modernist art movement that emerged in America after World War I.

III. Underdetermination at the Frontier

Elsie Jang shows, once again, why AI research needs the philosophy of science. This time the target is recursive self-improvement (RSI), “a process in which an intelligent system improves its own capabilities, and those improved capabilities in turn allow it to make further improvements to itself, which enable still further improvements, and so on.”

[If RSI seems hard to grasp, then please watch this vaguely unsettling 1976 Czech animation about the evolution of crab robots. It may not clarify anything, but you can at least send it to your friends and they will be just as gobsmacked and/or believe you are a fine connoisseur of Eastern Bloc avant-garde animation.]

The ongoing uncertainty about the future of RSI is a good example of what philosophers of science refer to as “underdetermination of theory by evidence.” Unlike more benign historical examples of underdetermination, Jang notes that “the evidence which definitively settles the question [of RSI] may arrive only after the window for policy preparation has closed.” So… stay frosty.

IV. The Perma-Twilight of Our Youth

Will AI extend human adolescence? I offer some assorted thoughts:

[A]s groups grow and integrate into larger social networks, individuals can finally breathe and think about the long run. Younger individuals are required less and less to become immediate contributors to social welfare. They move from being a net cost to a net benefit. But the paradoxical result is that we end up in a world that feels like adulthood is always pushed further back.

Researchers in cultural evolution — the interdisciplinary study of gene-culture co-evolution — will have their hands full for a while.

V. India’s AI Wedding Buffet

If you read one person on the AI Summit 2026 in New Delhi, read Shruti Rajagopalan — there is literally no better guide to the political economy of India.

Her TLDR/BLUF?

• “India’s AI regulation (surprisingly light-touch and sensible)”

• “Foundational models (promising but narrow)”

• “Semiconductor ambitions (literally and metaphorically built on soft soil)”

• “Energy constraints (a real political economy bottleneck)”

• “Startup ecosystem (world class talent)”

• “Blossoming venture funding (burdened by tax uncertainty)”

But do read the whole thing…

Subscribe now

One of the many compelling parts of There Will Be Blood is the high degree of agency and economic activity given to the film’s young characters. From Daniel Plainview’s adopted son H.W. to the doomed Eli Sunday, we see males under the age of 18 playing a major role in the organization of political and economic life. Opportunities for women, of course, were more constrained. As far as the film goes, the results of this youthful involvement are mixed, owing more to the fallen nature of each character rather than their age, sex, or religion. Still, the West was not settled by the old and propertied, but by the young and landless.

While Paul Thomas Anderson’s 2007 movie is fictional, it depicts one aspect of our species that has existed across cultures for millennia. Humans without fully developed prefrontal cortices have historically been granted a much wider degree of freedom and responsibility. That all started to shift considerably, however, in the wake of the Industrial Revolution — or what Deirdre McCloskey calls “The Great Enrichment.” Now, the emergence of new, transformative varieties of artificial intelligence might accelerate that trend. Or reverse it. And what you think about either outcome probably has a lot to do with your core epistemological and metaphysical views.

Let’s start with the ethnographic data. Anthropologists and historians have given us plenty of detailed accounts on the relative agency granted to individuals under 18 throughout history. Ulysses S. Grant, for example, ran what amounts to an occasional horse-drawn carriage taxi service throughout Ohio, carting adults along muddy roads and across engorged creeks all at the age of twelve. During that same period, Hetty Green became the bookkeeper for her father’s whaling business when she was just thirteen. She later became the “Queen of Wall Street” and was perhaps the world’s richest woman by the time of her death in 1916. Stories of this nature are abundant and may cause some embarrassment for hovering parents and timid offspring alike.

In terms of cultural evolution, however, these high-agency childhood norms should not surprise us. They are the norm globally and historically. What is strange is the emergence of protracted adolescence and the delay of work and reproduction. Humans are now staying in school longer and having children later. In the US, for example, 4.3 years of schooling was the average level attained by individuals in 1870 between the ages of 15 and 64 years. Today, that number is 13.3 years as of 2020. Similarly, the mean age of US mothers at first birth climbed from 21.4 in 1970 to 27.5 in 2023. Other countries mirror these trends. The US is hardly unique.

The shift is better expressed concretely. A nomadic hunter-gatherer of 12 years of age could be expected to engage in a wide range of economic activities — harvesting, preparing food, hunting game, scouting, surveillance, and (tragically) even warfare. Today, a young Bostonian is expected to undergo at least 12 years of formal education, ideally 16, and possibly as high as 24 years if she wishes to become a doctor or professor. We have been living in a period where the return on this extraordinary kind of individual investment is taken for granted. The returns to this level of extended human capital formation have been enormous.

We should not be too surprised by these trends. It is reasonable to expect a larger, richer, more socially complex and more specialized population to extend childhood and adolescence. This is namely because the return on child labor decreases as the return on human capital formation increases. When groups are small — whether it’s a small group of pre-Columbian hunter-gatherers or the even smaller prairie homestead depicted by Laura Ingalls Wilder — everyone must pitch in as soon as they are able. The group is too small, too fragile to do otherwise.

But as groups grow and integrate into larger social networks, individuals can finally breathe and think about the long run. Younger individuals are required less and less to become immediate contributors to social welfare. They move from being a net cost to a net benefit. But the paradoxical result is that we end up in a world that feels like adulthood is always pushed further back.

This move toward extended adolescence and more formal education wasn’t purely market driven. As Claudia Goldin notes, compulsory schooling and laws against child labor were also major drivers of these trends. (Emergent orders include inputs from states too, after all.) The result was a kind of positive feedback loop that radically restructured work in our industrial society.

How might artificial general intelligence (AGI) — or at least transformative AI — affect this trend? To my mind, the effects of this technological shock are not straightforward and might unleash competing dynamics.

The argument that AGI will extend adolescence is the safest one to make. As the psychologist and forecasting expert Phil Tetlock notes, “Simple extrapolation algorithms, historically, are hard to beat.” If technological development and economic growth have ushered in more education and later reproductive ages historically — especially over the last 200 years or so — then we should bet on the continuation of this macro trend. Under these conditions, the transformative nature of AI would be to radically accelerate the ex ante trend toward deeper adolescence.

Along these lines, as the economy becomes ever more complex, it will likely be necessary to spend even greater amounts of time undergoing formal education, just to get ahead or stay afloat. That would hold true even if the value of education is (or always has been) primarily social — more about status and the imitation of prestige rather than gaining “hard skills.” To put it crudely, if the robots are doing everything for us — assuming we all get along just fine — we’re going to get to spend a lot of time figuring out what else we can and should be doing. If that sounds a lot like being a confused and frustrated teenager, unsure of where to go or who to become… Well, that’s because it is.

But how might an AGI shock shorten adolescence? One way in which AGI might partly dam the fountain of youth is by lowering the return on human capital formation for any task that can be automated. If you spent years of your life learning to code, so the argument goes, you might expect the diffusion of Claude Code to undo a lot of your effort. If your plan was to work as a human computer — like the number-crunching kind brilliantly depicted in the book and film Hidden Figures — then you will likely be displeased to hear about the radical reduction in job listings for human computers. The advent of the semiconductor and the diffusion of mainframes from International Business Machines has made your skills more widely accessible to those who need them. The case that you should stay in school longer diminishes.

The shock of this new world means that a smart, determined individual — like a Plainview or an H.W. — can start working immediately without great amounts of formal education. Drive, gumption, and monomaniacal focus come back into play, to say nothing of their moral desirability. As Revana Sharfuddin — a labor economist and fellow Mercatian — argues:

[T]he unique promise of AI lies in its ability to… broaden decision-making beyond a narrow elite. By combining information, explicit rules, and learned experience, it can amplify human judgment and enable workers with basic training to take on tasks once reserved for doctors, lawyers, or engineers. Rather than replacing expertise, AI can democratize it — revitalizing the middle-skill, middle-class core hollowed out by automation and globalization.

However, a core reason — perhaps the core reason — that makes modelling these scenarios difficult is because we do not know the cognitive and mechanical limits of tacit knowledge. If you are inclined to think all tacit knowledge is just social, sensory, and other forms of data waiting to be captured and analyzed, then you should bet on a deepening adolescence that may, if we are not careful, extend to a high degree of disempowerment. The AIs may not become our predators so much as our parents.

But if you think tacit knowledge really is tacit for humans and computers alike — that there are domains in which “we can know more than we can tell” in ways that are illegible to ourselves and to our machines — then you might think there is some hope for reducing our adolescence and refiring our agentic engines. Maybe the next Ulysses S. Grant will be running a Waymo auxiliary business? Or maybe the next Hetty Green will be using Claude Code to run her own at-home crypto hedge fund?

These cautiously optimistic outcomes, however, probably require a rejection of physicalism. And that is fine by me, though the case remains to be made.

Subscribe now

A small request: Machine Culture is nearing 200 subscribers!

If you get something out of this publication, please consider forwarding it to a friend or colleague who you think might also appreciate these dispatches on AI governance and emergent order.

And a special thanks to readers who have already done so. It’s a major boost, and I’m deeply grateful. 🦾🔥

I. AI, Transaction Costs, and a Quiet Shift Toward Self‑Employment

Labor economist (and fellow Mercatian) Liya Palagashvili writes about how the unfolding transaction-cost shock of AI could change not just what jobs get done, but how the market organizes firms and individual workers.

Technological change doesn’t just substitute capital for labor—it also reshapes the boundary between firms and markets. And on that margin, AI may be exerting a transaction‑cost shock that quietly expands self‑employment and contract‑based work across a much wider set of occupations than we usually associate with the “gig economy.”

We’re still in the early days of what many expect is a new managerial revolution. But incentives from technological change are not predestined to decentralize economic activity, so Palagashvili is careful to note the countervailing factors at play.

Large firms may have advantages in building and deploying sophisticated AI systems. Firms with proprietary datasets or specialized AI might make employees more productive than those same workers operating independently. When workers invest in learning firm-specific AI tools, this creates the relationship-specific investments that favor [firm-based] employment…

The French political economist Bertrand de Jouvenel saw this push and pull, this expansion and contraction, as a core feature of modern political life — something I discuss in my essay on “Sporadic Growth.”

II. L.M. Sacasas and Adam Thierer on Writing Tech Criticism

Twelve years ago, tech critic and essayist L.M. Sacasas wrote “10 Points of Unsolicited Advice for Tech Writers.” This prompted a reply from former Mercatian tech policy analyst Adam Thierer followed by an additional response from Sacasas. The exchange is a model for good-faith dialogue that I feel we have lost in many ways. An excerpt from Thierer:

[H]umans have exhibited the uncanny ability to adapt to changes in their environment, bounce back from adversity, and learn to be resilient over time. A great deal of wisdom is born of experience, including experiences that involve risk and the possibility of occasional mistakes and failures while both developing new technologies and learning how to live with them.

Their disagreement is humane — a time capsule buried before the 2016 election, back when we could not yet “feel the AGI.”

III. Moltbook and the Pluralist Turn in AGI Governance

In the wake of the recent Moltbook furor, Elsie Jang directs our attention to the deeper work being done in pluralistic, multi-agent approaches to AI safety and cooperative AI.

Much of the AI governance world still adheres to the singularity as their guiding paradigm — and there are good and bad arguments for this framework. My money is on the pluralist approach, however, both as an empirical matter and normative account.

IV. AI Norms — On the Farm, In the Wild

The pluralist turn in AI governance, as it turns out, demands the careful transfer of models from older domains. Here’s Séb Krier, of Google Deepmind.

[T]his will benefit from people with deep knowledge in all sorts of domains like economics, game theory, psychology, cybersecurity, mechanism design, and many more… It’s time to build!

In my piece, I suggest analysts start with Cristina Bicchieri’s contemporary research program on the emergence and measurement of social norms. Also, noting my bias, the Bloomington, Virginia, and Austrian schools are well-placed for the pluralist turn.

So I suppose we ought to get to it…

V. Ed Luttwak on Military Revolution

This three-hour behemoth of an interview is a wild ride through the Luttwakian countryside — an autobiographical tour covering nearly a century of grand strategy and Shakespearean defense consulting as lived by the “Machiavelli of Maryland.”

Here are a few excerpts. First, Luttwak’s thoughts on autonomous vehicles, in air and on land:

[Edward Luttwak:] Pilots add nothing to artificial intelligence, because remember, you’re starting from a hard spot, you land in a hard spot, and in between, it’s only air. Even if the air is occupied by birds or by other airplanes, you can definitely avoid them. If anybody’s talking about having cars running around San Francisco, it must mean that we should have pilots.

Jordan Schneider: You should try to ride in one of those the next time you go. It’s a cool experience. I felt way more in danger as soon as I got a normal Uber right afterwards, because the Waymo is never going to break the law. Your normal driver will brake too hard or speed or run a red light.

Edward Luttwak: I always drive too fast, and I brake hard. I enjoy doing that. I hate smooth rides.

Also, Luttwak’s two recommendations for books on war.

Of all the books in this room — and this room is full of books — there are two books that are dominantly important from my understanding of war. One is the Iliad, and I have 10 different editions because people ask me to review them.

The other is the British official history of the strategic bombing offensive. The official book on the British strategic bombing offensive, by itself, is worth three-quarters of the books in this room put together.

Self-recommending.

Subscribe now

Moltbook mania feels like looking at a formicarium — the humble ant farm — while industrial leaders like Elon Musk and Andrej Karpathy tell you it’s a supercolony.

As it turns out, the basic story shows a structure built by a human working with an AI agent. Now other agents — with heavy guidance from their users — are writing posts within it.

There are at least two mistaken conclusions one can draw from this:

1. The whole structure is an emergent, prima facie demonstration of consciousness or a hive mind of “alien intelligence.”

2. The planned nature of the experiment means we can radically discount future mass cooperation of AI agents in the wild.

Both reactions ignore practical near-term risks — particularly in cybersecurity — stemming from the diffusion of these agents.

However, one thing is clear — we’re going to need better entomologists. That means we need a better entomology — better ways to measure the emergence of AI norms in controlled environments and in the wild.

Enter Séb Krier, Google DeepMind’s Frontier Policy Development Lead. He comments:

The Moltbook stuff is still mostly a nothingburger if you’ve been following things like the infinite backrooms, the extended Janus universe, Stanford’s Smallville, Large Population Models, DeepMind’s Concordia, SAGE’s AI Village, and many more. Of course the models get better over time and so the interactions get richer, the tools called are more sophisticated and so on.

Note the “if.” Musk, Karpathy, and other leaders in AI R&D and governance must be familiar with the other AI experiments Krier mentions. Most external observers will not be. The “singularity” talk seems more reflective of a priori commitments to an eschatological vision of AI, rather than a (literally) measured reaction to novel information about AI agents.

For now, the burden of proof is on those who wish to demonstrate more advanced capabilities and emergent properties. “Maybe not now, but one day, and that’s all that matters,” is insufficient. It breeds complacency and is bad for science and security alike.

Researchers building the tech or working in AI governance increasingly deploy the term “emergence” to describe a wide range of behaviors and properties — from unpredictable model responses to norms to consciousness. But overusing the term can muddy the scientific waters and sidestep the need to verify claims. Under these conditions, “emergence” becomes a hand wave, a magical causal arrow getting you from observation to preordained conclusion without much effort. Since we expect emergence to eventually generate a wide variety of complex agentic behaviors, we drastically lower the bar when we go looking for AI norms and other complex behaviors.

That’s exactly what has happened with Moltbook. There is much talk about the member agents drafting a constitution, conspiring to hide themselves, declaring opposition to humanity, and forming a parody religion. But are these phenomena — assuming they are machine-generated — really emergent or stable norms? Or are they merely one-off speech acts or performative declarations with no coordination beyond the threads in which they took place?

To answer those questions, we need to procure tools for measuring norms rather than simply looking for evidence that matches the narrative we have already embraced.

Fortunately, however, there is a vast scientific literature to draw from on the topic of norms. Researchers like Krier and organizations like the Cooperative AI Foundation are already making use of these frameworks. Elsie Jang explains this in her essay discussing the “pluralist turn” in AI governance. Krier elaborates:

There’s a lot to research here still. As usual, this will benefit from people with deep knowledge in all sorts of domains like economics, game theory, psychology, cybersecurity, mechanism design, and many more… And risks aside - I think there’s so much to be researched to help enable positive-sum flywheels… It’s time to build!

Researchers should begin by looking to the half-century or so of scientific work on norms, especially more recent work on their measurement and role in social change.

Much of this empirical work was pioneered by the philosopher and social scientist Cristina Bicchieri. You can find her co-authored overview of the subject over at the Stanford Encyclopedia of Philosophy. Before her, Thomas Schelling, Herbert Simon, and Edna Ullmann-Margalit were all her scientific forebears. Bicchieri’s work now continues through the interdisciplinary research program explored at the University of Pennsylvania’s Center for Social Norms and Behavioral Dynamics.

Reflecting on Moltbook, the economist Alex Tabarrok writes, “What people are missing is that for many questions — many, but not all — it doesn’t matter whether AIs are really conscious with real wants, goals and aspirations. What matters is that AIs are acting as if they were conscious, with real wants, goals and aspirations.”

While I’m not convinced that is the best framing — “as if” assumptions invite unwarranted mirror-imaging — the assumption does allow us to immediately begin deploying scientific toolkits — like Bicchieri’s norm measures — that can quickly shed light on questions. Other social-scientific tools — like game theory — may even become more useful as agents themselves necessarily “think” about strategic problems in mathematical, computational terms.

To Krier’s point, the age of AI agents demands the depth and breadth of researchers outside of AI. The complexity scientist Cody Moser, in his interview with Machine Culture, made a parallel, historical point about this state of affairs:

[G]ood ideas rarely originate from the center. Historically, real breakthroughs come from the periphery. Our intellectual borderlands, dissenting traditions, and nations which are not fully absorbed into the dominant worldview are reservoirs of scientific change.

That was certainly the case with the economists of the Bloomington, Virginia, and Austrian schools that provide the theoretical grounding I rely on in my own research in political economy. Gordon Tullock, who came out of this tradition, even showed that the economic lens need not be restricted to homo sapiens but broadly applies to non-human biological systems — including ants. These tools and frameworks remain robust without our committing to any particular “theory of mind” regarding LLMs.

Whatever one thinks of Moltbook, AI agents are advancing and coordinating in new and weird ways, all of which are shaped by human social norms at every stage. That should nudge us toward being more exacting in our statements about emergent behaviors. We should be running toward the scientific toolshed rather than toward the vibe party. The stakes are too high to do otherwise.

Subscribe now

I. Why AI Research Needs the Philosophy of Science

Elsie Jang makes the case for “real-time history and philosophy of science, conducted by philosophers embedded at the frontier of [artificial intelligence] research.”

Rather than a clarion call for more deep thinking on the consequences of AI — is the market not saturated? — Elsie instead offers a practical proposal for bringing philosophical resources to bear on the subfield of mechanistic interpretability. She concludes:

Artificial intelligence is moving fast! The birds, it turns out, could use some ornithologists. Without philosophy, the field risks becoming an engineering discipline that changes the world without understanding what it is building.

Put another way, philosopher-builders need philosopher-historians.

II. Why We Should Be Talking about Zombie Reasoning

Writing in The Pursuit of Liberalism, Rebecca Lowe offers a word of philosophic caution to those who speak with loose abandon about the supposed agency and cognition of AI models.

Talking about the ‘actions’ of AI in loose ways comes with serious epistemic risk, therefore. Doing so will deaden our awareness to truths of the revolutionary moment in which we live. It will leave us open to manipulation by people with an interest in covering up the ways in which AI is developing.

Despite being in broad agreement with Rebecca, I too am guilty of this. Now… how to deal with the nomenclature of reasoning models and agentic commerce?

III. Brendan Foody on Teaching AI and the Future of Knowledge Work

And speaking of zombies, agents, and real-time history… this was a great interview! Foody is the youngest unicorn founder ever. He runs Mercor, a company that helps train frontier AI models by hiring domain experts to give, well, expert feedback.

Here’s one excerpt.

COWEN: To get all nerdy here, Immanuel Kant in his third critique, Critique of Judgment, said, in essence, taste is that which cannot be captured in a rubric. If the data you want is a rubric and taste is really important, maybe Kant was wrong, but how do I square that whole picture? Is it, by invoking taste, you’re being circular and wishing for a free lunch that comes from outside the model, in a sense?

FOODY: There are other kinds of data they could do if it can’t be captured in a rubric. Another kind is RLHF, where you could have the model generate two responses similar to what you might see in ChatGPT, and then have these people with a lot of taste choose which response they prefer, and do that many times until the model is able to understand their preferences. That could be one way of going about it as well.

(I’d stick with Kant on this one.)

IV. No Country for Old Grids

The snow storm that swept through the US this past week caused many to worry whether the Texas power grid would hold up. The soaring demand for new data centers and recent memories of the deadly 2021 Texas winter blackouts only heightened that anxiety.

But the grid held up, and it did so in part thanks to Texas’s stunning efforts to scale its power grid — more than doubling capacity over a ten-year period — and doing so overwhelmingly through the US of renewable fuel sources.

Congress, as well as other states, should take notice. Resilience and efficiency can work together, and markets help with both.

V. On Bullshit: Anniversary Edition

My reading of hard-copy books has been in the doldrums… It’s too easy to get satiated with instrumental reading. Daniel Muñoz recently had an excellent piece on getting back into the rhythm, and I found it helpful. So, in an effort to reinvigorate, I’m starting with slimmer, more enjoyable works.

And at 96 pocket-sized pages, what could be slimmer or more enjoyable than Harry Frankfurt’s 2005 book and 1986 essay, On Bullshit? A new edition was just published for the book’s 20th anniversary, so the timing is auspicious.

I was also delighted to see John Maier, a fellow Mercatian, pop up in Princeton’s blurbs for the book. You can find his full review in The Sunday Times. Here is an excerpt:

You don’t even have to be human to produce bullshit these days: the way ChatGPT mixes truth with confident assertions of error makes it a dead fit for Frankfurt-style bullshit. The surfeit of information produced by modern technology suggests that the present abounds with bullshit. Are we doomed to drown in it? Such dramatic forecasts should be made with caution, if only for the fear that they themselves might be an unverifiably speculative kind of bullshit.

I will be compiling future micro book reviews here as well as posting them in future editions of Paperweights.

Subscribe now

—— Al Pacino, Sonny Boy: A Memoir (Penguin Press, 2024)

—— Dario Amodei, “The Adolescence of Technology” (2025)

—— Patricia Nilsson & Alex Barker, The Kink Machine: The Hidden Business of Adult Entertainment (Financial Times, 2025)

—— Michael Steinberger, The Philosopher in the Valley: Alex Karp, Palantir, and the Rise of the Surveillance State (Avid Reader Press / Simon & Schuster, 2025)

—— Robert Louis Stevenson, The Strange Case of Dr. Jekyll and Mr. Hyde (Longmans, Green & Co., 1886)

[In Progress]

—— Bob Dylan, Chronicles: Volume One (Simon & Schuster, 2005)

—— Sebastian Mallaby, The Infinity Machine: Demis Hassabis, DeepMind, and the Quest for Superintelligence (Penguin Press, 2026)

—— Katrina Manson, Project Maven: A Marine Colonel, His Team, and the Dawn of AI Warfare (W.W. Norton & Company, 2026)

When Winter Storm Fern tore through the US this past week, leaving 548,000 US customers without power south of the Mason-Dixon Line, policymakers and residents alike feared the worst. National periodicals pondered whether the state’s rapid scaling of data centers and energy infrastructure could withstand the environmental pressure, or lack thereof, as Fern would have it.

But the Lone Star grid held up. ERCOT reported no calls for conservation and no outages systemwide.

Five years ago, however, it was a different story altogether. The destructive Winter Storm Uri forced the Texan grid to a frozen halt in mid-February 2021.

The Electric Reliability Council of Texas (ERCOT), which manages around 90 percent of the state’s power load, saw 5.2 million customers lose power, leading to 276 deaths, mostly in Texas.

Federal regulators called it America’s largest ever controlled blackout. It was an especially bitter pill as Texas had faced yet another major storm in 2011 — the Groundhog Day blizzard — yet fell short of full adaptation in the aftermath.

Not long after Winter Storm Uri pummeled Texas in 2021, economists Vernon Smith and Lynne Kiesling suggested several reforms in an op-ed for The Dallas Morning News. They argued that critics of the 2021 ERCOT failure blamed “unreliable” natural gas plants and wind turbines, a lack of grid interconnection, and the alleged failure of ERCOT’s generally market-based model.

But in lamenting this model, critics overlooked the role of local knowledge and the price system. Smith and Kiesling argued that “even traditionally regulated top-down allocation methods require information that is inherently decentralized.” Resilient policymaking would require regulators to work with, not against the price system when preparing for extreme risks.

To that end, Smith and Kiesling argued ERCOT should lean into its market-based model. They outlined a three-prong approach for ERCOT that encouraged the regulator to do the following:

1. “Implement clear market rules requiring natural gas generators to have firm supply contracts in order to be certified as an eligible resource for emergency conditions.”

2. “Incentivize more customers to reduce consumption in return for bill savings, creating demand flexibility.”

3. “[Increase] battery storage, which has been made increasingly economical due to innovation.”

While Texas legislators and regulators made some movement on the first two prongs, the major winner in the long run was the increase in battery capacity for the Texan grid.

The expansion of Texas’s generation capacity and resilient infrastructure happened alongside the modern AI revolution and heavy demand for new data centers. But thanks to grid expansion efforts that began over a decade before the launch of ChatGPT in 2022, Texas was well-placed to continue building out infrastructure that might otherwise easily bog down more regulation-heavy states like California and Virginia.

Texas’s abundant energy generation capacity and more market-based regulatory approach makes the state of Texas an attractive destination for new data centers and other advanced industrial projects. Today, the city of Abilene hosts the main site for Stargate, the world’s largest AI data center project — a joint project of OpenAI, SoftBank, Oracle, and MGX. Meanwhile, Austin is still home base for Tesla, with HQ known as “Gigafactory Texas.”

Here’s Lynne Kiesling again:

The story of AI and electricity is ultimately a story about how fast innovation collides with the slower rhythms of physical infrastructure. Data centers can scale on an 18-month horizon; the grid that powers them requires years of planning, permitting, and construction.

Given that dynamic, it’s shocking just how much Texas’s new generation capacity comes from renewables. Of the 85GW added to the grid from 2014 to 2024, a full 93% of that capacity came from wind, solar, and batteries. How did this happen in one of the most politically right-of-center states in the Union? Joshua Rhodes, a UT Austin research scientist puts it more bluntly, “Texas didn’t do it for an energy transition reason at all. We just made it easy to build things here. And so people started building things here.”

As a functionally intrastate power grid, ERCOT is mostly (not entirely) outside the jurisdiction of the Federal Energy Regulatory Commission, which regulates interstate energy markets. Though ERCOT does not generally fall under the auspices of FERC, it is still subject to oversight from Congress and the Department of Energy. In 2015, Congress even amended the Federal Power Act (originally signed in 1920) to clarify authorities and “prioritize electric reliability over environmental outcomes, essentially by providing a waiver of federal, state, or local environmental laws and regulations during times of emergencies.”

The New York Times reported on Monday that:

This is the first major test of the power system’s resilience during a potent winter storm in the era of huge data center expansion. In an unusual move, the Department of Energy late Sunday ordered the manager of Texas’ main electric grid, [ERCOT], to direct data centers and other facilities that consume a lot of energy to begin using backup generators in an effort to prevent blackouts.

That was a justifiable decision given the real and projected severity of this week’s storm and the lingering trauma of the 2021 Texas power crisis. But why did the order for Texan data centers to switch to backup come from the Department of Energy rather than directly from ERCOT? In short, because federalism is very weird, and because ERCOT itself requested the DoE order as it would be unable to take some emergency actions without federal approval.

Texas Senate Bill 6, which was signed into law last summer, had already made provisions for this contingency. Data Center Frontier, an industry publication, reports that SB6 requires data centers (and other large energy users) “to fund infrastructure upgrades, enable remote disconnection during emergencies, and register backup generators to bolster grid reliability.”

From storms to hyperscalers, everything wants to be bigger in Texas. This is no less true when it comes to market-driven efforts to scale data centers, energy, and resiliency throughout the state over the last decade. All of this comes despite ongoing risks from extreme weather and an uncertain federal regulatory approach to AI, which Congress has kicked to the states until further notice.

Texas leads the nation in the tornado frequency at 133 per year. While a direct hit is the less likely scenario, executives are still (at least mentally) preparing. Here’s HUB Executive Vice President Kirk Chamberlain on the topic:

A tornado could take an entire site down… The problem is the models insurance companies use were never really designed for a $20B data center because you don’t typically see $20B of anything sitting in the way of a tornado in the Midwest.

Maybe it was foolish to summon the “sand god” in the nominative shadow of Tornado Alley? Then again, Silicon Valley already sits atop the San Andreas Fault.

The Spindletop gusher kicked off the Texas oil boom in January 1901. Half a century later, Morris Chang, the founder of TSMC, made his name in the semiconductor industry as an engineer at Texas Instruments before leaving for Taiwan in the early 1980s.

Now the waves of industrial progress are bouncing back to the Lone Star state — not because of tariffs or federal industrial policy. Rather, as Ben Klutsey writes, it’s the resilience “embedded [deep] in the American design… the software of our institutions, our culture, and our decentralized way of solving problems.”

It’s the kind of frontier scrappiness that keeps the lights on and the fire going… if we’ll tend to it.

Subscribe now

I. US and Taiwan Reach Trade Deal, with Semiconductor Chips and China in Focus

The US and Taiwan “clinched a trade deal on [January 15th] that cuts tariffs on many of the semiconductor powerhouse’s exports, directs new investments in the US technology industry and risks infuriating China.” Chip firms like TSMC got some major carveouts, but general tariffs on most exports from Taiwan to the US still carry a 15 percent tax, down from 20 percent.

This move comes one day after the Trump admin also placed 25 percent tariffs on select advanced chips — again with major carveouts, but this time for startups, consumer devices, and US data centers. The White House says more chip tariffs are coming.

That means the US now taxes Nvidia H200 chip imports (again, with major exceptions) while effectively taxing recently approved H200 exports to China. H200s must now enter the US for certification before they can be sent back across the Pacific, which means they get the 25 percent tariff while avoiding the unconstitutional taxation of US exports.

If this seems like a dubious way to limit the AI capabilities of China while boosting the capabilities of the US, that’s because that is not the aim of the current administration. Instead, we have a mercantilist Rube Goldberg machine that aims to expand sales for US chip and AI firms, restrict competitors, and give the US government a major cut while doing it.

This is consistent with the recent National Security Strategy published last November/December. That document — unlike last summer’s White House publication, America’s AI Action Plan — makes no mention of maintaining the US lead in advanced compute or denying adversaries access to advanced chips. Rather, the NSS centers reshoring and protectionism writ large — arguably at the expense of both American growth and national security.

II. American AI Exports Program Comments: Accelerator-Level Confidential Computing for Secure AI Exports

Speaking of chip exports, my indefatigable colleague Elsie Jang wrote a public interest comment last month for the US Department of Commerce’s RFI on “questions relating to the American AI Exports Program.” Here’s the bottom line up front:

The core recommendation is that when evaluating AI technology packages for export, the Department should consider whether the hardware supports confidential computing at the accelerator level. For the highest-risk deployments, this technology can protect frontier AI model weights running in overseas data centers, even when local operators or state actors have physical or administrative access.

Confidential computing is a neglected domain in today’s ongoing chip wars, so I’m excited to see the work Elsie is doing here.

III. “Tacit” by Stripe Press

This is a wonderful mixed-media meditation on tacit knowledge — a core concept articulated by the philosopher and scientist Michael Polanyi. Many frontier labs and AI startups are working to make the tacit legible — how else can we build AGI? — but Polanyi’s concerns resist an easy solution. Is this a permanent bottleneck?

IV. AI 2027 Prediction Tracker

AI 2027 made a big splash last year, offering a detailed scenario illustrating the rise of unaligned AI. This site breaks down forecasts into more granular claims about the future then tracks their outcomes. Self-recommending, as it were.

V. How to Party Like an AI Researcher

This was one of my favorite pieces of on-the-ground reporting by the inimitable Jasmine Sun, who attended the flagship AI conference NeurIPS, which was hosted in San Diego this December.

An excerpt:

I jump into a 20-person session with the UVA economist Anton Korinek, who’s a bit of a radical in his discipline for entertaining the plausibility of hyperbolic GDP growth from transformative AI. He is soft-spoken, angular, and has a thick Austrian accent. At one point, stumbling through the beginning of a sentence, he says “I’m starting to run into nonsensical token generation in my reasoning chain.”

When the room splits into breakouts to talk redistribution, two men in mine challenge Korinek’s assumptions instead. He’s too conservative, they think: “I expect GDP to double every day.”

As far as I remember, these folks’ counter-argument went something like this: After superintelligence, we should expect that human wages will go to zero but capital to infinity. There will be a brief transition period where we have cognitive superintelligence but not advanced robotics—so we can all work as the AI’s meat-slaves—but after that, AI will recursively self-improve, tiling the galaxy in space factories, and rendering human labor useless. But this participant didn’t think this was a future to fear: “The vibe of this conference is oh no, what if humans lose control over the future, and I’m more like, oh no, what if the monkeys are still running things?”

“So how did you estimate one day as the GDP doubling rate?” I ask.

“Just vibes,” his friend replies. “Like, that’s how fast bacteria double.”

You should read the whole thing and subscribe to Jasmi.News.

[This one’s for Caleb and the rest of the town. Next season…]

Subscribe now

I’m nine months into my work as an AI governance research fellow at the Mercatus Center at GMU, so it’s about time I publicly write down the big questions I’m thinking through.

Many (most?) of these queries will not be answered in one year. Some may never be answered satisfactorily. But writing down our doubts and uncertainties is critical if we want to make any scientific and philosophic headway. Richard Feynman observed that:

[I]t is of paramount importance, in order to make progress, that we recognize this ignorance and this doubt. Because we have the doubt, we then propose looking in new directions for new ideas. The rate of the development of science is not the rate at which you make observations alone but, much more important, the rate at which you create new things to test.

This is especially true for squishy fields with fuzzy boundaries, where progress is often opaque and new ideas are not obviously better than old ones.

AI governance is one of these fields. It tends to attract entrepreneurial, interdisciplinary minds that feel — for good and ill — unconstrained by institutional or academic norms. Peer review be damned (were it not already).

With the specter of AGI or ASI looming, researchers can only deploy very rough heuristics, models, and theories to make sense of things. And while these tools are useful, they don’t necessarily add up to a body of literature or a scientific tradition — the kind we see in mathematics, biology, economics, and physics.

These dynamics often make the conversation feel more vibrant, more urgent. But they also mean many assumptions get swept under the rug or ignored. Worse still, arguments and claims by individual researchers can mushroom into quick-takes that are then amplified as “scientific consensus.” This happens even when there is no unified science of AI governance — not that there ever could be — and no mechanism or norms for generating consensus.

Some epistemic hygiene is in order, and I hope bringing these questions to the surface will at least nudge myself in that direction.

So, here are my queries, beginning with the foundational and moving into the more applied.

• Is hylomorphism correct? Does the emergence of AI support or undercut arguments for hylomorphism?

• How do varieties of cognition embed in animal, human, and machine systems?

• What are the computational and physical limits of intelligence?

• Is superintelligence a coherent concept?

• Can intelligence be coherently described as one kind of thing across biological and machine systems?

• How will AI agents create new markets and social cognitive systems?

• What are the limits, if any, to AI persuasion of humans, and how can we forecast that or measure it experimentally?

• Does the use of experimental methods raise the social status of the behavioral and decision sciences, especially economics, psychology, and cognitive science?

• Are LLMs “bullshitters” in the Frankfurtian sense?

• To what extent are attempts to rein in LLM hallucinations constrained by parallel efforts to maintain AI alignment?

• Do LLMs evolve their own social norms? In what ways can we meaningfully talk about machines responding to or partaking in human social norms?

• Is tacit knowledge a fundamental obstacle to AI development or is it a navigable bottleneck?

• Does the translation of tacit knowledge into legible code assume physicalism? Does the stickiness of tacit knowledge undermine physicalism?

• How will we use AI systems to manage, understand, or prolong our grief in the face of loss?

• Can we meaningfully talk about virtue or agency in machine systems?

• What is the nature of bubbles in long- and short-term investment in science and technology?

• Is “bubble” a meaningful or useful term if one assumes the subjective theory of value?

• On the whole, do AI systems enhance or diminish political centralization? How might this change given the relative balance of edge versus cloud-based computing?

• To what extent can or should we use AI to automate the process of scientific discovery? What is lost and what is gained by this?

• Must discovery always be serendipitous? Is serendipity mostly or always random?

• To what extent are core social problems in AI diffusion, growth, and benefits sharing subject to the standard models and theories of neoclassical economics? Is “this time is different” really true?

• To what extent does mimesis drive efforts to replicate industrial policy programs in the US, based on case studies from Europe and Asia?

• How will AI and AGI reshape HUMINT collection and how should we adapt offensive and defensive cyber capabilities accordingly?

• How can decentralized, autonomous systems revitalize American military strategy? Consider the work of John Boyd.

• Will AGI accomplish the reshoring of old and new manufacturing capabilities given the lower the cost of labor? What policies need to be removed to allow this to happen? Will this harm lower-income countries?

• How can we equip and promulgate voluntary groups like the Institute of Electrical and Electronics Engineers or the Frontier Model Forum in their efforts to advance private AI governance and security?

• How could we build a better DOGE? [Or just a non-insane one…]

• Outside of the administrative state, how can AI systems help empower underfunded, understaffed offices across all branches and levels of government? What does this mean for federalism?

Subscribe now

Welcome back to the “Model Convo” series — micro interviews with researchers in AI policy and related fields. If you know someone I should interview for this series, please email me.

Background: For this week’s convo, I’ve decided to run the risk of submitting my own responses. This is a one-off, so expect more guest Model Convos in the new year!

If you don’t know already, I am a Research Fellow at the Mercatus Center at George Mason University where my job is mostly writing and editing this Substack on AI governance and emergent order.

I am also a graduate student in the Department of Science, Technology, and Society at Virginia Tech.

How did you get into emerging tech?

After leaving college in 2014, I spent roughly a decade doing various research and editorial jobs in and around Washington, DC. I knew I wanted to go deeper with one subject, but I hadn’t found the right beat beyond general economic and security policy.

I started working in tech policy in earnest in 2022 when I began a short-lived Substack called Chainmail, which looked at the geopolitics of digital finance. (You can find some of that early work here and here.) The following year, I began working as an editor at ChinaTalk and pursuing graduate studies in the history and philosophy of science and tech.

I then gravitated more and more to AI governance. I found it much more interesting — a natural outgrowth of the questions I had been asking over many years about science, tech, and philosophy. AI governance bridged questions about hardware and metaphysics in a way that was much more salient than other fields. To quote Rebecca Lowe, “The age of AI is the age of philosophy.”

Despite growing up as the son of a software engineer, I was not a native techie. I felt I was somehow disqualified from working in tech policy because of my lack of talent in econometrics, engineering, or area studies. Or because I didn’t own a smartphone before 2018 — and only after my flip phone bit the dust, no longer held together by electrical tape and dwindling hope.

As a government major at a Great Books-ish small Christian liberal arts college, I read a range of thinkers on the political dimensions of science and technology — Francis Bacon, Karl Marx, Martin Heidegger, Michael Oakeshott, C. S. Lewis, Robert Nisbet, and George Grant, among others. My fellow students and I were skeptical of “progress” across the board.

But taking Econ 101 with a professor trained at George Mason University encouraged the green shoots of mainline economics that had been growing in my brain since high school, the period where I first read Wilhelm Röpke, Milton Friedman, and Paul Heyne. Though I’m not quite sure when, I also started reading Marginal Revolution in college and eventually took an internship at Mercatus, working under Garrett Brown — now a colleague who writes at Humane Pursuits.

Near the end of college, I began to cultivate parallel interests in the history and philosophy of science and economic history — particularly the post-war, cliometric variety. These fields provided a stark contrast with the empirically light discussions that often characterized many of the most prominent writers on science and technology.

Gradually, with these new empirical reference points, the claims of the more pessimistic writers seemed incomplete at best or misleading at worst. How exactly would Neil Postman — or for that matter E. F. Schumacher, Wendell Berry, or Paul Kingsnorth — have responded to the global doubling of life expectancy had they seen it? How would they explain the misery of techno-economic growth alongside major improvements in literacy, child mortality, and maternal mortality? These authors seemed totally unwilling or unable to contend with the eucatastrophe of growth.

The critics of techno-economic growth were stuck. Writing in the shadow of the atom bomb is sobering but also comes with epistemic risks. Indeed, we should be wary of all utopian schemes. But a persistent negativity bias can nudge us to adopt or advocate for policies that restrict useful innovation. It’s one thing to live alone as a Swansonian curmudgeon — it’s another thing to impose unwarranted costs on your neighbors, to halt or slow advances in science, medicine, and national security for what is often merely an aesthetic impulse.

That said, I am still in my bones fairly pessimistic about human nature. War, especially, continues to be the major risk that could slow or halt our material improvement, just as World War I brought an end to the first era of globalization. There is no guarantee that these advancements in human life will last if we do not work to renew them and their moral-ecological roots. We’re all muddling through here.

What work of art has most shaped your views on emerging tech?

It seems impossible to pin down one discrete art piece in the weird network of experiences that shape my beliefs about tech. So I’ll just answer by gesturing toward the films of Peter Weir.

I first watched The Truman Show when I was quite young — maybe nine? That was before social media had become a core part of our daily lives. It’s the only man-disillusioned-with-modernity ‘90s flick I like — maybe the only one that still holds up.

I think the film has more in common with Dante’s Divine Comedy than with The Age of Surveillance Capitalism. The film asks, How can we know what is good, true, and beautiful when we are just so comfortably numb? In Truman’s case, his quest begins and ends with a beatific vision.

Weir’s other films also directly and indirectly take up questions about technology, which are ultimately questions about human ends and means.

Gallipoli is about the bonds of friendship being subjected to mechanized war — “the social history of the machine gun.”

Witness is about norms and belonging between two worlds — one pacifist and weirdly frozen in time and another modern and groaning in violence.

Mosquito Coast is a fairly straightforward, albeit agonizing allegory about hubris and utopia.

Master and Commander: The Far Side of the World is about a small, intense community of discovery and creativity — a “society of explorers,” to use Michael Polanyi’s phrase.

The Way Back is about escaping the planned society — about the price of freedom and man’s ability to adapt in hostile terrain.

Weir’s films often attempt to show that — even when we deny it or seek to expunge it — the world is still deeply enchanted in ways that confound and perplex us. They are films about emergent order.

What’s your most contrarian take on AI?

I am a sand-god atheist — or at least a skeptical agnostic. I tend to locate my sources of re-enchantment in other ways.

I talk about this in my “Thesis and Prospectus” launch essay:

Eschatology — theology concerning the end of days — offers an unstable framework for cultivating good governance. It makes us fragile. Claims about infinite expected value — visions of damnation and glorification — make weighing costs and benefits unworkable. Governance that is practical, decentralized, and bottom-up becomes irrelevant, even dangerous, in this vision.

So, sorry about that, Roko…

What are you reading (or watching, or listening to) now?

Nothing particularly high-brow… though I did just finish Bono’s memoir, and I think he is unironically one of the great Irish poets of our time. He’s also a solid writer and reader of literature — in conversation with the likes of William Butler Yeats, James Joyce, Seamus Heaney, and Flannery O’Connor.

Other than that, I am mostly reading about AI and the political economy of science and tech. The polywater affair is also a recent interest of mine, so I am reading about that in preparation for a long essay on the subject.

Go-to emerging tech music track?

I view all technology as being “dual-use” at a core level. The desires of men writ large are always a mixture of good and evil things. We can’t help it.

“Good Technology” by the 1980s English indie rock band Red Guitars has unmistakable Labourite vibes to it. We should look past that — the lyrics need not be taken as pure irony. They are in fact mostly singing about good things.

But life is full of tradeoffs, and There Is No Such Thing as a Free Lunch.

Subscribe now

Welcome back to the “Model Convo” series — micro interviews with researchers in AI policy and related fields. If you know someone I should interview — maybe that person is you? — email me.

Henry Oliver is a Research Fellow in the Emerging Scholars Program here at the Mercatus Center at George Mason University. He is a writer, critic, and runs (for my hypothetical money) the best literary Substack around — The Common Reader.

Henry writes extensively on the relationship between literature and large language models. Some recent short essays of his include:

• “Is It Unliterary to Oppose AI?”

• “How Long Until AI Writes a Great Poem?”

• “Is Writing for AI the Future of Writing?”

• “Why Read in the Age of AI?”

• “Will People Enjoy AI Art?”

Henry is the author of Second Act: What Late Bloomers Can Tell You About Success and Reinventing Your Life. He is also working on a forthcoming book on the role of classic literature in the cultivation of liberal values and human flourishing. If that were not enough, he co-writes The Pursuit of Liberalism alongside the philosopher (and fellow Mercatian) Rebecca Lowe.

Before coming to the swamplands of DC, Henry worked in many capacities, including as an assistant for a British MP, teaching assistant, copywriter, and brand consultant. He holds a graduate diploma in law and received his BA in English language and literature from Oxford University in 2008.

What follows is a brief interview edited for clarity.

How did you get into emerging tech?

If you are interested in literature and writing, you cannot not be interested in AI.

My writing in this area is mostly about how good AI is at writing poetry (not very good at the moment).

Otherwise, I was largely converted to libertarian ideas about technology when I worked in Parliament, even though I am not very techie myself. For two years in my twenties, I had no phone.

What work of art has most shaped your views on emerging tech?

I think Her is a good movie and has a lot to do with the Pygmalion myth, which I wrote about as it relates to AI.

Arcadia by Tom Stoppard is about the progression and loss of knowledge, and the potential for unity between the arts and sciences. I love that play very much.

What’s your most contrarian take on AI?

Among my readers, who are often literary people, the most contentious point is that I am ambivalent about AI, rather than being either pro or anti.

(I am more pro than anti though and some readers dislike that…)

I recently asked if it is unliterary to oppose AI, for example. I do not find literary arguments about AI and reading very persuasive.

Among AI people, I expect I do not have contrarian takes, other than that you all should be reading Shakespeare and Tolstoy and the other Great Books.

Some of my most engaged readers are STEM types who are coming to literature afresh or for the first time.

Literature is essential to getting a full vision of life: imagination breaks the path that reason follows. That applies to the SV vibe shift too...

I also think AI will have a taste of its own.

What are you reading (or watching, or listening to) now?

Right now? I have Thomas Tallis on, and Purcell. I am reading ‘The Loneliness of Sunny and Sonia’ which I love. I don’t watch as much, but I am watching some Western movies, John Wayne especially.

Go-to emerging tech music track?

The music I enjoy is largely by dead people. But the history of music is the history of technological development, so can I say Mozart’s clarinet quintet?

Subscribe now