Are LLMs overconfident? (just like humans)
Can LLMs accurately adjust their confidence when facing opposition? Building on previous studies measuring calibration on static fact-based question-answering tasks, we evaluate Large Language Models (LLMs) in a dynamic, adversarial debate setting, uniquely combining two realistic factors: (a) a multi-turn format requiring models to update beliefs as new information emerges, and (b) a zero-sum structure to control for task-related uncertainty, since mutual high-confidence claims imply systematic overconfidence. We organized 60 three-round policy debates among ten state-of-the-art LLMs, with models privately rating their confidence (0-100) in winning after each round. We observed five concerning patterns: (1) Systematic overconfidence: models began debates with average initial confidence of 72.9% vs. a rational 50% baseline. (2) Confidence escalation: rather than reducing confidence as debates progressed, debaters increased their win probabilities, averaging 83% by the final round. (3) Mutual overestimation: in 61.7% of debates, both sides simultaneously claimed >=75% probability of victory, a logical impossibility. (4) Persistent self-debate bias: models debating identical copies increased confidence from 64.1% to 75.2%; even when explicitly informed their chance of winning was exactly 50%, confidence still rose (from 50.0% to 57.1%). (5) Misaligned private reasoning: models’ private scratchpad thoughts sometimes differed from their public confidence ratings, raising concerns about faithfulness of chain-of-thought reasoning. These results suggest LLMs lack the ability to accurately self-assess or update their beliefs in dynamic, multi-turn tasks; a major concern as LLMs are now increasingly deployed without careful review in assistant and agentic roles.
That is by Pradyumna Shyama Prasad and Minh Nhat Nguyen. Here is the associated X thread. Here is my earlier paper with Robin Hanson.
o3 pro
It is very, very good. Hallucinates far less than other models. Can solve economics problems that o3 cannot. It can be slow, but that is what we have Twitter scrolling for, right? While we are waiting for o3 pro to answer a query we can read about…o3 pro.
Sam Altman comments on his blog. And Nabeel comments.
Tuesday assorted links
1. Mapping the power elite in Greenland.
2. Did “tanking” work for the 76ers?
3. Scott Sumner on the demise of the penny.
4. Claims from Nievas and Piketty: “A 20% increase in commodities prices from 1970 to 2025 period – very moderate compared to the very low market exchange rates of many global South countries – could turn Sub-Saharan Africa into a larger creditor than East Asia by 2025.” Here are some comments.
5. Extremely life-like dolls in the culture that is Brazil (NYT). Ross Douthat, telephone!
6. Justin Sandefur estimates the foreign lives costs of proposed budgetary changes to foreign aid. Here is a redux of my 2023 column on PEPFAR.
Racial Disparities in Mortality by Sex, Age, and Cause of Death
Racial differences in mortality are large, persistent and likely caused, at least in part, by racism. While the causal pathways linking racism to mortality are conceptually well defined, empirical evidence to support causal claims related to its effect on health is incomplete. In this study, we provide a unique set of facts about racial disparities in mortality that all theories of racism and health need to confront to be convincing. We measure racial disparities in mortality between ages 40 and 80 for both males and females and for several causes of death and, measure how those disparities change with age. Estimates indicate that racial disparities in mortality grow with age but at a decreasing rate. Estimates also indicate that the source of racial disparities in mortality changes with age, sex and cause of death. For men in their fifties, racial disparities in mortality are primarily caused by disparities in deaths due to external causes. For both sexes, it is racial disparities in death from healthcare amenable causes that are the main cause of racial disparities in mortality between ages 55 and 75. Notably, racial disparities in cancer and other causes of death are relatively small even though these causes of death account for over half of all deaths. Adjusting for economic resources and health largely eliminate racial disparities in mortality at all ages and the mediating effect of these factors grows with age. The pattern of results suggests that, to the extent that racism influences health, it is primarily through racism’s effect on investments to treat healthcare amenable diseases that cause racial disparities in mortality.
In other words, much of the discourse on this topic is quite off. That is from a new NBER working paper by Robert Kaestner, Anuj Gangopadhyaya, and Cuiping Schiman.
Hayek Goes Supersonic
When I post about lifting the ban on supersonic flight, smart commenters show up with charts: optimal fuel burn is at Mach 0.78–0.84, they say, or no one wants to pay thousands to save a few hours. Maybe. But my reply is always the same: Bottled water!
In 2024, Americans spent $47 billion a year on H₂O that they could get for nearly free. That still boggles my mind—but bottled water has passed the market test. I argue for lifting the SST ban, and similar policies, not because we know supersonics will work but because we don’t. Hayek reminds us that competition is a discovery procedure. Like science, markets generate knowledge by experiment—hypotheses are posted as prices, and the public accepts or rejects them through revealed preference. Fred Smith’s FedEx plan got a “C” in the classroom, but the market graded the experiment and returned an A in equity. Theory is great, but just as in science, there is no substitute for running the experiment.
The great Sly Stone has passed away
There is indeed a riot going on, do listen to the whole album, one of the very best. Here is a NYT obituary.
Adam Tooze on European military spending
Now, you might think that the US figure is inflated by the notorious bloat within the American military-industrial complex. I would be the last person who would wish to minimize that. But the evidence suggests that the bias may be the other way around. American defense dollars likely go further than European euros.
Look for instance at the price of modern, third-generation battle tanks and the cost of self-propelled howitzers, which have been key to the fighting in Ukraine. German prices are far higher than their American counterparts.
And, as work by Juan Mejino-López and Guntram B. Wolff at the Bruegel policy think tank has shown, these higher costs have to do with smaller procurement runs and smaller procurement runs are, in turn, tied to the fragmentation of Europe’s militaries and their strong preference for national procurement.
Right-now there is often lamentation about the tendency of European militaries to import key weapons systems from the US. And there is, of course, plenty of geopolitical and political maneuvering involved, for instance, in Berlin’s initiative to build an air defense system heavily reliant American and Israeli missiles. As the data show, Germany does have a strong preference for imports from the US rather than its European neighbors.
But, on average, across the entire defense budget, the besetting sin of European militaries is not that they rely too heavily on foreign weapons, but that they import not enough. They are too self-sufficient. The problem is not that Germany buys too many weapons from the US, but that it buys too many in Germany.
National fragmentation creates the balkanized defense market, the inefficient proliferation of major weapons systems and in terms of global industrial competition, the small size of European defense contractors.
Here is the full Substack, very good throughout. Via Felipe.
Monday assorted links
1. The motherhood mental health advantage.
2. “Cape Town residential property prices have risen by 160% since the start of 2010.” (FT)
3. Can taking photos impair your memories of events? (2021)
4. Short Greg Brockman video on the checks and balances in our AI future.
5. Criticism of the capabilities of AI reasoning models. And Kevin Bryan responds. And from Rohit.
7. Jennifer Burns reviews the new William F. Buckley biography (NYT).
The High Cost of Self-Sufficiency
Mike Riggs and his wife dreamed of returning to the land. It wasn’t as easy as it looks on Tik-Tok:
How many square feet of raised beds do you need to meet a toddler’s strawberry demand? I still don’t know. We dedicated 80 square feet to strawberries last season. The bugs ate half our harvest, and the other half equaled roughly what our kid could eat in a week.
Have you ever grown peas? Give them something to climb, and they’ll stretch to the heavens. Have you ever shelled peas? It is an almost criminal misuse of time. I set a timer on my phone last year. It took me 13 minutes to shell a single serving. Meanwhile, a two-pound bag of frozen peas from Walmart costs $2.42. And the peas come shelled.
…In addition to possums and deer, we’ve faced unrelenting assaults from across the eukaryotic kingdoms: the tomato hornworm caterpillar, the cabbage looper caterpillar, the squash vine borer, the aphid, the thrip, the earwig and the sowbug; cucurbit downy mildew, powdery mildew, collar rot, black rot, sooty mold, botrytis gray mold and stem canker; the nematode, the gray garden slug, the eastern gray squirrel, the eastern cottontail rabbit and the groundhog. All of these organisms reside in the North Carolina Piedmont and like to eat what we eat. Many of them work toward this existential goal while humans sleep, which is why the North Carolina State Agriculture Extension advises growers to inspect their plants at night. No, thank you.
…. In the early 1900s, one of my paternal great-grandfathers moved from urban Illinois to a homestead in Oklahoma. Our only picture of him was taken shortly before the Dust Bowl destroyed his farm. After his farm failed, he abandoned my great-grandmother and their children and migrated to California with thousands of other Okies. When my crops fail, I go to Whole Foods.
Some good lessons here in self-sufficiency, comparative advantage and the productivity of specialization and trade. Of course, it might have been easier for Mike had he read Modern Principles:
How long could you survive if you had to grow your own food? Probably not very long. Yet most of us can earn enough money in a single day spent doing something other than farming to buy more food than we could grow in a year. Why can we get so much more food through trade than through personal production? The reason is that specialization greatly increases productivity. Farmers, for example, have two immense advantages in producing food compared with economics professors or students: Because they specialize, they know more about farming than other people, and because they sell large quantities, they can afford to buy large-scale farming machines. What is true for farming is true for just about every field of production—specialization increases productivity. Without specialization and trade, we would each have to produce our own food as well as other goods, and the result would be mass starvation and the collapse of civilization.
Oh, and by the way, don’t forget Adam Smith, “What is prudence in the conduct of every private family can scarce be folly in that of a great kingdom.”
Deport Dishwashers or Solve All Murders?
I understand being concerned about illegal immigration. I definitely understand being concerned about murder, rape, and robbery. What I don’t understand is being more concerned about the former than the latter.
Yet that’s exactly how the federal government allocates resources. The federal government spends far more on immigration enforcement than on preventing violent crime, terrorism, tax fraud or indeed all of these combined.
Moreover, if the BBB bill is passed the ratio will become even more extreme. (sere also here):
Don’t make the mistake of thinking that immigration enforcement is about going after murderers, rapists and robbers. It isn’t. Indeed, it’s the opposite. ICE’s “Operation At Large” for example has moved thousands of law enforcement personnel at Homeland Security, the FBI, DEA, and the U.S. Marshals away from investigating violent crime and towards immigration enforcement.
I’m not arguing against border enforcement or deporting illegal immigrants but rational people understand tradeoffs. Do we really want to spend billions to deport dishwashers from Oaxaca while rapes in Ohio committed by US citizens go under-investigated?
Almost half of the murders in the United States go unsolved (42.5% in 2023). So how about devoting some of the $167 billion extra in the BBB bill to say expand the COPS program and hire more police, deter more crime and to use Conor Friedersdorf’s slogan, solve all murders. Back of the envelope calculations suggest that $20 billion annually could fund roughly 150 k additional officers, a ~22 % increase, deterring some ~2 400 murders, ~90 k violent crimes, and ~260 k property crimes each year. Seems like a better deal.
Early North America was more agricultural than we had thought?
A new study has found that a thickly forested sliver of Michigan’s Upper Peninsula is the most complete ancient agricultural location in the eastern United States. The Sixty Islands archaeological site is recognized as the ancestral home of the Menominee Nation. Known to the members of the tribe as Anaem Omot (Dog’s Belly), the area is a destination of pilgrimage, where remains of the settlement date to as far back as 8,000 B.C.
Located along a two-mile stretch of the Menominee River, Sixty Islands is defined by its cold temperatures, poor soil quality and short growing season. Although the land has long been considered unsuitable for farming, an academic paper published on Thursday in the journal Science revealed that the Menominee’s forbears cultivated vast fields of corn and potentially other crops there.
Here is more from the New York Times. The data came from drone-based LIDAR, which has been possible for only a few years. Most likely, much of the early history of the New World will need to be rewritten, as similar efforts are being pursued elsewhere.
The wisdom of Ezra Klein
What both forms of populism share is a tendency to treat virtue as a fixed property of groups and policy as a way of redistributing power from the disfavored to the favored. When I said we needed “a liberalism that builds,” David Dayen, the editor of The American Prospect, responded that “we need a liberalism that builds power” and that the way to get it is for the government “actively supporting the very groups that have been left out of past economic transitions, building the necessary coalition for long-term transformation.”
Every policy, in this telling, has two goals. One is the goal of the policy or the project; perhaps you’re trying to decarbonize the economy or build affordable housing or increase competition in the market for hearing aids. But the other is the redistribution of power among groups: Does this policy leave unions stronger or weaker? Environmental justice groups? Corporations?
Under the populist theory of power, bad policy can be — and often is — justified as good politics. In California, the California Environmental Quality Act is defended by unions that use it to “greenmail” all manner of projects. CEQA is meant to protect the environment, but the threat of unending litigation can be used to win non-environmental concessions on virtually any building project in California.
Here is the full NYT piece, interesting throughout, for instance:
My view of power is more classically liberal. In his book “Liberalism: The Life of an Idea,” Edmund Fawcett describes it neatly: “Human power was implacable. It could never be relied on to behave well. Whether political, economic or social, superior power of some people over others tended inevitably to arbitrariness and domination unless resisted and checked.”
Worth a ponder.
The convent where the Salamancans wrote their great works
Convent San Esteban. It is still there, you can just walk right in, though not between 2 and 4, when the guards have off. Arguably the Salamancans were the first mature economists, and the first decent monetary theorists, as well as being critically important for the foundations of international law, natural rights, and anti-slavery arguments. It is also difficult to find issues where they were truly bad.
You can just walk right in, and you should.
Sunday assorted links
1. “Particulate matter local air pollution from road transport is much less due to petrol exhaust than most people think. More is due to brake, tyre, and road wear.” There is some disputation going on here, but at the very least there is something directionally correct in this point.
2. Milei gives Hayek’s Fatal Conceit to the Pope.
3. Walton heirs to start a new STEM-focused university in Arkansas.
4. Mathematicians encounter o4-mini.
5. What will happen to the federal GSA art collection? (NYT…I am, by the way, happy to auction it off).
Supersonics Takeoff!
In Lift the Ban on Supersonics I wrote:
Civilian supersonic aircraft have been banned in the United States for over 50 years! In case that wasn’t clear, we didn’t ban noisy aircraft we banned supersonic aircraft. Thus, even quiet supersonic aircraft are banned today. This was a serious mistake. Aside from the fact that the noise was exaggerated, technological development is endogenous.
If you ban supersonic aircraft, the money, experience and learning by doing needed to develop quieter supersonic aircraft won’t exist. A ban will make technological developments in the industry much slower and dependent upon exogeneous progress in other industries.
When we ban a new technology we have to think not just about the costs and benefits of a ban today but about the costs and benefits on the entire glide path of the technology
In short, we must build to build better. We stopped building and so it has taken more than 50 years to get better. Not learning, by not doing.
… I’d like to see the new administration move forthwith to lift the ban on supersonic aircraft. We have been moving too slow.
Thus, I am pleased to note that President Trump has issued an executive order to lift the ban on supersonics!
The United States stands at the threshold of a bold new chapter in aerospace innovation. For more than 50 years, outdated and overly restrictive regulations have grounded the promise of supersonic flight over land, stifling American ingenuity, weakening our global competitiveness, and ceding leadership to foreign adversaries. Advances in aerospace engineering, materials science, and noise reduction now make supersonic flight not just possible, but safe, sustainable, and commercially viable. This order begins a historic national effort to reestablish the United States as the undisputed leader in high-speed aviation. By updating obsolete standards and embracing the technologies of today and tomorrow, we will empower our engineers, entrepreneurs, and visionaries to deliver the next generation of air travel, which will be faster, quieter, safer, and more efficient than ever before.
…The Administrator of the Federal Aviation Administration (FAA) shall take the necessary steps, including through rulemaking, to repeal the prohibition on overland supersonic flight in 14 CFR 91.817 within 180 days of the date of this order and establish an interim noise-based certification standard, making any modifications to 14 CFR 91.818 as necessary, as consistent with applicable law. The Administrator of the FAA shall also take immediate steps to repeal 14 CFR 91.819 and 91.821, which will remove additional regulatory barriers that hinder the advancement of supersonic aviation technology in the United States.
Congratulations to Eli Dourado who has been pushing this issue for more than a decade.