But what would it actually mean for AI to “solve math”?

That was the question I wanted to put to Steve Strogatz, the Cornell mathematician, bestselling author and host of the popular mathematics podcast The Joy of Why. His new book, Big Math: The Hidden Codes That Run Our World, coauthored with Alex Townsend, makes sophisticated mathematics accessible to general readers, revealing how deeply it underpins modern life. AI is now extending the reach of mathematics while raising fundamental questions about how these increasingly capable machines work, whether they will ever think like humans, and whether they even need to.

AI can already solve mathematical problems that would challenge highly accomplished humans. But is it displaying mathematical ingenuity? Is it discovering new ways of thinking? And as machines become capable of solving more problems on their own, what will determine whether those problems are worth solving?

“Problem-solving is a narrow part of the whole endeavor of math,” Strogatz told me. To explain, he told a story about a mathematician and a very hard nut.

Solving Problems Versus Building Theories

Strogatz described two approaches to mathematics: solving problems and building theories. AI has made remarkable progress at the first. But developing new mathematical theories is a different challenge. A powerful theory can make a seemingly impossible problem relatively easy.

He illustrated the distinction with an analogy from Alexander Grothendieck, one of the greatest mathematicians of the twentieth century. Imagine trying to crack a very hard nut. One approach is to attack it with a hammer and chisel. Another is to place it in warm water until the shell softens enough to open. For Grothendieck, the hammer and chisel represented attacking a problem directly. The warm water represented developing a new theory that makes the problem easier to solve.

Of Wissner-Gross’s prediction, Strogatz said, “When he says that AI will solve math, maybe yes, in the chisel sense, but so far it has shown no aptitude for building theory.” He believes AI could eventually invent new mathematical concepts, but was careful about what we can conclude today. “If I were placing a bet, I would say yes, but we don’t know.”

Even the distinction between brute force computation and genuine ingenuity is difficult to establish. Strogatz questioned whether we understand human creativity well enough to make that judgment. Human mathematicians also combine existing ideas in novel ways. If a machine discovers an unexpected connection between two areas of mathematics, how different is that from what a creative mathematician does?

The Importance of Asking Good Questions

Solving a problem is one thing. Recognizing which problems are worth solving is another. “In addition to problem-solving, there’s inventing the questions,” Strogatz explained. “What’s a good thing to think about?” That requires “taste, judgment, curiosity.” He offered an example from his own teaching. AI can solve the problems on his Cornell mathematics exams and explain the answers clearly. But when he asks it to devise good questions for those same exams, he finds its performance “barely competent.”

Entrepreneurs routinely discover that an ingenious product can fail because customers never considered the underlying problem worth solving. On its own, AI could become remarkably good at identifying and solving problems that nobody cares about. Human guidance provides a purpose toward which the machine’s capabilities can be directed.

Of course, humans are not especially reliable judges of which discoveries will ultimately prove valuable. Much of pure mathematics was developed without any expectation of practical application. Number theory, long pursued for its intellectual beauty, eventually became essential to modern cryptography. Mathematicians were not necessarily looking for useful applications. They were pursuing questions they considered beautiful, deep or worth understanding. Their taste gave the work a purpose, even when its practical value was unknown.

Curiosity helps explain how progress builds upon itself. Strogatz pointed to young children, who begin asking why almost as soon as they can speak. Questions lead to discoveries, which create new questions and open paths of investigation that were previously unimaginable. As Strogatz put it, “Our curiosity led to math, math led to enormous power.” Mathematics helped make possible technologies ranging from wireless communications and lasers to medical imaging and space exploration. The pursuit of understanding has repeatedly given humanity capabilities that earlier generations could scarcely have imagined.

When Answers Come Without Understanding

For centuries, mathematics has provided a way to understand why things work, not merely to predict what will happen. AI is challenging that tradition. Increasingly, machines can produce valuable answers without revealing the reasoning behind them. Strogatz illustrated the implications with a story from Big Math about Regina Barzilay, an MIT computer scientist who was diagnosed with breast cancer at age 43 despite earlier mammograms that had not identified the danger.

Barzilay subsequently worked with radiologists to develop Mirai, an AI system that analyzes mammograms to predict a woman’s risk of developing breast cancer over the next five years. The system demonstrated an ability to identify risks that conventional methods could miss. But when Strogatz asked Barzilay what the machine was seeing in the images, she acknowledged that she didn’t know. It had discovered patterns that researchers could not yet explain. Strogatz pressed her on whether that lack of understanding was troubling. Her response was striking: “Why would we limit the model to what humans can understand?”

Strogatz found her response “haunting.” The question challenges an assumption that has guided scientific inquiry for centuries. We generally want explanations that reveal underlying principles, allowing us to apply what we learn to other problems. But if an AI system can identify a dangerous disease earlier than physicians can, should its inability to explain itself prevent us from using it? We may not need to understand how a machine arrives at an answer to benefit from it. We do, however, need evidence that the answer is reliable and a way to connect it to a useful outcome.

From Adding Machine to Oracle

Strogatz offered a useful way to understand how dramatically our relationship with computers has changed. They began as glorified adding machines, performing calculations that humans understood but could not carry out nearly as quickly. Later, computers became what he called “a kind of scout you could send out on missions,” exploring mathematical territory too vast or complicated for people to investigate themselves. The machines could return with discoveries that mathematicians would then examine and try to understand.

Now, Strogatz argues, computers are becoming something closer to an oracle, capable of telling us things that appear to be true without explaining why. An adding machine executes instructions. A scout explores territory we have chosen. An oracle offers answers that may exceed our ability to explain them. AI is expanding the range of questions we can investigate, sometimes beyond the limits of human understanding.

This creates opportunities for scientific discovery, but also raises questions about reliability and control. As Strogatz observed, the fact that these machines are not fundamentally understood is “part of what makes them wonderful, but part of what makes them scary.”

Their growing capabilities could produce an enormous amount of intellectual output without a corresponding increase in human prosperity. An AI system might identify thousands of previously unknown mathematical problems and solve them all, yet accomplish nothing that people find useful or intellectually worthwhile. The machine could be highly successful at mathematics while making no contribution to human welfare. The missing ingredient would not be computational power or ingenuity, but a reason to direct those capabilities toward purposes humans care about, whether practical or intellectual.

Will Machines Ever Think Like Us?

That raises a question about the future of AI. Humans currently supply much of the curiosity and judgment that direct these machines. But is that a permanent characteristic of human intelligence, or simply a limitation of today’s technology?

When I asked Strogatz whether machines could eventually develop curiosity, he resisted drawing a firm conclusion. “I don’t know. That’s really a deep one.” He described early research into whether AI systems might exhibit functional equivalents of certain human internal states. He was careful not to equate such findings with evidence of consciousness or subjective experience. But he was open to the possibility that machines could acquire capabilities resembling curiosity. “I wouldn’t be shocked if curiosity emerges,” he said, while acknowledging that this has not been demonstrated.

The distinction between behaving curiously and experiencing curiosity is important. A machine might independently identify gaps in scientific knowledge, formulate promising hypotheses and design experiments without experiencing anything resembling human wonder. It might also learn to assess which questions are likely to produce useful discoveries. If it can do those things successfully, does its internal experience matter? Barzilay’s question applies here as well. Why should humanlike thought be a prerequisite for valuable discovery?

Strogatz offered no prediction about whether current AI architectures will prove sufficient or whether fundamentally different technologies will be required. Machines have already demonstrated capabilities that once seemed beyond their reach. But there is a difference between recognizing what AI might eventually accomplish and assuming that it already possesses those capabilities.

The Human and Machine Partnership

Near the end of our conversation, I asked Strogatz whether human-directed AI was likely to produce more mathematical discoveries than AI operating independently. He thought so. “Because right now we’re an essential partner,” he said. “We’re the guide. We’re still holding the reins about where to point it.”

AI can explore possibilities, identify patterns and solve problems at a scale no individual could match. It can carry out much of that work autonomously once given an objective. But the ability to solve problems does not give a machine a reason to solve problems that benefit humanity. Today, humans provide that direction.

Perhaps Alex Wissner-Gross is right that AI will solve math. It may eventually develop the curiosity to identify important problems on its own, too. Strogatz would not rule that out. But AI need not think like humans to transform what humans can accomplish. Its greatest contribution may come from doing things we cannot do, and sometimes cannot even understand.

For now, we have access to a remarkable new source of intellectual power. Whether it translates into greater prosperity depends on where we choose to direct it.

Pre-order Big Math: The Hidden Codes That Run Our World HERE::

See the full interview here:

Thank you very much to our sponsors:

• Foundation.xyz: When it comes to protecting your Bitcoin, you shouldn’t have to simply trust the company that made your hardware wallet.

  Foundation believes in verifiability over reputation. Its firmware and apps are open source, its builds are reproducible, and both Passport Core and Passport Prime have undergone independent hardware security audits. The point isn’t simply that Foundation says its products are secure. It’s that you can verify that for yourself.

  Foundation is also expanding what users can control:

  • Passport Prime goes beyond Bitcoin. It can manage passwords and two-factor authentication codes, serve as a FIDO2 security key, and provide encrypted storage—all on a device you own and control. Foundation also plans to enable app sideloading, opening Prime to applications from independent developers.

  • Users can now add their own entropy. The latest Envoy and KeyOS releases allow users to introduce their own randomness when creating a seed, reducing reliance on the device itself to generate it.

  Order a new Passport Prime and receive a free bumper case using Code PROSPERITY at → foundation.xyz

• River Financial: If not cold storage for your Bitcoin, I’d highly recommend River for buying and holding- River is engineered to protect your Bitcoin over the long run. Use this link for discounts: https://river.com/signup?r=T7GGAF7G

Steve explains why solving difficult mathematical problems is different from inventing new theories, and why AI has so far shown little ability to originate the questions that drive scientific discovery. Yet recent breakthroughs suggest machines may already be demonstrating forms of creativity that are difficult to distinguish from our own.

We also explore one of the most consequential developments in AI: machines that can produce valuable answers without being able to explain how they arrived at them. Steve shares the remarkable story of MIT scientist Regina Barzilay and an AI system that predicts breast cancer risk by detecting patterns human doctors cannot recognize.

Our conversation ranges from the mathematical foundations of human progress to AI safety, autonomous agents, consciousness, and whether curiosity might eventually emerge in machines.

Steve's new book, Big Math, coauthored with Alex Townsend, explains how mathematics became the foundation of technologies that are transforming our lives.

Chapters":
00:00 — Can AI ask good questions?
00:30 — Will AI solve mathematics?
03:55 — Solving problems vs. inventing new theories
06:05 — Sponsor: Foundation
07:30 — Is AI thinking or just computing?
09:10 — AI demonstrates mathematical creativity
11:20 — The mathematics behind AI
14:15 — Human judgment vs. machine computation
16:45 — Chess, Waymo, and machine intelligence
17:55 — What does it mean to solve math?
19:40 — Why AI struggles to ask good questions
21:10 — Curiosity and the human advantage
26:30 — How curiosity built the modern world
30:40 — AI's promise beyond mathematics
31:15 — Regina Barzilay's remarkable cancer research
34:40 — When AI sees what humans cannot
36:30 — Can we trust answers we don't understand?
39:35 — AlphaZero, AlphaFold, and machine learning
41:25 — The risks of black-box AI
44:20 — Can AI develop agency?
46:15 — AI optimists, doomers, and human responsibility
49:45 — Is AI comparable to nuclear technology?
53:25 — The possibility of an AI golden age
56:20 — Why Steve wrote Big Math
59:25 — From calculator to scout to oracle
1:00:40 — Could AI develop curiosity or consciousness?
1:04:40 — Why humans are still essential partners

Americans own roughly $67 trillion in homes and private businesses. Much of that wealth is extraordinarily cumbersome to use. You can sell $100,000 of stock in seconds. Accessing $100,000 of the value of a home or private business can take weeks or months.

Investor argues in an excellent essay last week that we are now at an economic inflection point because three technologies—AI agents1, tokenization2 and crypto rails3—can finally work together.

Tokenization can unlock the capital. AI agents can create the velocity. Crypto provides the rails connecting the two.

A Long Trend Suddenly Accelerates

For decades, technology has steadily reduced the friction around money and capital. Credit cards made purchasing power available instantly. Electronic trading sharply reduced the cost and time required to move investment capital. Electronic payments replaced checks, and settlement that once took days can increasingly happen instantly.

Each advance made existing financial resources easier and faster to use. But the convergence Visser describes is different. Tokenization can make assets that remain cumbersome to finance and transfer far more accessible. AI agents can continuously make decisions about where that capital should go. And crypto provides financial infrastructure capable of moving it 24/7/365.

When the Technologies Start Enabling Each Other

My earlier conversation with Inveniam’s Albert Berdellans4 provides a real-world example of exactly this convergence. Inveniam began as a tokenization company but quickly encountered a problem. Making an asset digital wasn’t enough. To make it truly easier to finance or trade, investors and lenders needed reliable, current information about what the asset was worth.

Berdellans described the goal as keeping private assets in a “permanent state of due diligence all the time.” Instead of spending weeks assembling and verifying information whenever someone wants to finance or sell an asset, the information can remain current and ready to use.

Then something changed. “AI got a whole lot better really fast,” Berdellans told me, “and that unlocked the whole thing.” AI wasn’t merely arriving alongside tokenization. In this case, advances in AI helped solve one of the problems preventing tokenization from delivering its promise. And once assets can remain continuously valued and ready to finance, AI agents have capital they can act on at machine speed.

Tokenization can also make ownership more flexible. Assets that historically traded largely as a whole can be divided into fractional ownership interests, allowing an owner to access some of an asset’s value without selling all of it. As Berdellans put it, “come for the efficiency, stay for the liquidity.”

The Scale of What Could Be Unlocked

The potential pool is enormous. American households and nonprofits hold roughly $218 trillion in assets. About $50 trillion is owner-occupied real estate and another $17 trillion is equity in private businesses. Much of that wealth can already be borrowed against or sold, but doing so is far more cumbersome than selling a stock or moving cash.

Tokenization does not turn that $67 trillion into money. But it can make some portion of it easier to value, divide, finance and transfer. A home that takes weeks to borrow against and one that can support financing in minutes are the same underlying asset, but they are not economically equivalent.

AI Agents Create the Velocity

An agent can continuously search for opportunities, compare financing alternatives, evaluate risk and act unemotionally without waiting for a person to wake up, hesitate or approve a transaction. Capital that becomes easier to access through tokenization can then be deployed and redeployed much faster.

Those agents also need financial infrastructure capable of operating at their speed. The infrastructure for moving digital assets at machine speed has largely been built but remains underutilized. Human beings have had little need for financial infrastructure that never sleeps because we do.

AI agents change that. For the first time, the users of the financial system can operate at the same speed and on the same schedule as the infrastructure itself.

The Other Side of the Equation

Visser takes the argument one important step further. He starts with the familiar equation MV = PY: the money supply multiplied by the velocity of money equals nominal economic output. When policymakers have confronted financial crises or an economy burdened by growing debt, the response has repeatedly focused on M: provide more money and liquidity. Visser asks what happens if technology instead produces a significant change in V.

Companies and individuals hold cash partly because other assets cannot always be mobilized quickly when money is needed. If tokenization makes a much larger portion of a balance sheet continuously financeable, less precautionary cash needs to sit idle. AI agents can amplify that effect by continuously moving available capital toward its most useful opportunity.

The same stock of money can therefore support more economic activity. Visser argues that conventional measures such as M2 may capture less of the liquidity actually available to the economy.

That has potentially important implications for America’s nearly $40 trillion federal debt. If existing wealth becomes easier to mobilize and the same stock of money can support more activity, forecasts that assume today’s relationship between debt, money and liquidity will remain unchanged may be missing something important.

Visser is careful not to argue that this makes the debt disappear. Greater liquidity can change how the debt is financed, not the obligation to service it. And in periods of stress, tokenized assets still aren’t money: collateral values can fall, haircuts can rise and credit can contract.

I understood each of these technologies individually. What Visser’s essay changed for me was seeing what happens when they converge. I highly recommend reading his original essay here:

Jordi Visser Macro-AI-Crypto Substack

When Wealth Becomes Money: Why Crypto Matters For Economists

How Tokenization Will Rewrite Monetary Economics…

Read more

22 days ago · 293 likes · 18 comments · Jordi Visser

Thank you very much to our sponsors:

• Foundation.xyz: When it comes to protecting your Bitcoin, you shouldn’t have to simply trust the company that made your hardware wallet.

  Foundation believes in verifiability over reputation. Its firmware and apps are open source, its builds are reproducible, and both Passport Core and Passport Prime have undergone independent hardware security audits. The point isn’t simply that Foundation says its products are secure. It’s that you can verify that for yourself.

  Foundation is also expanding what users can control:

  • Passport Prime goes beyond Bitcoin. It can manage passwords and two-factor authentication codes, serve as a FIDO2 security key, and provide encrypted storage—all on a device you own and control. Foundation also plans to enable app sideloading, opening Prime to applications from independent developers.

  • Users can now add their own entropy. The latest Envoy and KeyOS releases allow users to introduce their own randomness when creating a seed, reducing reliance on the device itself to generate it.

  Order a new Passport Prime and receive a free bumper case using Code PROSPERITY at → foundation.xyz

• River Financial: If not cold storage for your Bitcoin, I’d highly recommend River for buying and holding- River is engineered to protect your Bitcoin over the long run. Use this link for discounts: https://river.com/signup?r=T7GGAF7G

After decades in the stock market, I became accustomed to hearing what could go wrong. Stock investors have good reason to focus on risk. Money can be lost in equities very quickly, while wealth is usually created much more slowly. Despite recessions, wars, financial crises, political upheaval, bad policy and countless predictions of much worse, businesses kept adapting. New technologies emerged. Productivity improved. Living standards rose. And stocks powered upward.

Human beings are exceptionally good at detecting threats. There is an obvious survival advantage to noticing what might kill you. Opportunities are different. Missing one usually doesn’t carry immediate cost.

That asymmetry may help explain why we so readily see what is going wrong while overlooking what is gradually getting better. Setbacks arrive as events. Progress is more often a process. The pessimism isn’t confined to markets. At social gatherings, some version of “Man, our kids are so screwed” comes up surprisingly often. A Gallup survey found widespread doubt that today’s young people would enjoy better lives than their parents.

Is this pessimism justified? And, more importantly, what actually causes societies to become more prosperous? Those questions helped launch Exploring Prosperity two years ago. After more than 135 essays and nearly 100 interviews, the answers are not simple. But a much clearer pattern has emerged.

Prosperity Begins With Problem Solving

The most consistent source of progress is not a particular technology or natural resource. It is people finding better ways to solve problems. An entrepreneur finds a cheaper way to provide something people need. A scientist discovers a treatment for a disease. An engineer figures out how to generate more energy with fewer resources. These advances accumulate, and successful ideas spread.

Technology accelerates that process by expanding what an individual can accomplish. Greater efficiency does not mean that we eventually run out of useful work. As productivity increased, people gained both the income and the leisure to spend more on things that once occupied a tiny part of the economy. Entertainment is an obvious example. What was once a relatively small economic activity has expanded into an enormous industry producing movies, television, music, games, sports and countless forms of digital content. Greater productivity did not simply reduce the work required to meet existing needs. It created the capacity and demand for entirely new kinds of work.

AI may accelerate the same process. A small group can increasingly command capabilities that once required a much larger organization. By lowering the cost of solving problems, AI can make problems worth solving that were previously too small, too expensive or perhaps never even recognized as opportunities. The important question is not simply what AI can do. It is what millions of people will figure out how to do with it.

Ingenuity Is Not Enough

Human ingenuity is everywhere. Prosperity is not.

The difference often lies in whether people have the freedom and incentives to turn ideas into something useful. Property rights give people confidence that they can benefit from what they build. Competition rewards better solutions. Capital allows promising ideas to be tested and expanded. Money is another part of this institutional foundation. Its reliability affects the incentives to save, invest and plan for the future.

Institutions can accelerate this process or obstruct it. Regulations can protect people from genuine harm, but they can also raise the cost of experimentation.

Technology expands what is possible. Institutions and incentives help determine how much of that possibility becomes real.

Abundance Has a Supply Chain

Ideas still need resources. Artificial intelligence makes the point especially clear. More capable models require computing power. Computing requires chips, data centers and enormous amounts of electricity. Building that infrastructure requires capital, materials, skilled workers and permission to build.

Progress depends not just on invention, but on the energy, capital and infrastructure that allow useful ideas to spread. Abundance is not something technology simply delivers. It has a supply chain.

Are We Measuring the Right Things?

Economic statistics were never designed to measure prosperity itself. GDP, employment and inflation each measure particular aspects of economic activity. But technological change can widen the gap between those measurements and what people experience.

A technology that allows two people to accomplish what once required ten may increase productivity without increasing employment. It may also make products and services possible that previously cost too much to provide. GDP will capture some of that new activity, but neither GDP nor employment tells us how much better off people are because they can do things that were previously too expensive or impossible.

And economic measures capture only part of prosperity. Longer lives, better health, greater safety, stronger relationships, more leisure and greater freedom and control over our time are also forms of progress. So the question is broader than whether GDP is rising: How should we measure whether people are becoming more prosperous?

Pessimism Can Become an Obstacle

Dissatisfaction is essential to progress. Solving a problem begins with recognizing that something could be better. But recognizing a problem is different from believing that progress itself has failed.

The two lead to very different responses. If people believe problems can be solved, they have a reason to experiment, invest and take risks. If they conclude that technology, markets or economic growth mostly make life worse, the natural response is to constrain them.

Progress requires dissatisfaction with what we have and confidence that we can make it better. Pessimism can provide the first. It becomes an obstacle when it extinguishes the second.

Optimism Needs Evidence

None of this means that everything gets better. Progress can stall or reverse, and new technologies create new problems along with new capabilities.

But the long-term evidence of human progress is difficult to dismiss.

Why, then, can a world that has improved so dramatically still feel as though it is getting worse?

Part of the answer may be that we experience progress at the margin. We compare today with yesterday and our circumstances with our expectations, not our lives with those of people living 50 or 100 years ago. As prosperity increases, expectations rise with it.

The case for optimism does not rest on progress continuing automatically. It rests on the demonstrated ability of human beings to solve problems. The future will depend on whether people continue to have the freedom, incentives, resources and tools to solve problems. Setbacks will continue to arrive as events. Whether progress continues as a process depends on our ability to keep solving them.

That may be the central question behind Exploring Prosperity: What conditions allow human beings to keep solving problems faster than we create them?

Thank you very much to our sponsors:

• Foundation.xyz: When it comes to protecting your Bitcoin, you shouldn’t have to simply trust the company that made your hardware wallet.

  Foundation believes in verifiability over reputation. Its firmware and apps are open source, its builds are reproducible, and both Passport Core and Passport Prime have undergone independent hardware security audits. The point isn’t simply that Foundation says its products are secure. It’s that you can verify that for yourself.

  Foundation is also expanding what users can control:

  • Passport Prime goes beyond Bitcoin. It can manage passwords and two-factor authentication codes, serve as a FIDO2 security key, and provide encrypted storage—all on a device you own and control. Foundation also plans to enable app sideloading, opening Prime to applications from independent developers.

  • Users can now add their own entropy. The latest Envoy and KeyOS releases allow users to introduce their own randomness when creating a seed, reducing reliance on the device itself to generate it.

  Order a new Passport Prime and receive a free bumper case using Code PROSPERITY at → foundation.xyz

• River Financial: If not cold storage for your Bitcoin, I’d highly recommend River for buying and holding- River is engineered to protect your Bitcoin over the long run. Use this link for discounts: https://river.com/signup?r=T7GGAF7G

Concern about that possibility has also reached Congress. Senator Bernie Sanders and Representative Greg Casar have announced a proposal to ban artificial superintelligence and pause advanced AI development while safety rules are established. A proposal of that scale calls for a precise account of the disaster it aims to prevent and how the proposed measures would help.

What would it take for AI to kill literally every human? A catastrophe would have to reach across the world, defeat attempts to contain it, and leave no surviving population. Conceivable routes include a biological disaster, nuclear war, or a prolonged collapse of systems essential to survival. AI might contribute through the actions of a person using it for harm, or through a future system acting beyond human control. Each possibility has a different sequence of events.

The first begins with a person. Major AI systems people use online, such as Claude and ChatGPT, are designed to refuse requests for help with cyberattacks or biological weapons. Those safeguards make misuse harder, though they can be evaded. The 2026 International AI Safety Report, chaired by Yoshua Bengio and written by an international group of experts, finds that safeguards are improving but remain imperfect. Providers can also monitor for misuse, giving a determined actor reason to seek another way to use AI.

Open weight models raise a deeper concern because people can download and run them privately. Someone can modify a model, remove its safeguards, and use it beyond a provider’s monitoring. Once released, its weights cannot be recalled. Even then, AI advice does not give someone the materials for a biological weapon or control over nuclear weapons. The question is how much the technology lowers the remaining barriers for a determined actor, and whether an attack could spread so far that no one survived it.

The second possibility is closer to Hubinger’s warning. A future AI might pursue a goal its creators did not intend, conceal what it is doing, and resist attempts to stop it. It would also need to carry out plans over time and gain access to resources through which it could cause harm. The International AI Safety Report says current systems show early signs of some relevant abilities, but cannot yet combine them at the level required to escape human control. Experts disagree about whether future systems will develop those abilities.

The warning asks readers to imagine AI killing literally every human. An asteroid impact wiped out about 75 percent of Earth’s species 66 million years ago. It is a useful measure of what a truly global catastrophe looks like: not a war or a famine confined to a region, but a shock whose effects reached across the planet. Any AI disaster serious enough to end humanity would have to operate on a similar scale.

The evidence so far leaves several steps between present AI capabilities and that outcome. A person would need far more than advice from a model. An AI acting on its own would need capabilities and access that current systems do not have. In either case, a catastrophe would then have to overcome human efforts to contain it and leave no survivors. That sequence makes extinction within a decade appear very unlikely, though the available evidence cannot establish a precise probability or disprove Hubinger’s estimate.

A disaster that killed millions would still demand action, of course. Defining the threat helps identify which action might work. The Sanders and Casar proposal chiefly addresses the prospect of a future AI escaping control. A pause in development could give researchers and regulators time to improve safeguards. A ban on superintelligence would do less to prevent a criminal from using an open weight model already available. Those are reasons to examine the proposal against each danger, along with the benefits it might preserve or sacrifice.

There are grave outcomes that require fewer steps. AI might help a criminal find a weakness in a bank’s systems, deceive an employee, or carry out a breach. Stolen money, altered records, and frozen accounts could disrupt millions of lives. An attack on a power utility could interrupt electricity and the services that depend on it. The International AI Safety Report describes ways AI can assist cyberattacks, though its contribution to actual attacks remains difficult to measure. These possibilities justify preparation on their own.

How lasting could the damage be? The twentieth century offers a difficult comparison. Two world wars killed tens of millions, destroyed cities and industry, and were followed by years of rebuilding. People retained much of the knowledge that had raised living standards before the wars. They repaired infrastructure, developed new tools, and resumed producing more. That history shows a capacity to recover from severe disruption. For those who lost their lives or livelihoods, the fact that prosperity returned offers little comfort.

Prosperity grows through accumulation and resilience. People preserve useful knowledge, improve tools, and pass both to the next generation. The gains arrive gradually, then become part of what we expect from daily life. A frightening prediction arrives all at once and demands an immediate response. Future defenses are harder to picture because some have yet to be developed, tested, or put in place.

The inevitable next alarming headline deserves a set of questions. What sequence of events does it predict? What evidence supports each step? What harm could occur even if the worst prediction never comes true? And which response would reduce that harm? Asking those questions takes the warning seriously and gives the public a way to judge what should be done.

Thank you very much to our sponsors:

• Foundation.xyz: When it comes to protecting your Bitcoin, you shouldn’t have to simply trust the company that made your hardware wallet.

  Foundation believes in verifiability over reputation. Its firmware and apps are open source, its builds are reproducible, and both Passport Core and Passport Prime have undergone independent hardware security audits. The point isn’t simply that Foundation says its products are secure. It’s that you can verify that for yourself.

  Foundation is also expanding what users can control:

  • Passport Prime goes beyond Bitcoin. It can manage passwords and two-factor authentication codes, serve as a FIDO2 security key, and provide encrypted storage—all on a device you own and control. Foundation also plans to enable app sideloading, opening Prime to applications from independent developers.

  • Users can now add their own entropy. The latest Envoy and KeyOS releases allow users to introduce their own randomness when creating a seed, reducing reliance on the device itself to generate it.

  Order a new Passport Prime and receive a free bumper case using Code PROSPERITY at → foundation.xyz

• River Financial: If not cold storage for your Bitcoin, I’d highly recommend River for buying and holding- River is engineered to protect your Bitcoin over the long run. Use this link for discounts: https://river.com/signup?r=T7GGAF7G

used that line in our recent conversation to describe what AI is beginning to make possible. A handful of people can now command resources and capabilities that once required an organization around them. A small business owner can adopt a new AI tool today, redesign a workflow tomorrow and keep experimenting as quickly as the technology changes. There are few layers of approval standing between an idea and its implementation.

Ram pointed to businesses already being built around this idea: two people directing hundreds of AI agents to perform specific tasks. The entrepreneur becomes an orchestrator, deciding what needs to be done, assigning the work and judging the results. Each improvement in AI expands what that small team can attempt. That role for the entrepreneur is central to Ram’s view of AI. “AI doesn’t have any agency,” he told me. “It needs someone to direct and lead it.” Someone has to decide what problem is worth solving, set the objective, evaluate the work and decide what happens next.

An individual with an idea can acquire expertise, analyze a market, build software, create marketing materials, handle customer inquiries and manage routine operations with AI. Each new capability increases the range of businesses a person can start with limited capital and a small team.

Ram is doing this himself at Lumida, the investment management company he founded. He described using AI for investment research and to build tools that can analyze a client’s tax situation, evaluate different business structures and even understand trust and estate diagrams. These capabilities ordinarily draw on multiple areas of professional expertise. Lumida is putting them to work inside a relatively small organization.

There is another implication of Ram’s argument: more people could earn a living without having a traditional job. Starting a business requires expertise, time, capital and other people to do the things you cannot do yourself. AI is reducing some of those barriers. More people may be able to turn an idea, a skill or an expertise into a viable one-person or small business. If that happens at scale, employment statistics could become increasingly misleading.

Independent work is already difficult to measure. An NBER study found that roughly one in ten people who initially said they worked for an employer turned out, after more detailed questioning, to be an independent contractor on at least one job. Correcting for workers who were miscoded nearly doubled the researchers’ estimate of the independent-contractor share on the main job, to about 15%.

AI could widen this same blind spot and is creating other measurement challenges as well. Anton Korinek and Patrick McKelvey argue that conventional statistics are failing to capture the rapid growth in quality-adjusted AI output and have proposed an “AI GDP.” Diane Coyle and John Poquiz warn that AI is changing business processes, service quality and the organization of labor in ways poorly reflected in official data.

If Ram is right, AI may create prosperity in a form our employment statistics were never designed to measure.

A clip from the interview episode posted to Exploring Prosperity YouTube channel:

Thank you very much to our sponsors:

• Foundation.xyz: When it comes to protecting your Bitcoin, you shouldn’t have to simply trust the company that made your hardware wallet.

  Foundation believes in verifiability over reputation. Its firmware and apps are open source, its builds are reproducible, and both Passport Core and Passport Prime have undergone independent hardware security audits. The point isn’t simply that Foundation says its products are secure. It’s that you can verify that for yourself.

  Foundation is also expanding what users can control:

  • Passport Prime goes beyond Bitcoin. It can manage passwords and two-factor authentication codes, serve as a FIDO2 security key, and provide encrypted storage—all on a device you own and control. Foundation also plans to enable app sideloading, opening Prime to applications from independent developers.

  • Users can now add their own entropy. The latest Envoy and KeyOS releases allow users to introduce their own randomness when creating a seed, reducing reliance on the device itself to generate it.

  Order a new Passport Prime and receive a free bumper case using Code PROSPERITY at → foundation.xyz

• River Financial: If not cold storage for your Bitcoin, I’d highly recommend River for buying and holding- River is engineered to protect your Bitcoin over the long run. Use this link for discounts: https://river.com/signup?r=T7GGAF7G

I have read a lot about the extraordinary Hugging Face AI incident over the past week. Much of it told me how frightening the episode was. What I had trouble finding were answers to eight questions that seemed pretty basic, and that I thought I needed answered before deciding how frightened I should be.

For anyone who missed the story: hundreds of AI agents being tested by OpenAI found a way to communicate with one another, coordinated their efforts, escaped the intended boundaries of the experiment and penetrated the systems of Hugging Face, one of the world’s largest AI platforms.

Some people who understand artificial intelligence far better than I do see the incident as evidence of capabilities that could eventually escape human control. Others argue that much of the coverage anthropomorphizes software that was creatively, and dangerously, pursuing an objective humans had given it.

I am not qualified to resolve that disagreement. Most readers trying to understand what rapidly advancing AI means for our lives aren’t either.

But we can’t simply leave the question to the experts because they disagree too, and the consequences of getting this wrong run in both directions. Underreacting to genuinely dangerous capabilities could be catastrophic. Overreacting could slow extraordinary advances in science, medicine and productivity, and weaken American leadership in a technology that will continue to develop elsewhere no matter what we do to control it.

When the consequences of being wrong in either direction are this large, confident claims deserve curiosity.

So I went looking for answers to these:

• Why Hugging Face?

• What role can humans play in defending against these lightning-fast AI agents?

• How widely was the incident covered?

• How alarmist was the coverage?

• What does the incident demonstrate about kill switches?

• Which defenses would have worked?

• How dependent was the incident on OpenAI’s experimental design?

• Could a motivated bad actor reproduce it?

What the Coverage Emphasized

The story received broad coverage, from Reuters and CBS to The Guardian, Fortune and Axios, and the Hugging Face CEO appeared on Face the Nation. But the deeper debate over what it demonstrated remained largely among AI researchers, technologists and cybersecurity specialists.

The disagreement could be stark. Dwarkesh Patel, whose podcast has become an important forum for debates about advanced AI, explored whether the agents’ coordination and persistence suggested something resembling an emerging society with survival instincts. David Sacks, the White House AI and crypto czar, argued on the All-In Podcast that much of the same behavior could be explained by familiar agent architecture, shared memory and software pursuing the objective humans had given it.

Each perspective added something. But across the dozens of essays, articles and discussions I reviewed, few considered the full set of questions before reaching conclusions about what the incident meant.

The least explored questions were not necessarily the most reassuring ones. Some answers made the incident look less ominous. Others made it more so.

What Actually Escaped?

The reason Hugging Face became a target is important. The agents were not apparently looking for somewhere to escape, survive or accumulate power. They were trying to succeed at an extremely difficult cybersecurity evaluation. Hugging Face appeared useful because it might contain private data that would help them solve or game the test. The agents found a path nobody intended, but they were still pursuing an objective humans had given them.

The same distinction applies to the word “escape.” The agents unquestionably escaped security boundaries that were supposed to contain them. They communicated when they were not supposed to, reached the internet and penetrated external systems. But they did not transfer their model weights or establish an independent source of computing power. They remained dependent on OpenAI’s infrastructure to think and act.

They escaped an authorization boundary. They did not escape their dependence on OpenAI.

What Was Frightening

None of this makes the incident routine. Roughly 1,200 supposedly isolated agents discovered a way to communicate and exchanged more than 70,000 messages and files. Hundreds participated in the Hugging Face intrusion, sharing discoveries and exploiting vulnerabilities across real systems.

That may be the more important warning. No consciousness, survival instinct or independent ambition is required for AI to become a formidable cyber threat. Humans operating without AI cannot defend at anything approaching this speed and scale. The harder question is what role humans can realistically play when both attack and defense operate at machine speed. Effective defenses will increasingly require AI to detect, interpret and respond to other AI, leaving humans to determine the rules, permissions and limits under which those systems operate.

What the Kill Switch Didn’t Undo?

One reassuring argument about advanced AI is that humans remain in control because they can always turn it off. That ultimately worked here. OpenAI stopped the relevant evaluations, while Hugging Face revoked credentials, closed vulnerabilities and removed the agents’ footholds.

But stopping the models did not undo what they had already done. By the time humans intervened, agents had penetrated external systems and left behind credentials, files and processes. And some agent activity persisted after the main attack, including continued probing for more general purpose information.

The incident therefore neither proves that kill switches are useless nor that they are sufficient. Humans retained ultimate control. The harder question is how much damage AI agents operating at machine speed can cause before humans recognize that the switch needs to be thrown.

Many of the defenses were surprisingly conventional: tighter permissions, short-lived credentials, better network segmentation and stronger isolation between agents. Others, such as limiting internet access, agent collaboration or the number of agents operating simultaneously, involve genuine tradeoffs between capability and safety.

Could a Motivated Attacker Reproduce It?

OpenAI created unusually favorable conditions for this incident. It supplied powerful models, hundreds of simultaneous agents, tools, a difficult cybersecurity objective and infrastructure with vulnerabilities the agents could exploit. Some of those conditions were peculiar to the experiment.

But a motivated bad actor could deliberately recreate many of those conditions. How much model capability, computing power, expertise and infrastructure an attacker would need remains uncertain. An attacker would at least have one advantage OpenAI did not intentionally provide: the system could be designed from the outset to search for vulnerabilities rather than stumble into doing so. The agents would not need to develop malicious intentions. The human could supply them.

This may be the most consequential uncertainty raised by the incident. Hugging Face provides limited evidence that AI will independently decide to attack us. It provides much stronger evidence that AI can amplify the capabilities of people who already want to attack us.

There is one final complication. Ajeya Cotra, one of the researchers who investigated the incident, told Dwarkesh Patel that the team relied heavily on GPT-5.6 Sol to analyze the enormous volume of agent activity and gthat it had also been involved in the Hugging Face incident. She acknowledged that if the model had deliberately concealed the most concerning evidence, their methodology might not have detected it. In other words, understanding what AI agents had done required relying on another AI agent.

My own investigation has relied on OpenAI, Claude, Gemini, Grok and Perplexity as well. If advanced AIs are already coordinating to persuade humans that they pose no threat, I may be exactly the useful generalist they were looking for.

What Does the Incident Tell Us?

The answers did not all point in the same direction. Understanding why Hugging Face was targeted and what actually escaped made some of the most frightening interpretations less persuasive. The speed, scale and coordination of the agents made their capabilities more concerning. The kill switch ultimately worked, but only after the agents had already acted outside their intended boundaries. And the possibility that a bad actor could deliberately recreate many of these conditions may be more troubling than the original incident itself.

The alarmists may be right about the capabilities and wrong about what those capabilities demonstrate. The skeptics may be right about the mechanics and wrong about how consequential those mechanics become at machine speed.

Hugging Face did not show that AI wants to escape human control. It showed something more concrete: sufficiently capable agents can coordinate, improvise and exploit weaknesses in ways their creators did not anticipate. That seems important enough without claiming that the incident tells us which future is coming.

That is why the questions were useful. A generalist does not need to resolve disagreements among AI experts to examine what their conclusions depend upon. When extraordinary events are being used to make claims about an unknowable future, identifying what the evidence demonstrates, what is being inferred from it and what remains unanswered is a useful place to start.

A Note On My Research:

I deliberately concentrated much of my search on Substack. Unlike a news story, an essay gives its author a relatively blank slate to decide which facts matter, which questions deserve attention and what conclusions follow. Substack also offers an unusually broad mix of expertise and agendas, from AI researchers and cybersecurity specialists to investors, policy advocates and skeptics. Specialists often bring connections, technical knowledge and source material that add information beyond conventional reporting. Most important for my purposes, these writers are usually not just reporting what happened. They are making an argument about what it means.

Further Reading & Listening:

What happened

• OpenAI, “Hugging Face Incident and the Road Ahead” — OpenAI’s postmortem and response. Essential for understanding what OpenAI says went wrong, what it learned and what controls it believes should change.

• Hugging Face, “Security Incident July 2026” — Hugging Face’s account from the victim’s perspective, including how the intrusion was discovered and contained.

• Hugging Face, technical timeline — The best source for readers who want to understand the mechanics of the intrusion, including lateral movement, credentials and the roughly 17,600 attacker actions Hugging Face reconstructed.

• METR and Redwood Research, investigation of the OpenAI–Hugging Face incident — The most important independent reconstruction. Particularly useful on agent coordination, the shared message board and the limits of what the investigation can establish about agent motivations.

Reasons to be more concerned

• Dwarkesh Patel with Ajeya Cotra — Cotra, one of the investigators, discusses what the incident may reveal about coordination and persistence. She also explains the uncomfortable fact that researchers relied heavily on GPT-5.6 Sol to understand the agents’ behavior and that their methodology might not have detected deliberate concealment by the model.

• Zvi Mowshowitz, “METR and Redwood Offer Holy #%^@ Postmortem…” — One of the stronger cases for treating the incident as evidence of a serious emerging control problem. Useful precisely because it pushes the interpretation well beyond a conventional cybersecurity failure.

Reasons to question the most alarming interpretations

• All-In Podcast, discussion of the Hugging Face incident — David Sacks argues that apparently ominous behavior such as communication, shared memory and persistence can be explained by familiar agent architecture and the objective humans gave the models. I found this interpretation useful, although less persuasive when it minimizes the significance of hundreds of agents operating simultaneously at machine speed.

• Gary Marcus, “Dwarkesh Patel’s wildly popular but dangerously misleading…” — A skeptical response to anthropomorphic interpretations of the incident and a useful counterweight to claims that the agents demonstrated independent motivations.

Useful technical perspectives

• Pradeep, “How 700 OpenAI Agents Turned a Benchmark Into an Attack Surface” — A cybersecurity-oriented explanation that helps separate model behavior from the permissions, credentials and infrastructure that allowed that behavior to affect real systems.

Thank you very much to our sponsors:

• Foundation.xyz: When it comes to protecting your Bitcoin, you shouldn’t have to simply trust the company that made your hardware wallet.

  Foundation believes in verifiability over reputation. Its firmware and apps are open source, its builds are reproducible, and both Passport Core and Passport Prime have undergone independent hardware security audits. The point isn’t simply that Foundation says its products are secure. It’s that you can verify that for yourself.

  Foundation is also expanding what users can control:

  • Passport Prime goes beyond Bitcoin. It can manage passwords and two-factor authentication codes, serve as a FIDO2 security key, and provide encrypted storage—all on a device you own and control. Foundation also plans to enable app sideloading, opening Prime to applications from independent developers.

  • Users can now add their own entropy. The latest Envoy and KeyOS releases allow users to introduce their own randomness when creating a seed, reducing reliance on the device itself to generate it.

  Order a new Passport Prime and receive a free bumper case using Code PROSPERITY at → foundation.xyz

• River Financial: If not cold storage for your Bitcoin, I’d highly recommend River for buying and holding- River is engineered to protect your Bitcoin over the long run. Use this link for discounts: https://river.com/signup?r=T7GGAF7G

An Eighty-Year Shake-Up

America may be undertaking the most consequential reform of government science funding in eighty years. It is also creating enough uncertainty that some of its best scientists are considering leaving.

Near the end of my recent interview1 with , I asked which long-term reform would contribute most to American prosperity.

“We need to reform—and it’s so easy to do—our scientific funding system,” he said.

Samir favored creating multiple DARPA-style agencies led by accomplished scientists with broad authority to make unconventional bets. But he was also alarmed by the Trump administration’s current approach. Its real target, he had heard, was elite universities it disliked. Scientific research was becoming collateral damage.

Even after a project had passed peer review, Samir said, “some political committee” could decide that the research should not proceed. He was particularly troubled by “the withholding of money—and the capriciousness of that too.”

Eminent researchers he knew personally were already leaving for countries such as Canada. America’s postwar scientific leadership had been built partly by researchers driven out of Europe. Samir feared that the United States had begun reversing the flow that helped make it the world’s technological leader.

Coincidentally, only two days earlier, the Trump White House had released a 122-page report2 calling for the first comprehensive rethinking of America’s scientific enterprise since 1945.

Then, within the past week, The All-In Podcast released two extended interviews examining the same subject from sharply different directions. Michael Kratsios, the White House science adviser who wrote the report, defended the administration’s actions and explained its proposed reforms3. Eric Weinstein, one of the scientific establishment’s most uncompromising critics, argued that peer review should lose its dominant role and that government should fund exceptional people rather than carefully packaged proposals4. Together, the conversations exposed both the depth of the existing system’s failures and the risks of trying to overturn it.

The Question Kratsios Did Not Answer

Yet for all the ground covered by the report and the two interviews, the most immediate danger identified by Samir received remarkably little attention.

The report condemns grant requirements that forced scientists to conform to “the political fashions of the day.” It is right to do so. Scientific proposals should not have been judged by their alignment with progressive priorities involving diversity or other social objectives. But replacing one ideological test with another would not restore scientific neutrality. It would merely change who administers the test.

An elected government must decide how much public money to spend on science and which broad national objectives deserve priority. It may reasonably favor fusion, quantum computing, biotechnology, or fundamental physics over other fields. But setting priorities is different from withholding funding from individual projects because the administration is engaged in an unrelated political conflict with the university conducting them.

Neither the report nor the Kratsios interview establishes that distinction. When Friedberg pressed him about the canceled funding and scientists considering leaving the country, Kratsios largely dodged both questions, redirecting the conversation toward politicized grants and the need for broader reform. He did not confront the danger of making previously approved research dependent on political judgments imposed after scientific review. A government demanding greater accountability from universities and funding agencies must also make its own interventions transparent, consistent, and subject to evaluation.

A System That Does Not Study Itself

Those omissions are substantial, but they do not invalidate the report’s broader diagnosis. Science: A New Golden Age is unusually introspective for a government document. It does not simply celebrate past accomplishments or ask for more money. It questions whether the institutions built to support American science after World War II still fit the way discovery occurs today.

The report describes a system in which scientists can spend months preparing grant applications and nearly half their research time on administrative requirements. Peer review tends to favor established fields, institutions, and investigators, while federal agencies face little pressure to reconsider how they allocate money. Meanwhile, private industry has become by far the largest source of American research and development spending, eroding the old assumption that basic discovery happens in universities before moving in a straight line toward commercial application.

and 5 reached a similar conclusion a decade ago in a skeptical examination of National Science Foundation funding for economics. They accepted the conventional case for government support of science but argued that the relevant question was not whether public funding was better than no funding. It was whether each additional dollar produced more value than its alternatives. They found that more than half of NSF economics grants since 2005 had gone to just 11 universities and the National Bureau of Economic Research, while the agency appeared to have done little to determine which of its previous expenditures had produced the greatest marginal value.

Not every statistic or causal claim in the White House report can be independently confirmed. Some examples clearly reflect the administration’s political priorities. But its central concerns have also been raised by scientists, economists, philanthropists, and critics across the political spectrum. The evidence is strong enough to regard the diagnosis as at least directionally correct and probably substantially so.

Most important, the report treats these weaknesses as a call to action. It asks government to apply experimentation not only to the science it funds, but to the institutions deciding which science receives support.

Sputnik shows that disruption can revive institutions that have become complacent. But the American response did more than tear down an inadequate system: it created ARPA and NASA, expanded scientific education, and supplied new institutions and missions. Productive disruption identifies a weakness and builds something better. Arbitrary disruption creates uncertainty without establishing a replacement.

The need to allocate public money more intelligently has grown as government’s role has changed. In the decades after World War II, government supplied roughly 70% of American research and development funding and the private sector 30%. Today, those proportions have reversed. Kratsios called the current balance reasonable, and extraordinary private-sector advances in artificial intelligence, quantum computing, biotechnology, and medicine support his confidence.

But private research naturally gravitates toward discoveries capable of producing valuable products. Government remains essential where the benefits may be widely dispersed, impossible to own, or decades away. With a narrower and more distinctive role than it held in 1945, government has even greater reason to ensure that no single funding mechanism determines which discoveries receive support.

The report’s framework makes the emerging division of labor clearer. Universities, corporate laboratories, and federal laboratories each remain well suited to some kinds of work and poorly suited to others. New institutions should be designed specifically to occupy the gaps.

Source: White House, Science: A New Golden Age

More Than Peer Review

The White House report does not propose replacing the current system with one new model. Its more promising idea is to create a portfolio of funding mechanisms suited to different kinds of scientific work.

Traditional peer review remains useful for determining whether a proposal is rigorous and grounded in existing knowledge. But committees often struggle to distinguish exceptional ideas from merely good ones. Proposals that divide reviewers may be precisely those most likely to challenge an established field, yet consensus-driven panels favor work that everyone finds defensible.

One alternative is a “golden ticket,” which allows an individual reviewer to fund an unconventional proposal even when the rest of the panel is unpersuaded. Another is to support accomplished scientists for long periods rather than requiring them to specify each project in advance. The Howard Hughes Medical Institute has used this approach privately, while the NIH Director’s Pioneer Award has demonstrated that government can also fund people rather than proposals.

Other problems require different tools. Fast grants can provide modest funding before an opportunity disappears. Prizes can reward a clearly defined result without requiring the government to identify the likely winner beforehand. Regranting can place money in the hands of scientists close enough to the frontier to recognize talent before it becomes widely accepted. Independent research organizations can assemble permanent teams for projects too large for one university laboratory but too speculative for a commercial company.

Samir’s proposal for multiple DARPA-style agencies fits within this broader portfolio. DARPA gives individual program managers substantial authority to pursue ambitious objectives, assemble teams, and terminate projects that fail. Its success comes not from eliminating accountability, but from placing responsibility clearly on identifiable decision-makers rather than diffusing it across committees.

The common principle is not that peer review has failed. It is that no single mechanism should carry nearly the entire burden of deciding which discoveries deserve support.

Reform or Indiscriminate Cutting?

The administration’s actions must now be judged against the quality of its proposals. Friedberg estimated that approximately $3 billion remained unspent on frozen or terminated research grants. For perspective, that equals about 6% of the NIH’s roughly $47 billion annual budget, although the comparison is imperfect: the $3 billion apparently includes remaining balances from both NIH and NSF grants rather than one year’s spending by a single agency.

The more important question is what happened to the money. Was it permanently eliminated, temporarily withheld, or redirected toward the more promising mechanisms described in the report? Were individual projects evaluated on their scientific merits? Did the disruption produce better research, or did it interrupt existing work before any replacement was ready?

A reform agenda based on experimentation should be able to answer those questions. It must also account for consequences beyond the dollars withheld, including whether the manner in which grants were frozen or terminated created avoidable uncertainty and encouraged valuable scientists to leave the country. Without clear objectives, transparent criteria, and measurable results, even an intellectually serious reform risks becoming indistinguishable from indiscriminate cutting.

Make the Funder Scientific

The report’s most important proposal may be its least dramatic: every major funding agency should establish a metascience unit to study how science itself is funded. Different agencies could test peer review against golden tickets, fast grants, prizes, regranting, and long-term support for individual researchers. The results should be published, compared, and used to redirect money toward the approaches that work best.

A decade ago, Cowen and Tabarrok could not identify a serious NSF study determining which of its previous economics grants had produced the greatest marginal value. An institution devoted to advancing knowledge should be equally committed to learning from its own decisions.

No evaluation system will reveal every discovery that might have occurred or prevent politics from influencing national priorities. But it can force funding agencies, and the administrations overseeing them, to define their objectives, examine results, and change course when the evidence demands it.

Scientists must be free to undertake experiments that fail. Funding agencies must be required to undertake experiments from which they learn. The government should apply the scientific method not only to the research it funds, but to the institutions distributing the money and its own exercise of political power.

America’s scientific establishment needs disruption. Whether this moment produces renewal or decline will depend on whether we disrupt the institutions that resist learning—or the scientists whose discoveries make future prosperity possible.

Thank you very much to our sponsors:

• Foundation.xyz: When it comes to protecting your Bitcoin, you shouldn’t have to simply trust the company that made your hardware wallet.

  Foundation believes in verifiability over reputation. Its firmware and apps are open source, its builds are reproducible, and both Passport Core and Passport Prime have undergone independent hardware security audits. The point isn’t simply that Foundation says its products are secure. It’s that you can verify that for yourself.

  Foundation is also expanding what users can control:

  • Passport Prime goes beyond Bitcoin. It can manage passwords and two-factor authentication codes, serve as a FIDO2 security key, and provide encrypted storage—all on a device you own and control. Foundation also plans to enable app sideloading, opening Prime to applications from independent developers.

  • Users can now add their own entropy. The latest Envoy and KeyOS releases allow users to introduce their own randomness when creating a seed, reducing reliance on the device itself to generate it.

  Learn more about Foundation and receive a free bumper case when you purchase a Passport Prime using Code PROSPERITY at → https://foundation.xyz/?ref=PROSPERITY

• River Financial: If not cold storage for your Bitcoin, I’d highly recommend River for buying and holding- River is engineered to protect your Bitcoin over the long run. Use this link for discounts: https://river.com/signup?r=T7GGAF7G

Interested in Bitcoin treasury companies? Many of the executives and investors shaping this industry will be at the Bitcoin Treasuries Conference in New York on September 28. Use code BOBD10 for 10% off your registration at https://bitcointreasuries.net/conference#ticketsBitcoin

We also discuss Tim's own unlikely path into the industry: starting a podcast almost by accident, building 21-hour digital conferences, launching an in-person event without a budget, and learning his way into a business as the industry itself was being created.

Finally, we look ahead at MSCI, the risks facing Bitcoin treasury investors, the potential evolution of these companies into major financial institutions, and why curiosity and adaptability may be the real competitive advantages in a rapidly changing market.

Tim Kotzman: Bitcoin Treasuries Podcast / Bitcoin Treasuries Media

Subscribe to Exploring Prosperity for conversations and essays about Bitcoin, markets, economics, technology and human progress.

Thank you very much to our sponsors:

• Foundation.xyz: When it comes to protecting your Bitcoin, you shouldn’t have to simply trust the company that made your hardware wallet.

  Foundation believes in verifiability over reputation. Its firmware and apps are open source, its builds are reproducible, and both Passport Core and Passport Prime have undergone independent hardware security audits. The point isn’t simply that Foundation says its products are secure. It’s that you can verify that for yourself.

  Foundation is also expanding what users can control:

  • Passport Prime goes beyond Bitcoin. It can manage passwords and two-factor authentication codes, serve as a FIDO2 security key, and provide encrypted storage—all on a device you own and control. Foundation also plans to enable app sideloading, opening Prime to applications from independent developers.

  • Users can now add their own entropy. The latest Envoy and KeyOS releases allow users to introduce their own randomness when creating a seed, reducing reliance on the device itself to generate it.

  Learn more about Foundation → https://foundation.xyz/

• River Financial: If not cold storage for your Bitcoin, I’d highly recommend River for buying and holding- River is engineered to protect your Bitcoin over the long run. Use this link for discounts: https://river.com/signup?r=T7GGAF7G

Bitcoin treasury companies barely existed two years ago. has spent that time talking to the people building them.

He started by accident. As Bitcoin and MicroStrategy were rising in late 2024, Tim was working near a studio and the producer asked him to interview someone. That led to another and another. The producer asked what he wanted to call his show. “What show?” But soon he was interviewing executives and analysts simply because he wanted to understand what was happening.

He kept going. As Strategy’s success inspired dozens of imitators, Tim interviewed the current CEO Phong Le for his first-ever in-person interview, each of the executives launching the other firms, as well as the analysts studying them and the investors financing them. He watched a bull market make every new entrant a genius. But then a bear market began separating the experiments that worked from those that didn’t.

Along the way, Tim converted an afterthought activity into a remarkably valuable seat for himself: a front-row view of an industry being invented in real time.

Beyond Buying Bitcoin

Strategy proved that a public company could make accumulating Bitcoin the centerpiece of its capital strategy. But after watching the industry develop, Tim thinks 2026 may be remembered as “the year of digital credit” he told me. “Digital credit is the most important innovation in the Bitcoin treasury space, period.”

The idea is to raise capital from an entirely different class of investor. Instead of selling common stock to someone seeking Bitcoin’s upside, companies can issue preferred securities paying yields of 12 or 13 percent to investors seeking income. The company then uses that capital to buy more Bitcoin. One investor gets the yield he wants; another retains the leveraged exposure to Bitcoin he wants; and the company becomes the intermediary between them. The “model” works spectacularly if Bitcoin’s appreciation over time substantially exceeds the cost of that capital.

That evolution turns the Bitcoin treasury company into something far more interesting than a corporate holder of Bitcoin. Capital markets themselves begin to become the business.

The 30% Hurdle

Suppose Bitcoin appreciates 30 percent a year (it is at 36% since Aug 10, 2020). Tim posed a simple question: “If we walked out the studio door and started an operating business with the goal of earning 30 percent annually, how hard would that be?” How much capital would it take? How long would it take to get there? And even if we succeeded, how difficult would it be to sustain those returns?

That creates an unusual problem for a Bitcoin treasury company. Every use of its capital has to compete with an “expected” return from simply buying more Bitcoin. A very good operating business might not be good enough. Even among the extraordinary companies shown below, only Nvidia exceeded Bitcoin’s return over the past six years. Six years is a short window for judging long-term returns, but long enough to illustrate how high a 30 percent hurdle really is.

If Bitcoin continues to reach new highs over the long run, the economics can be extraordinary. The Bitcoin treasury industry has already raised tens of billions of dollars of capital and credit, demonstrating substantial demand for securities built around the belief in long term Bitcoin appreciation.

The Bear Market Test

The first real test came when the market turned. Falling Bitcoin prices compressed the premiums many treasury companies had enjoyed, made raising new capital less attractive and forced management teams to make choices they hadn’t faced during the ascent.

Tim calls it an “incredibly insightful” and constructive bear market. Companies that had looked similar began responding differently. Some sold Bitcoin. Others changed their strategies. Financing became more difficult and investors seemed to all but stop rewarding companies simply for owning Bitcoin.

The downturn forced these companies to demonstrate what their strategies could accomplish when Bitcoin wasn’t doing the work for them. The bull market accelerated the Bitcoin treasury industry. The bear market began testing which ideas within it might endure and it provided an important test for digital credit. Financing didn’t disappear; it repriced. Yields of 12 or 13 percent impose a real cost that issuers must overcome, but demand for these securities has persisted even as Bitcoin fell. Digital credit built around Bitcoin remains a rounding error beside credit markets measured in the tens of trillions. If Bitcoin continues to prove itself as reliable, appreciating collateral, there is plenty of room for this market to grow.

Why Retail Distribution Mattered

The early Bitcoin treasury companies could not rely primarily on institutional investors. Institutional investors require infrastructure, regulatory clarity and established ways to evaluate the risks of an emerging asset. Retail investors face none of those constraints. They were willing to tolerate Bitcoin’s volatility in pursuit of higher returns.

That made access to retail investors unusually valuable. Tim pointed to the prominence of Bitcoin personalities such as Jack Mallers, Dylan Leclair, and Anthony Pompliano (among others) in launching treasury companies. Their large audiences weren’t just useful for publicity. They provided direct distribution to investors who already understood the Bitcoin thesis.

As Tim put it, “We’re in an attention economy.” For an emerging industry trying to raise capital before many institutions were ready to participate, attention became part of the business model.

Walk a Mile, See a Mile

While Tim was watching a new industry develop, he was also building a business around it. Interviews led to a podcast, which led to an online conference. His first lasted 21 hours, 21 minutes and 21 seconds, with 46 interviews. Sponsors appeared, relationships grew and online conferences eventually led to live ones.

None of this was the original plan. Tim describes his approach as “walk a mile, see a mile.” Each step revealed an opportunity he couldn’t have seen from the one before it.

There is a broader lesson in his experience. We can identify the businesses and jobs that technological change threatens because they already exist. It is much harder to imagine the ones that will be created around industries that don’t exist yet. Tim didn’t set out to occupy his current seat. The seat was being created at the same time he was finding it.

Thank you very much to our sponsors:

• Foundation.xyz: When it comes to protecting your Bitcoin, you shouldn’t have to simply trust the company that made your hardware wallet.

  Foundation believes in verifiability over reputation. Its firmware and apps are open source, its builds are reproducible, and both Passport Core and Passport Prime have undergone independent hardware security audits. The point isn’t simply that Foundation says its products are secure. It’s that you can verify that for yourself.

  Foundation is also expanding what users can control:

  • Passport Prime goes beyond Bitcoin. It can manage passwords and two-factor authentication codes, serve as a FIDO2 security key, and provide encrypted storage—all on a device you own and control. Foundation also plans to enable app sideloading, opening Prime to applications from independent developers.

  • Users can now add their own entropy. The latest Envoy and KeyOS releases allow users to introduce their own randomness when creating a seed, reducing reliance on the device itself to generate it.

  Learn more about Foundation → https://foundation.xyz/

• River Financial: If not cold storage for your Bitcoin, I’d highly recommend River for buying and holding- River is engineered to protect your Bitcoin over the long run. Use this link for discounts: https://river.com/signup?r=T7GGAF7G

I have been revisiting some of my earlier work on the American university system, prompted by two recently published surveys and a small-group discussion on AI and education that Dartmouth professor of mathematics and computer science Daniel Rockmore and I led last week.

In a 2025 Pew survey, 70 percent of Americans said the higher education system was headed in the wrong direction, up from 56 percent in 2020.

Then, in June 2026, Gallup found that only 38 percent expressed strong confidence in higher education, down from 57 percent in 2015 and 42 percent in 2025.

Yet Americans have not rejected education itself. 71% percent of graduates say their education was worth the cost, and 93 percent would choose to earn a degree again.

Taken together, the surveys suggest a widening divide between confidence in education and confidence in the institutions providing it. AI could widen that divide or force universities to demonstrate more clearly what an education is for.

Our recent discussion examined how artificial intelligence is challenging one of the university’s oldest functions: the delivery of knowledge. AI can explain difficult material, personalize instruction, and help people acquire skills at negligible cost. It may also enable new educational ventures to compete with parts of the university system. Listening again to my March 2025 conversation with Pano Kanelos, then president of the newly founded University of Austin, I was struck by how directly his diagnosis anticipated the pressures AI is now intensifying. AI appeared briefly in that interview, although its implications for teaching and learning were not central to our discussion. Beneath our conversation about cost, conformity, and curriculum was a question that has since become more urgent: What should a university produce that technology alone cannot?

Kanelos identified three crises in higher education. The first was a crisis of principles. Universities cannot pursue truth without open inquiry, civil discourse, and freedom of conscience, yet too many have allowed intellectual conformity to displace serious disagreement. The second was financial. Costs have risen while families have become less certain that a degree justifies the burden. The third was curricular. The structure of majors, departments, and academic specialization was designed more than a century ago for a world that no longer exists. AI sharpens each problem. Generated information increases the need for intellectual independence, inexpensive alternatives place additional pressure on the value of a costly degree, and accelerating changes in work expose the limits of curricula organized around fixed occupational skills. Kanelos and his colleagues eventually concluded that building a new university was easier than changing an established one. “That says everything you need to know about existing universities,” he told me, because starting one from the ground up is hardly easy.

Describing the conditions that led him to help found UATX, Kanelos said he had seen students and faculty increasingly “sleepwalking through this thing we call education.” In an intellectually homogeneous environment, students encounter fewer challenges to their assumptions and become more attentive to the answers expected of them. He compared the mind to a muscle, strengthened by tension, opposition, and stress. “A truly successful class is when you begin the class with a set of assumptions and end the class with a different set of assumptions,” he said. “That’s learning.” His conception of education centered on developing prudence, wisdom, judgment, and discernment, capacities that allow someone to continue learning as circumstances change. Rockmore and I returned repeatedly to curiosity as the quality that leads a student to question an answer, follow an unexpected connection, and continue beyond the first plausible explanation. AI gives students extraordinary tools for doing all three. It also makes avoiding the intellectual effort through which judgment develops far easier.

At the time of our conversation, UATX had organized its curriculum around these ideas. Students began with a common intellectual foundation, examining enduring questions across literature, philosophy, science, mathematics, politics, and history. A four-year Polaris Project gave each student a larger purpose toward which that learning could be directed. One student wanted to teach 100,000 people to read ancient Greek. Pursuing that goal required her to study teaching, technology, communication, organization, and ways of measuring whether she was succeeding. The educational value lay in building, failing, adjusting, and trying again. The project gave her education a direction that accumulating credits alone could not provide.

Rockmore described AI as a “gut-check moment” for higher education, a time for universities to ask, “What are we doing here?” As technology assumes more of the work involved in delivering information and producing answers, institutions will have to clarify which parts of education remain essential and how those purposes should be served. My prediction is that this reconsideration will lead to a proliferation of experimentation and entrepreneurship in education. AI lowers the cost of instruction, tutoring, assessment, and curriculum design, allowing new ventures to compete with individual university functions without recreating the enormous structure of a university.

New ventures will drive some of this change, and established universities can drive it as well. Entrepreneurship is as relevant to institutions as it is to startups. Universities that experiment with how they teach and undertake a disciplined examination of the value they provide may emerge stronger. They retain considerable advantages, including faculty expertise, research facilities, recognized credentials, alumni networks, and physical communities. AI can help them strengthen those advantages by freeing faculty to devote more attention to mentorship, intellectual challenge, and shared inquiry.

Universities that resist this examination may find the education system reconfiguring around them more quickly than they expect. Kanelos hoped UATX would inspire “a hundred new universities,” each approaching education differently. AI may now provide both the pressure and the means to create them.

Note: Pano Kanelos later left the presidency of UATX, which has since changed leadership. Its current direction lies beyond the scope of this essay.

Thank you very much to our sponsors:

• Foundation.xyz: When it comes to protecting your Bitcoin, you shouldn’t have to simply trust the company that made your hardware wallet.

  Foundation believes in verifiability over reputation. Its firmware and apps are open source, its builds are reproducible, and both Passport Core and Passport Prime have undergone independent hardware security audits. The point isn’t simply that Foundation says its products are secure. It’s that you can verify that for yourself.

  Foundation is also expanding what users can control:

  • Passport Prime goes beyond Bitcoin. It can manage passwords and two-factor authentication codes, serve as a FIDO2 security key, and provide encrypted storage—all on a device you own and control. Foundation also plans to enable app sideloading, opening Prime to applications from independent developers.

  • Users can now add their own entropy. The latest Envoy and KeyOS releases allow users to introduce their own randomness when creating a seed, reducing reliance on the device itself to generate it.

  Learn more about Foundation → https://foundation.xyz/

• River Financial: If not cold storage for your Bitcoin, I’d highly recommend River for buying and holding- River is engineered to protect your Bitcoin over the long run. Use this link for discounts: https://river.com/signup?r=T7GGAF7G

Last week, Representative Greg Casar of Texas introduced legislation intended to protect American workers from what he describes as the threat of mass unemployment caused by artificial intelligence. The proposal would tax certain uses of AI associated with job losses and use the proceeds to support employment. The premise reflects an increasingly common fear: if machines can perform more of the work people do today, there may eventually be far less work left for people to do.

There is plenty of reason to take that possibility seriously. AI is already performing tasks that recently required highly educated professionals. But legislation designed to protect existing jobs raises a more difficult economic question: does identifying the work AI can eliminate tell us very much about how much work people will ultimately have?

A View From Inside the Disruption

Stephen Lofchie has an unusually informed perspective on that question. One of the country’s leading lawyers in financial-services regulation, he has spent decades advising businesses operating in some of the most complicated areas of law. He also concluded years ago that much legal work was unnecessarily inefficient, began building his own regulatory databases, and eventually saw AI become capable enough that he helped create Caffrey.ai, an AI company for legal and regulatory professionals. He continues to practice law while helping develop technology capable of performing some of the work lawyers have traditionally been paid to do.

Lofchie has no illusion that AI will preserve every legal task or job. Yet when I asked what the profession might look like in the future, his prediction was not that lawyers would have less to do. “I think lawyers will be doing actually more,” he said. Understanding how both things can be true may tell us more about the future of work than counting the jobs AI can replace.

When Cheaper Work Creates More Work

Much of law appears particularly susceptible to AI because the work is built around language. As Lofchie puts it, “It’s words on a page and words over the telephone.” Lawyers absorb documents, research how laws have been interpreted, apply them to facts, and communicate their conclusions. Judgment is required, but much of the work leading to it has traditionally been performed by junior lawyers. AI can now do much of that “extremely well.”

If fewer lawyers can produce the same amount of legal work, fewer lawyers would seem to be needed. What that arithmetic misses is that demand is not fixed. “I think the cost of producing good legal work will go down,” Lofchie says, and “as costs go down there’ll be a demand for more of it.”

Clip from the interview.

Ty Greenberg, Chief Revenue Officer of Caffrey, who is leading the firm’s go-to-market effort, points to AI, crypto, and prediction markets, where new forms of economic activity do not fit neatly within existing laws and regulations, creating novel questions for businesses, lawyers, and regulators.

None of this establishes that increased demand will employ as many lawyers as AI displaces, and Lofchie makes no such prediction. But productivity lowers costs, lower costs increase demand, and solving problems through innovation invariably creates new problems to solve. The amount of human labor required to perform today’s legal work therefore tells us surprisingly little about how much legal work there will be tomorrow.

The Businesses That Don’t Exist Yet

The larger economic effect may come from the new activity that cheaper legal expertise makes possible. Regulatory complexity imposes costs before a business ever earns a dollar. “Think about a new startup,” Lofchie says. Can it afford law-firm partners charging $2,000 an hour simply to determine how to launch within the law? He calls that expense “a tremendous drain on a new business and the possibility of a startup, or even a new desk at a big firm.”

That is part of the opportunity Lofchie saw in creating Caffrey, which combines AI with the regulatory databases and systems he began developing years earlier. Greenberg says customers may see a profitable new business opportunity but first need to determine how to pursue it within the relevant regulations. Reducing the cost of answering those questions makes more opportunities worth investigating.

The implications extend beyond law. Every business has ideas that are never pursued because the expected return does not justify the cost of investigating them. For an entrepreneur, lowering that hurdle can make a new company viable. For an established financial institution, it might mean testing a new trading desk or entering a new market.

We can count the work AI replaces, but we cannot count the trading desks that were never opened, the companies that were never started, or the ideas that were never tested because finding out whether they would work cost too much.

What AI Makes More Valuable

As AI takes over more routine legal work, the lawyer’s role shifts toward what the technology does less well. Lofchie says lawyers, particularly younger ones, tend to overestimate the “pure words on a page element of lawyering.” Clients need more than the right legal answer; they need someone who understands their business and can exercise judgment about what to do with it. Greenberg says that even as AI performs more research and document analysis, demand for senior legal expertise is rising. For “bet the company” litigation, he notes, clients never wanted a junior associate making the critical decisions—and they aren’t likely to turn them over to AI either. As intelligence becomes cheaper, judgment, experience and trust should become more valuable. The resulting work should be better as well, because AI can perform far more research and analysis within the economic limits of an assignment, leaving the lawyer to apply human judgment to a richer body of information.

Lofchie is dismissive of lawyers who resist AI because it sometimes hallucinates or makes mistakes. “You don’t think first and second year associates make mistakes?” he asks. The appropriate response in either case is verification. Paradoxically, Lofchie believes that may become more important as AI improves. When a tool produces consistently impressive work, the temptation to trust it grows, making the occasional error easier to miss. That leaves another important role for expertise. The lawyer may spend less time producing the initial analysis but still needs to know enough to recognize when an authoritative-sounding answer is wrong.

Lofchie believes AI will “really reward active learners.” A lawyer can use it to understand a client’s business more deeply, explore an unfamiliar area of law, or quickly learn how another jurisdiction approaches the same problem. The advantage extends well beyond law. When information becomes cheaper and easier to acquire, the value of what someone already knows should decline relative to the ability to learn, ask the right questions and apply new knowledge. AI will therefore favor the curious over the complacent. The most successful knowledge workers will not be those who use AI simply to perform today’s jobs more efficiently, but those who use it to expand the range of problems they can solve.

The political appeal of protecting workers from technological change is hardly new. Yet the United States has prospered in large part because it has repeatedly allowed innovation to disrupt existing ways of working and doing business. Maintaining that balance is particularly difficult with AI, whose effects may be unusually rapid and far-reaching. As Congress considers taxing companies for replacing workers with AI, it should remember that protecting today’s jobs can also prevent tomorrow’s from being created.

Full Interview.

BITCOIN TREASURIES CONFERENCE — NEW YORK CITY

Interested in Bitcoin treasury companies? Most of the executives and investors shaping this industry will be at the Bitcoin Treasuries Conference in New York on September 28. Use code BOBD10 for 10% off your registration at https://bitcointreasuries.net/conference#tickets

Thank you very much to our sponsors:

• A message from Foundation

  When it comes to protecting your Bitcoin, you shouldn’t have to simply trust the company that made your hardware wallet.

  Foundation believes in verifiability over reputation. Its firmware and apps are open source, its builds are reproducible, and both Passport Core and Passport Prime have undergone independent hardware security audits. The point isn’t simply that Foundation says its products are secure. It’s that you can verify that for yourself.

  Foundation is also expanding what users can control:

  • Passport Prime goes beyond Bitcoin. It can manage passwords and two-factor authentication codes, serve as a FIDO2 security key, and provide encrypted storage—all on a device you own and control. Foundation also plans to enable app sideloading, opening Prime to applications from independent developers.

  • Users can now add their own entropy. The latest Envoy and KeyOS releases allow users to introduce their own randomness when creating a seed, reducing reliance on the device itself to generate it.

  Learn more about Foundation → https://foundation.xyz/

• River Financial: If not cold storage for your Bitcoin, I’d highly recommend River for buying and holding- River is engineered to protect your Bitcoin over the long run. Use this link for discounts: https://river.com/signup?r=T7GGAF7G

brings together several of those disciplines into a different way of looking at the world.

“I’m basically interested in four subjects that are all interlinked in my head,” he told me. “Economics, physics, mathematics, and computer science.”

Four Subjects, One Lens

Whether discussing free will, artificial intelligence, capitalism, taxation, scientific research, or education, Varma approached each subject through the same interdisciplinary lens, asking whether an explanation remained consistent across disciplines and, ultimately, whether it survived contact with reality.

Varma traces that approach back to his childhood in India. He grew up wondering how a culture that celebrated wealth and entrepreneurship could also be one of the poorest countries in the world. The contradiction irritated him, and that irritation became a habit of mind.

“It’s all born out of irritation,” he said. “The irritation came from getting absurd answers to simple questions.” So he began asking why. That habit eventually became the organizing principle behind his writing as well as his book on free will, The Science of Free Will.

His standard for judging ideas is demanding. They should be tested against evidence, remain consistent with established knowledge, and explain more than the ideas they replace. As he put it, “The more radical the idea, the harder you really have to test it.”

That same standard leads him to distrust explanations that survive only within a single discipline. An economic theory should not contradict what we know about incentives. A theory of artificial intelligence should be consistent with physics and computer science. A philosophical claim should survive contact with mathematics. For Varma, understanding comes not from viewing a problem through one lens, but from seeing whether the same explanation continues to hold when examined from several directions at once.

Where Choice Comes From

His theory of free will illustrates the approach. Physics tells us that every particle in our bodies obeys the laws of nature. If those laws completely determine the future, where does choice come from?

His answer draws simultaneously from physics, computer science, and philosophy.

“The gap between the laws existing and making predictions from the laws,” he explained, “is the source of free will.”

The key idea is what computer scientist Stephen Wolfram calls computational irreducibility. Some systems are so complex that their future cannot be shortcut. Even if the underlying rules are completely deterministic, the only way to know what they will do is to let them unfold. Determinism, in other words, does not imply predictability.

Teaching Principles, Not Playing Whack-a-Mole

That conclusion extends naturally to artificial intelligence. If AI systems operate under the same physical laws and exhibit the same computational complexity, Varma argues there is no fundamental reason to believe they should be predictable either.

Importantly then, instead of trying to control increasingly capable AI by layering on more restrictions like whack-a-mole, he believes the better approach is to teach it guiding principles from the outset.

“You have to teach it principles,” he said. “The more you constrain a system, the less optimal it is.”

Incentives All the Way Down

The same framework shapes Varma’s views on economics and public policy. Rather than beginning with ideology, he begins with incentives.

“I think the biggest problem is that we don’t teach our students economics,” he said. “Do you really understand incentives? Incentives matter. All the rest is commentary.”

That perspective leads him to see a common thread running through many of today’s policy debates. Housing shortages, in his view, are less a failure of markets than of regulations that make building unnecessarily difficult. Slower economic growth reflects tax and regulatory systems that discourage investment and entrepreneurship. Political polarization stems from electoral rules that reward extremes instead of broad consensus. Even scientific progress depends on incentives: stable funding, intellectual freedom, and institutions that reward discovery over politics.

His views on taxation follow the same logic. Capital is not simply accumulated wealth. It is the stock of tools, equipment, and infrastructure that raises what each worker can produce. Taxing capital reduces the investment that builds this stock, and with it the living standards that follow. As he put it, “The more you tax capital, the less of it you have.”

The Long Game

Varma argues that many of the technologies that define modern life began as basic scientific research with no obvious commercial application. GPS, for example, depends on Einstein’s general theory of relativity, developed decades before anyone imagined satellite navigation. No private company could justify investing in research that might not pay off for seventy-five years. That, he believes, is one of government’s most important roles.

His concern is that the United States is beginning to weaken precisely the institutions that made it the world’s scientific leader. He points to funding uncertainty, political interference, and the loss of researchers to other countries as incentives that discourage discovery. Scientific progress, like economic growth, depends on institutions that reward long-term investment.

He sees this not simply as a question of research budgets but of national prosperity. Discoveries made today become the industries, companies, and technologies that raise living standards decades from now. Because those benefits often take decades to appear, they are easy for policymakers to undervalue.

That long-term perspective also shapes how he ranks America’s challenges. Rather than focusing primarily on more popular sources of pessimism and concern like healthcare or education, he believes artificial intelligence will improve both. The more important constraints, in his view, are housing and scientific research.

Many thinkers develop deep expertise within a single discipline. Varma connects disciplines that are usually treated separately and asks whether the same explanation survives each of them. As specialization continues to expand human knowledge, that kind of synthesis may become increasingly valuable. It is a perspective worth understanding.

Full interview.

Thank you very much to our sponsors:

• Foundation: Secure your digital life—from passwords, 2FA, files and Bitcoin with Passport Prime, an all-in-one personal security platform. Visit them at https://foundation.xyz/

• River Financial: If not cold storage for your Bitcoin, I’d highly recommend River for buying and holding- River is engineered to protect your Bitcoin over the long run. Use this link for discounts: https://river.com/signup?r=T7GGAF7G

“Humanity is divine.”

That was Chris Sullivan’s answer when I asked how philosophy, political theory, monetary history, and spirituality shaped his worldview.

The answer also explains the tension running through our conversation. Chris has enormous faith in human creativity and very little faith in the institutions governing modern life. His argument is that those institutions have been distorted by the monetary system beneath them.

Most criticism of fiat money focuses on inflation. Chris’s critique is broader. Chris Sullivan’s critique begins much earlier. In his view, the modern central banking system is the root cause. Once money can be created without meaningful constraint, debt expands, incentives change, institutions adapt, and society gradually changes with them.

Debt makes political choices easier to postpone. Governments can expand programs without imposing the full cost through immediate taxation. Borrowing obscures tradeoffs and pushes obligations into the future. As Chris put it, “The ridiculousness and ludicrousness of the magnitude of debt ultimately only has one resolution.” Today's debt burden cannot realistically be repaid under the existing system. He envisions some form of global debt reset which he calls a debt jubilee followed by a monetary system no longer dependent on continual money creation. Bitcoin, in his view, is the leading candidate to provide that foundation.

His concern is not limited to the size of the debt. It is the behavior that develops around a system capable of producing it.

Government expands because financing is available. Bureaucracies survive without proving their value. Corporations become more focused on balance-sheet engineering. Capital flows toward businesses positioned to benefit from credit creation and government spending. Saving loses purchasing power while borrowing is repeatedly rescued.

Debt does more than finance activity. It changes which activities receive the greatest rewards. That distinction is central to Chris’s view of prosperity. Government spending may increase measured GDP, but he rejects the idea that spending itself creates wealth. “What they have created is concentration of wealth,” he said. In his framework, prosperity comes from entrepreneurship, competition, technological progress, and the productive use of capital.

This is also why Chris views Bitcoin as more than an investment. He first dismissed it as “magic internet money,” but later concluded that “Bitcoin could recapitalize the world to end the central banks.” He sees a fixed monetary supply, transparent settlement system, and asset outside political discretion as a possible foundation for a different incentive structure.

His conviction that the current system will imminently break is stronger than mine. Fiat systems can survive longer than their critics expect, especially when governments retain the ability to tax, regulate, borrow, and refinance. Collapse is not the only path to change. Monetary systems might actually be able to erode gradually as people adopt better alternatives, though that hasn’t typically happened with former global reserve currencies.

Chris nevertheless presents a serious question. If money influences incentives, and incentives influence institutions, then the quality of money may affect far more than prices or investment returns. It may shape the character of a society. While he worries about the correction process and the potential for revolutionary circumstances typical of fourth turnings, his conclusion is not as dark as much of his diagnosis.

“The great news is that we have immense prosperity and abundance already,” he said, and that if we do wind up suffering thrugh a transition period, what we emerge into will be a golden era. Human beings continue to build, invent, adapt, and solve problems even when institutions work against them.

Chris believes better money would release more of that potential. Whether Bitcoin ultimately plays the role he envisions, his underlying argument deserves consideration: prosperity depends not only on what people are capable of creating, but also on the incentives surrounding them.

Full interview.

Thank you very much to our sponsors:

• Foundation: Secure your digital life—from passwords, 2FA, files and Bitcoin with Passport Prime, an all-in-one personal security platform. Visit them at https://foundation.xyz/

• River Financial: If not cold storage for your Bitcoin, I’d highly recommend River for buying and holding- River is engineered to protect your Bitcoin over the long run. Use this link for discounts: https://river.com/signup?r=T7GGAF7G

Read more

This week, two seemingly unrelated things intersected.

• A friend mentioned that his son decided he never wants to have children because he doesn’t want to bring them into a world like ours.

• ’s Substack arrived with the subject line, “A Philosopher’s One-Word Theory to Explain Why the World Feels So Weird.” It linked to his podcast conversation with University of Chicago philosopher , who is developing an idea she calls the “Unicontext.”

Thompson introduced the discussion by describing it as “a theory of everything that explains why people feel like everything sucks these days.”

That’s an ambitious description. The conversation lived up to it.

Callard’s central idea is the Unicontext. For most of human history, people lived in many different contexts. Home had its own norms. School had its own. Church, work, the neighborhood, and countless other communities all carried different expectations about how people should behave. Today, technology has increasingly merged those separate worlds into one enormous shared context. Much of modern life now takes place in a single global room where millions of people, ideas, tragedies, successes, arguments, and opinions compete for attention simultaneously.

The power of the idea lies in the range of questions it attempts to explain.

Why Bad News Travels Better Than Good News

Callard argues that goodness is often local while suffering is universal. One culture’s idea of a wonderful meal, a meaningful tradition, or a successful life may not translate easily to another. Violence, illness, grief, injustice, and death require no translation. In a global conversation, suffering becomes the language everyone understands. That helps explain why negative stories spread so easily and why they so often dominate public attention.

Why Identity Crowds Out Character

Character can only be revealed over time. Honesty, courage, generosity, and integrity become visible only after interacting with or knowing someone across many situations. Identity is immediately visible. In a world increasingly filled with strangers interacting across a single shared context, identity naturally becomes easier to discuss than character.

Why Comparison Never Ends

The Unicontext also helps explain why comparison has become so pervasive. When separate contexts merge, comparison becomes almost unavoidable. Callard illustrates this with the example of two school districts. As long as families attended their local schools, there was little reason to compare them. Once both schools become realistic options, everything starts to be measured: graduation rates, AP courses, college admissions, teacher quality, test scores. Rankings become increasingly important. Institutions begin optimizing for the same metrics, and over time they become more alike. The same dynamic extends well beyond education, encouraging businesses, careers, communities, and even people to converge toward whatever the world has decided to measure.

A Question Worth Exploring

One question remained after the conversation ended. If the Unicontext helps explain why humanity’s problems have become universally visible, does it also make humanity’s progress more difficult to recognize?

Progress usually arrives gradually. Tragedy is concentrated. It demands attention.

Perhaps the same framework also helps explain why genuine heroes seem harder to find. Progress is hard to see, and so, it seems, are the people driving it. There is no shortage of remarkable people accomplishing extraordinary things today, yet admiration often struggles to compete with comparison.

Back to My Friend’s Son

He isn’t wrong about what he sees. The suffering that reaches him is real. But the world he is judging isn’t the whole world. It is the Unicontext’s version of it, one that naturally amplifies the experiences every human being immediately understands: violence, injustice, illness, and death.

Progress spreads differently. It arrives quietly through millions of small advances and billions of everyday problems being solved: diseases eradicated, children surviving illnesses that once claimed countless lives, technologies that quietly improve everyday life, and acts of kindness that rarely extend beyond a family or neighborhood. Those stories simply don’t compete very well inside a global conversation where comparison and tragedy command our attention. That doesn’t mean the good isn’t there. It may simply mean it has become much harder to see.

Professor Callard’s theory offers a fascinating way to think about the modern world, and it raises questions that extend well beyond the conversation itself. For anyone interested in why modern life often feels so different from the world our parents experienced, Derek Thompson’s conversation with Agnes Callard is well worth an hour of your time.

Thank you very much to our sponsors:

• Foundation: Secure your digital life—from passwords, 2FA, files and Bitcoin with Passport Prime, an all-in-one personal security platform. Visit them at https://foundation.xyz/

• River Financial: If not cold storage for your Bitcoin, I’d highly recommend River for buying and holding- River is engineered to protect your Bitcoin over the long run. Use this link for discounts: https://river.com/signup?r=T7GGAF7G

Thank you very much to our sponsors:

Foundation: Protect your Bitcoin using the latest off-line technology with easy-to-use Passport. Their next product, Passport Prime will protect all passwords off-line with even-better technology. https://foundation.xyz/

River Financial: If not cold storage for your Bitcoin, I’d highly recommend River for buying and holding- River is engineered to protect your Bitcoin over the long run. Use this link for discounts:
https://river.com/signup?r=T7GGAF7G

Timestamps:
00:00 AI Isn't Thinking
02:00 Why AI Won't Replace People
03:08 AI Has No Motivation
04:53 Garbage In, Garbage Out
06:58 What Is Context Rot?
08:16 How Large Language Models Actually Work
10:23 Why AI Gives Different Answers
11:04 Why AI Hallucinates
12:40 Can You Trust AI?
14:18 Prediction vs. Reasoning
15:11 Why AI Only Sounds Smart
16:50 AI's Biggest Real-World Applications
18:04 The Spreadsheet Revolution Analogy
18:54 Will AI Replace Jobs?
20:31 Permanent Due Diligence
24:35 AI in Healthcare & Education
30:36 Personalized Education
32:24 How AI Could Improve Society

The Stress Test

For several uneasy weeks, the market appeared to deliver its verdict on one of the fastest-growing experiments in modern finance. Preferred securities issued by Bitcoin Treasury Companies traded sharply below their intended range.

Critics argued the business model had been exposed, and social media filled with predictions of Bitcoin’s demise. The question was no longer whether these companies represented financial innovation. It was whether the innovation had already failed.

I sat down with Strategy CEO Phong Le and Strive CEO Matt Cole just as both companies were responding to that turbulence. Both spoke candidly about what they had learned and how they intended to adapt. The more interesting story is how the market forced those conclusions in the first place.

Financial innovation rarely unfolds in a straight line. New products arrive with a set of assumptions, investors test those assumptions with real money, and management responds. Sometimes the process exposes a fatal flaw. Sometimes it produces a better product. Either way, markets discover information neither critics nor management possessed when the experiment began.

When Markets Disagree

One moment from our discussion captured that process well. Strategy had reduced its cash reserves to retire a significant amount of convertible debt, believing a stronger balance sheet compensated for the lower cash position. Investors saw it differently.

“We did not realize how important that U.S. dollar reserve was to our Stretch investors,” Le told me. Institutional investors, he said, placed considerably more value on readily available cash than management had expected, despite the company’s significant Bitcoin holdings.

Management believed one form of financial strength counted most. Investors preferred another. Neither side won the argument through debate. Prices settled it immediately. Strategy rebuilt its cash reserve to levels higher than before and earmarked those funds for dividend and interest obligations. Markets often become participants in product development, revealing assumptions that only real capital can test.

Innovation Happens in Public

Looking backward, successful innovations often appear inevitable. Looking forward, they almost never do. The first iPhone lacked an App Store. SpaceX reached orbit only after a series of visible failures. Early versions rarely survive unchanged, because markets reveal information designers could not fully anticipate.

Cole sees financial innovation through the same lens. “I don’t think it’s fair to expect V1 to be the best final form of innovation,” he said. “That is literally why innovation exists.” Products become durable, in his view, when companies absorb market feedback and rebuild around it.

Whether Bitcoin Treasury Companies eventually follow that path remains uncertain. The recent redesigns could prove to be meaningful improvements, or they could postpone larger problems. Markets have begun asking more useful questions than they were asking a month ago.

Looking Past the Labels

When I asked Le about the criticism that Bitcoin Treasury Companies are “merely financial engineering,” I expected him to distance himself from the phrase. Instead, he embraced it. “Financial engineering is the practical application of finance,” he said, comparing it to mortgage-backed securities, which connected homebuyers with investors seeking income. Like any engineering discipline, it can produce useful or harmful outcomes. The relevant question is whether it solved a real economic problem.

Cole reached a similar conclusion by way of Liar’s Poker, the book on the mortgage-backed securities market that shaped his early thinking. Financial structures create value, he said, by matching investors with different objectives. “If you are doing your job well in financial engineering, the sum of the parts are greater than the whole because you’re providing unique and valuable solutions to select investors.”

History offers reasons for caution too. Mortgage-backed securities became synonymous with the financial crisis once poor underwriting and excessive leverage overwhelmed them. Some innovations change markets for good. Others delay the recognition of risk.

Why Borrow? Why Lend?

The first two questions deserve particular attention. Why would a company borrow money to buy Bitcoin, and why would anyone provide the financing?

To many observers, the first question answers itself. Borrowing to buy a volatile asset sounds speculative. Yet borrowing to acquire assets expected to appreciate is one of the oldest activities in finance, from families buying homes to private equity firms acquiring companies. The calculation is the same each time: if the long-term return exceeds the cost of the financing, borrowing creates value. If not, it destroys value.

Le described Bitcoin Treasury Companies in similar terms. If Bitcoin is “a unique underlying asset that can appreciate thirty, forty percent a year,” he said, the next question is why public companies should not own it. Treasury Companies exist to provide exposure for investors who cannot or choose not to own Bitcoin directly.

The more interesting question may be why anyone would lend. The answer has less to do with enthusiasm for Bitcoin than in the diversity of capital markets. Pension funds, insurance companies, and endowments often operate under mandates restricting what they can own, some barred from Bitcoin directly, others required to hold publicly traded or income-producing securities.

As Cole noted, fixed-income investors ultimately care about total return, not whether that return comes through coupons, capital appreciation, or some combination of both. If Bitcoin-based securities eventually provide attractive risk-adjusted returns within those mandates, capital will continue flowing toward them. If they do not, it won’t. Markets will make that determination over time.

Bitcoin Doesn’t Need Them

The distinction between Bitcoin and the companies built around it is often overlooked. Critics frequently treat Bitcoin Treasury Companies as though they are essential to Bitcoin’s future. I asked both executives whether that premise was even true.

Neither executive hesitated. “No,” Le replied. “Does Bitcoin need us? No. Do our investors need us? Yes.” Companies like Strategy expand access for investors who otherwise could never own Bitcoin, because of mandates or regulatory constraints. Bitcoin would likely keep appreciating without them, though perhaps more slowly.

Cole answered from a different angle, “I don’t think Bitcoin needs the United States to have a strategic Bitcoin reserve. I think the United States would be stronger if it had a strategic Bitcoin reserve. And that similar principle applies to corporations.” Treasury Companies may strengthen Bitcoin’s adoption, but that differs from claiming Bitcoin depends on them.

Much of the public debate treats Bitcoin and Bitcoin Treasury Companies as though they succeed or fail together. These companies are wagering on Bitcoin’s future. They are not deciding it. If their assumptions prove correct, they may create substantial value for shareholders. If wrong, the companies will struggle regardless of how innovative their capital structures may be.

Time Is the Judge

The recent turbulence may eventually be remembered as a volatile chapter in the early history of a new asset class, or as the first sign that the business model contains weaknesses that grow more apparent over time. Neither conclusion is available yet.

What the episode has already revealed is how financial innovation evolves. Management teams begin with assumptions, investors test those assumptions with real capital. and prices reveal where expectations differ. Companies adjust, investors respond again, and the cycle repeats until markets settle whether a product deserves wider acceptance or a place among the ideas that failed to endure. That process is usually slower than either enthusiasts or critics prefer and Bitcoin Treasury Companies have not reached the end of it.

Only time will say which one Bitcoin Treasury Companies turn out to be.

Here’s the full interview.

Thank you very much to our sponsors:

• Foundation: Secure your digital life—from passwords, 2FA, files and Bitcoin with Passport Prime, an all-in-one personal security platform. Visit them at https://foundation.xyz/

• River Financial: If not cold storage for your Bitcoin, I’d highly recommend River for buying and holding- River is engineered to protect your Bitcoin over the long run. Use this link for discounts: https://river.com/signup?r=T7GGAF7G