Data centers are, in a word, unpopular. They’ve quickly gone from the sorts of infrastructure projects that communities are neutral towards or welcome to being the sorts of projects that enjoy nearly universal revulsion. This is exaggeration, but not by much: the latest polling suggests about 75% of the population dislikes data centers.

Why? Data centers are clean, they pay high taxes, during their construction they pay a lot of blue collar workers and after construction they continue to stimulate the local economy, and they have very little impact on things like local utility costs or housing prices. But this is not what people believe. When you poll them, they seem to consistently believe that data centers are bad: they pollute and they increase costs.

But data centers do not greatly stress local resources or increase costs, so this is an odd set of views. If you’re very online, you might expect that the top reason for opposition to data centers would be something AI-related; that’s at least something real and worthy of concern, whether it’s an AI extinction risk fear or dismay about having one’s job replaced by machines. Alas, that is simply not it, and to suggest otherwise is to suggest that survey respondents are lying, and much worse, it’s a view that can’t even be falsified: if we believe survey respondents are actually worried about AI and they’re just fibbing or answering in confusion, then we can’t justify that view.

Because so many surveys have now been run, we can triangulate what portion of the resistance to data centers is driven by specifically AI-related sentiment. The triangulation evidence is as follows, from strong evidence to evidence that’s merely contextual:

• Explicit AI Wording. Generic wording produces considerable opposition, and adding explicit AI wording adds only a relatively small penalty. For example, Change Research in April 2026 found generic wording produced 51% opposition versus AI-specific wording, which produced 55% opposition.

• Matched AI-Purpose Comparisons Within Surveys. Across different survey instruments, AI-specific purposes consistently make data centers less popular, but not by a lot compared to the baseline levels of resistance. Echelon recorded AI purpose wording made support “less likely” by 17 points in February; Groundwork found it reduced favorability by 12 points in March; Rainey found that it added 6 points of opposition in May; Echelon found that it added 9 points of opposition in June. To contextualize those values, recall how much people already disliked data centers.

• Combined Study-Level Estimates Across the Polling Corpus. 41 direct-siting studies have been done between March 2025 and August 2026, and in the narrowest date-adjusted comparisons of exact local questions, predicted decided opposition is 65.1% generic versus 71.7% AI-specific. Across all distances, that increment averaged 7.7 points, so it’s fair to say AI adds 6-8 points over an already low level of support and high level of resistance.

• Temporal Divergence. In Annenberg’s February-July 2026 polls, local opposition to data centers increased by 12 points, but negative expectations about AI dropped by 3 points in the same period. Opposition rose across every subgroup they analyzed while negativity about AI moved in both directions, but generally down.

• Repeated Generic and AI-Specific Polling Over Time. All 13 comparable, repeated series became more negative on data centers by between 3 and 39 points, with the largest changes happening under generic wording, not AI-specific wording. Moreover, the opposition followed the same pattern we would expect if NIMBYism were to blame for opposition: it increased from 14 to 51% for a national generic question, from 26 to 65% for a local one, and from 40 to 73.5% for a near-home one. Were AI to blame, locality shouldn’t matter.

• Randomized Facility Comparators. Searchlight’s March polling found that opposition at five miles was 39% for warehouses, 54% for data centers, and 69% for workerless automated factories. Puck News’ June polling found more that data centers had 4% less opposition that nuclear power plants, but AI-specific data centers were just 5% more disliked than those plants. The gaps weren’t large and they’re not consistently tilted against AI-specific worries.

• Spontaneous Supplied Reasons. In Gallup’s Spring poll and Fox’s July poll, when opponents of data centers were provided with the opportunity to supply their reasons, AI concerns ranked much lower than environmental ones. In the Gallup poll, 14% cited negative AI views and 13% cited specific concerns; in Fox’s polling, 11% cited negative AI views. 50% of each poll’s respondents cited environmental concerns.

• Respondent AI Worldviews. Milltown’s polling suggested that negative AI views were common among opponents and not among data center supporters, and harm to the economy was associated with huge swings against data centers. People who thought they would be harmful were 28.1 points more likely to be opponents, and those who had career pessimism were 15.2 points more likely to oppose. Notice the contrast there: general fears versus personal ones, and swings so large that if these were common motivations, AI-specific data centers should have even less support and greater opposition.

• AI Opinions versus Local Opposition. In Fox’s July polling, 54% of participants had unfavorable views of AI versus the 70% who opposed data centers.

• Prompted Concern Batteries. Across different pollsters, environmental concerns lead, and AI concerns range from 5% as the single biggest to an outlier 52% in a multiple-response battery, with an average far below many other issues.

• Material Condition Moderation. Presenting participants with data centers that run off of clean or have their own dedicated power improves support, local electricity and water concern alleviation outperforms talking about winning the race with China, and cost and resource protections move support among 65-85% of people (note: this is not saying that it moves 65-85% to support, but that most people move in that direction).

• Age and AI Labor Patterns. AI job losses are expected to lead among ages 18-24, but opposition and environmental worry are nearly flat across ages. In some polls, like Fox’s recent one, the under-30 and 65+ differ substantially on sentiment towards AI, but they’re almost identical in how much they locally oppose building.

• The ‘Literally Who?’ Reality. When you poll people on which corporate leaders, companies, and products they recognize and how they feel about them, the results don’t comport with the idea that AI woes drive concern.
  
  99% or more of people have heard of AI, 99% or more have heard of Google, about 98% know of Microsoft, 95% know of Meta, 97% know of Amazon and Apple, etc. But move down from household names and awareness falls off: 82% for OpenAI, 64% for Nvidia, 62% for Anthropic, 51% for DeepSeek, 45% for Palantir, and 30% for DeepMind. The products don’t match the corporation awareness: ChatGPT sits at 92%, Gemini at 84%, Copilot at 81%, Grok at 64%, Claude at 63%, Perplexity at 53%, and specialized products are usually at 40% or less. Musk and Zuckerberg enjoyed high recognizability (98% and 95%, respectively), but beyond that, major AI company CEOs are not widely known: 57% for Sam Altman, 52% for Peter Thiel, 49% for Larry Ellison, 35% for Jensen Huang, and 26% for both Alex Karp and Dario Amodei. Then, the ratings don’t even match! Karp is viewed positively among those who know him, and Dario is, too!1
  
  But then, problematically for the narrative that AI is driving concern, it’s polling better over time, more people are using it over time, and favorability is higher among users than nonusers, with signs of some increase over time, unrelated to history of prior use. Moreover, people simply don’t even seem to believe that massive AI job loss or extinction concerns are realistic possibilities!

There are several lines of evidence that point to the most realistic conclusion: data center concerns are about the things people say they’re about. People are actually worried about things that are not realistic concerns, like electricity prices and water availability. Opposition is not due to worries about AI extinction risk, AI job loss, AI threatening people’s happiness and family time, or anything else like that. AI only adds a small bit of extra ‘oomph’ to the opposition, and the reason is ambiguous. For example, when asked about data centers versus AI-specific data centers, people could oppose it more because they think AI data centers have worse environmental impacts!

There are a lot of things we need to consider in order to get the public to support building more data centers. To do that, I suggest we stop drawing up and relying on narratives that explain why people in our circles might be against data centers, and start, instead, listening to the people who are actually on the ground and in the polls telling us—explicitly—why they oppose them. All signs suggest they’re not lying; they point instead to the much simpler fact that the opposition is mostly just NIMBYs.

Here is the list of surveys I reviewed and/or referenced, listed in field dated order. 2023: Washington Post/George Mason Schar School, 2023; CNBC/Generation Lab youth-and-money study, October 2023.

2024: Axios Harris Poll 100, 2024; CNBC/Generation Lab youth poll, January 2024; Morning Consult AI Super Users, March 2024–March 2025; CNBC/Generation Lab youth poll, April–May 2024; Data for Progress/Accountable Tech, July 2024.

2025: Anne Arundel Community Survey, 2025; HostingAdvice, 2025; Wason Center Virginia survey, January 2025; Public Policy Polling/Tech Oversight Project, January 2025; Change Research, March 2025; Airedale/Modine neighborhood survey, March 2025; Illinois data-center tax-incentive survey, March 2025; Overton Insights, June 2025; North Carolina LCV Foundation, July 2025; Data for Progress/WE ACT, July 2025; Ohio data-center tax-exemption survey, July 2025; South Carolina Policy Council, July 2025; Embold Research/Heatmap, August 2025; Arizona data-center survey, August–September 2025; Minnesota data-center accountability survey, August–September 2025; AARP national 50+ utility-cost survey, September 2025; AP-NORC/EPIC Climate Change survey, September 2025; Mississippi Omnibus Survey, September 2025; Georgia data-center tax-and-bills survey, September 2025; Fairleigh Dickinson University New Jersey survey, October 2025; Marquette Law Wisconsin Poll, October 2025; Data for Progress Alaska, October 2025; Morning Consult, October 2025; Texas data centers and electric-grid survey, October 2025; Harvard Salata Institute/YouGov, November 2025; Hoosier data-center and energy survey, November 2025; Ipsos/Redfin, November 2025; Elon/North Carolina data-center electricity-cost survey, November 2025; Emerson Pennsylvania data-centers-and-AI survey, November 2025; DHM Washington energy survey, November 2025; Morning Consult, November 2025; Caltech/YouGov, December 2025; Navigator Research/Global Strategy Group, December 2025; CCAN Action Fund/Global Strategy Group, December 2025; Change Research, December 2025; Harris Poll data-center survey, December 2025; Blue Rose Research, December 2025; Emerson national poll, December 2025; YouGov daily survey, December 2025; Louisville Metro Office of Planning survey, 2025.

2026: Value of Water Index, 2026; ACSI AI Platforms Study, 2026; Cherokee Nation citizen survey, 2026; Reach3 Insights, 2026; YouGov AI Index/BrandIndex, 2026; Wason Center Virginia survey, January 2026; Georgia pre-legislative survey, January 2026; Morning Consult/Stateline, January 2026; Public First/POLITICO, January 2026; Pew Research Center, January 2026; South Carolina Policy Council, January–February 2026; Climate Power/Blue Rose Research, January–February 2026; Amber Integrated/Oklahoma Watch, February 2026; Embold Research/Heatmap, February 2026; Glengariff Michigan survey, February 2026; SLU/YouGov Missouri technology survey, February 2026; Ohio voter energy survey, February 2026; Marquette Law School Poll, February 2026; Axios Harris Poll 100, February–March 2026; Nexus Pulse, February 2026; Annenberg Public Policy Center/SSRS, February–March 2026; Echelon Insights, February 2026; Quinnipiac University, February 2026; Rainey Center, February 2026; Clean Virginia, March 2026; Gallup, March 2026; South Dakota issues poll, March 2026; Elon University North Carolina poll, March 2026; Marquette Law Wisconsin Poll, March 2026; Illinois data-center regulation survey, March 2026; Searchlight near-home facility comparison, March 2026; Pew energy-bill survey, March 2026; Quinnipiac University, March 2026; Fairleigh Dickinson University, March 2026; Washington Post/George Mason Schar School, March 2026; AARP Oklahoma 50+ utility survey, April 2026; Montana Conservation Voters, April 2026; University of Houston SPACE City Panel, April 2026; Center for American Progress utility-price message testing, April 2026; Zencity/TransformGov, April 2026; Change Research, April 2026; South Dakota News Watch/Chiesman Center, April 2026; Climate Power/Global Strategy Group, April 2026; Searchlight energy-price survey, April–May 2026; Glengariff/Detroit Regional Chamber, April–May 2026; Alabama Republican-primary survey, May 2026; Conservation Colorado, May 2026; Hahn/Escalent, May 2026; Ipsos/Redfin, May 2026; Morning Consult consumer-AI reputation study, May 2026; Reuters/Ipsos, May 2026; Ipsos Consumer Tracker, May 2026; Embold/Compass on-site energy experiment, May 2026; Milltown Partners, May 2026; Change Research Florida, May 2026; UMass Lowell Maine survey, May 2026; Embold Research/Heatmap, May 2026; Rainey Center experiments, May 2026; Utah Stratos survey, May 2026; FM3/Net Zero California, May 2026; UNH New Hampshire AI survey, May 2026; Justice Research Group/Working Families Power, May 2026; The Argument/Embold company-favorability survey, May–June 2026; Environmental Polling Consortium and partners, June 2026; University of Virginia Data Center Opinion Project, June 2026; Utah SD-21 Republican-primary survey, June 2026; YouGov Millennial crossbreaks, Q2 2026; YouGov fame and popularity trackers, through Q2 2026; CBS News/YouGov, June 2026; Reuters/Ipsos AI survey, June 2026; University of Texas/Texas Politics Project, June 2026; AI Policy Institute, June 2026; Echelon Insights, June 2026; Data for Progress moratorium survey, June 2026; Fox News/Beacon Research/Shaw & Company, June 2026; Annenberg Public Policy Center/SSRS, June–July 2026; VCU Commonwealth Poll, June–July 2026; PPIC/Ipsos California survey, June–July 2026; ARI/Morning Consult, July 2026; Hancock County open survey, July 2026; Public First/POLITICO, July 2026; Harvard Harris Poll, July 2026; Empire Center New York energy poll, July 2026; Quinnipiac University, July 2026; Marquette Law Wisconsin Poll, July 2026; Fox News/Beacon Research/Shaw & Company, July 2026; Emerson national poll, July 2026; DHM Oregon Pulse, July 2026; Elon University North Carolina poll, July 2026; Wyoming public-lands poll, July 2026; Data for Progress Alaska, July–August 2026; Napolitan News, July 2026; CNBC/Generation Lab AI-leader trust poll, August 2026; Embold Research/Heatmap, August 2026.

This was a timed post. The way these work is that if it takes me more than an hour to complete the post, an applet that I made deletes everything I’ve written so far and I abandon the post. You can find my previous timed post here.

A message from my sponsor, Mechanize:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

Depending on the poll, the two biggest issues the public has with data centers are:

1. They take all the water;

2. They increase electricity prices.

The former complaint is scientifically illiterate, dead and done, and gets its limited intellectual backing from analyses that were shown to have included mathematical errors that inflated water use estimates and others that have spuriously attributed water use elsewhere to data centers. The latter complaint has been found to be not just false but backwards for the period before today’s hyperscale buildout, so proponents of the idea have shifted their defense to stating ‘That might’ve been true, but now the buildouts are radically increasing electricity prices!’

I’ve decided to take this possibility seriously. To that end, I’ve made a continuously-updated website that assembles that latest available data on data center energizations and electricity prices. You can explore the website—wattsmybill.org—by clicking here.

What did I find? Firstly, there is an effect of new data centers on electricity prices. However, the effect is small, and it appears to plateau. Secondly, the effect of new data centers on electricity prices is driven most sharply by large, recent buildouts. Finally, projections based on the existing data suggest there’s plenty of room to build out data center capacity without major effects on electricity bills.1

Here’s how average price effects from a “hyperscale”—250MW—data center being built locally look for a typical household as of August, 2026; in short, they’re small:

If you want to play around with the data, check the expected effects of data centers on your bill, or just explore all the analyses I ran, go check out the site by clicking here! If you have requests for additions to the website, then go and leave a comment below.2

This was a timed post. The way these work is that if it takes me more than an hour to complete the post, an applet that I made deletes everything I’ve written so far and I abandon the post. You can find my previous timed post here.

A message from my sponsor, Mechanize:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

Data centers are very unpopular right now, for all sorts of reasons. The latest polling is abysmal: people want a data center in their community less than they want a new nuclear power plant, and they really don’t want to live near a new nuclear power plant.

The reasons people supplied for not wanting a data center will be familiar to you if you’ve seen other polling. The most slanted beliefs were that data centers strain the electrical grid and that they consume extreme amounts of water, pushing up the utility bills of local residents. People even tend to believe things that are extraordinarily obviously untrue, like that data centers tend to lead to increased taxes and lower property values. Apparently they don’t see the contradiction, but I digress.

The unique thing in this poll was that the pollsters sought to see if they could change people’s views. They wanted to test if they could unravel people’s misinformed views about data center harms by telling them the truth. Unfortunately, people don’t tend to believe the truth, even when you present it to them directly: they don’t think data centers generate tax revenues, that they make AI more widely available, that they create high-paying jobs, power modern business applications, or—critically—that they actually don’t use very much water.

Given how much disdain people hold at baseline, how strongly they hold their most errant beliefs, and how unconvinced they are by statements of fact, you might expect supplying facts to have little effect. You’d be right! People’s opinions moved in the direction of supporting data centers, but by a barely significant amount.1

How Can We Usher in the Future?

There can be little room for doubt that AI is the future of the global economy. To develop and power that AI, we need to increase the available amount of computational power, which means more data centers. But the public opposition to data centers is acting as a very effective Luddite revolt against realizing the benefits of this future by bringing the promise of AI to fruition. How exactly do we stop it?

My first thought was to look at the results of other persuasion experiments. When researchers have assessed the impacts of political campaign messaging on persuasion, they’ve generally come up empty-handed. Experiments on persuading people to be less hateful through interpersonal conversation sometimes turn up results, but in the best analyses, the effects are meager. The effects of online political advertising on the propensity to vote for a given party have also been estimated to be small and they fade fast, much like TV ads. Common shortcuts like deploying people with shared demographic characteristics to do persuading also don’t seem very valuable.

Persuasion experiments in the published literature turned out to be mostly irrelevant, so I was disappointed by them. But, I did find some NIMBY messaging experiments.

Monkkonen and Manville asked 1,300 LA County adults about hypothetical apartment buildings proposed on their block. They presented the participants with arguments against buildings the apartments, and all such arguments helped to convince people to be against building, often by a very considerable degree. For example, telling people that the developer serves to make large profits increased opposition by 20 whole percentage points. To get people to support building required basically bribing them: most participants (63%) increased their support for projects if developers offered community benefits.2

Walker, Wiersma and Bailey ran a postal survey in England, where they asked people about an offshore wind farm near them. Some participants were asked about it without any framing, others were told about community benefits, and the rest were told about the wind farm’s community benefits and common counterarguments. Without the counterarguments, people became more supportive of the wind farm when they were presented as a boon to the area. The whole support boost evaporates the moment wind farm project benefits are labeled ‘bribery’.

Liebe, Bartczak and Meyerhoff asked Germans and Poles about wind farms near them and found that, firstly, Poles are more accepting or new turbines than are Germans, and secondly, people more readily accept wind farms when they’re owned by local citizens, when the electricity is consumed regionally instead of being exported, and when citizens perceive that their community is involved in the planning process. A lucky finding was that those who had already been exposed to wind turbines were more accepting of additional wind turbines, but it’s not exactly clear why this is. It could be that exposure indexes prior support rather than an exposure effect per se.

Vuichard, Stauch and Dällenbach asked Swiss people about what would increase support for local wind farms, focusing on financial inducements. They found that the only thing that worked to raise acceptance was a local resource tax that benefitted everyone in the community. This was politically split. Conservatives just weren’t convinced at all, and leftists had generally high support for the wind farm and only had their support boosted with the wind resource tax idea. However, centrists actually saw notably increased support with every inducement, they just saw the most with the resource tax idea.

Knauf ran a conjoint experiment in Germany where they found something similar: financial benefits for wind energy projects increased support, but only among those who already supported them or who didn’t have any strong views. Among those who were opposed, benefits failed to move them.

García et al. found that, in Norway, the amount of required compensation to increase local support for wind farms was contingent on how the benefits were delivered. Private payments needed to be much larger than compensation delivered in the form of local services or infrastructure improvement, in order to generate the same support increases.

Muñoz and Tormos ran a survey in Catalonia where they found that framing a wind or solar project as being done “somewhere in Catalonia” led to more support compared to when it was framed as being in one’s own municipality. Opposition to projects also increased with scale, but it decreased when benefits to the environment or municipal tax reductions were offered.

Doberstein, Hickey and Li found that Canadians in a fast-growing, car-dependent mid-sized city were strongly opposed to new developments on their street, but were less opposed to developments elsewhere in town. They softened their stance and became more supportive when the message noted the public benefits of residential densification, but they were unconvinced by private benefits or by the presentation of expert views on benefits. They were also able to have their views shifted by learning what their neighbors felt about the development.

To summarize the literature on NIMBY/YIMBY messaging experiments:

• People care about the community benefits of new building.

  • They want public, widely-shared benefits.

  • They don’t want private benefits per se.

• People care about feeling involved in the process of development.

• Some people cannot be convinced in the short term by any reasonable message.

• When it seems like proponents of building are personally benefitting, opposition increases dramatically.

Opposition to building is driven by a mixture of disdain for unsightly builds, demands to preserve community and environmental character, pure resent, and so on. These are generally not rationally motivated concerns, which helps to explain why fact-based messaging doesn’t work that well: you can fully convince someone that they were wrong about some particular empirical claim about building’s harms or lack of benefits, but they still won’t budge. I would avoid persuasion techniques generally.

It is likely that enough of the fact-based pushing can win out in the longer term, as always happens with technologies. People resisted the internet, the automobile, the telephone, and even electrification writ large, but familiarity, aging, and progress eventually wear down resistance and absurd myths tend to die—who besides crazies now believes that electrical wires make the communities they’re in sick? But this is a poor assurance to people who want to bring that progress about sooner.

For near-term progress, the strategy is clear: give people what they think want, first of all, even if you don’t give them what they’d get through a real negotiation process. Run roughshod over them with any permitting workarounds possible while giving the appearance that locals have been listened to, in messaging. Ensure they know building is costless for them: pay for electricity, pay for water. Then talk about the implicit and actual bribery that comes in the form of construction jobs, permanent white-collar jobs, and property tax abatements, more funding for schools and parks, whatever.

At every stage in history, progress has been resisted, and in some places and times it’s been halted entirely, resulting eventually in a humiliating, unnecessary national defeats. We now face the same challenge and rising to it is our choice to make.

This was a timed post. The way these work is that if it takes me more than an hour to complete the post, an applet that I made deletes everything I’ve written so far and I abandon the post. You can find my previous timed post here.

A message from my sponsor, Mechanize:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

I recently read an article that claimed climate change research has peaked, as measured by the number of papers being written on the subject since 2020 rapidly declining. The paper’s recorded decline is dramatic: the number of peer-reviewed publications labeled “climate change” went from a peak of about 900,000 in 2020 to a near-term nadir below 200,000 in 2026.

Science Is Not The Answer

We Passed Peak “Climate Change” (um) Research

Here is a graph of the yearly number of total papers that have the keyword “climate change” in them…

Read more

12 days ago · 86 likes · 32 comments · William M Briggs

The article explains this dramatic change in terms of “the teats [being] squeezed dry”, the field of climate change research being “on repeat”, totally hollowed out of new territory to look into, data to model, and hypotheses to propose. I don’t believe this explanation, but I could believe some explanation like ‘People have realized that we’re going to beat climate change, so the energy is going elsewhere’ and ‘The decline could have accelerated recently due to papers reclassifying themselves away from being about climate change in order to avoid the ire of the Trump administration.’

However, before we try to explain phenomena, we must determine if there’s anything to explain. As I’ve noted before, abrupt changes are usually fake and we have a duty to determine if what we’re critiquing or supporting is valid to the greatest extent we can.

So, how did I check if this article was correct? Three ways. First and weakest, I looked at the article’s own source: Google Scholar. I checked if there was a generalized service-wide decline that mirrored what the blogpost noted for climate change-related research. As it turns out, there was! Fields with more hits in 2020 had considerably fewer hits in 2025. This is odd, but it’s not explicable by something specific to climate change research changing.1

Second, I looked into another repository, OpenAlex. I tried to replicate the peak the blogpost observed, but, well, take a look:

Third, I replicated this again. After all, we don’t want a game of warring repositories; we don’t want to just play Google Scholar off against OpenAlex, so I also looked at Semantic Scholar, Crossref, OpenAIRE, PubMed, Europe PMC, and DOAJ.2

The question of whether climate change research paper numbers have declined in the COVID era is an empirical one that’s worthy of looking into, and the answer is ‘no’. Accordingly, we have no reason to speculate about why it’s declined, as the truth is that it hasn’t. But we do have reason to repeat once again that we must verify things need to be explained before we explain them; we must avoid that harmful impulse.

This was a timed post. The way these work is that if it takes me more than an hour to complete the post, an applet that I made deletes everything I’ve written so far and I abandon the post. You can find my previous timed post here.

A message from my sponsor, Mechanize:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

# Root fields (OpenAlex level-0 concepts)
- Computer science
- Physics
- Medicine
- Biology
- Engineering
- Materials science
- Political science
- Chemistry
- Philosophy
- Geography
- Mathematics
- Psychology
- Art
- Business
- Sociology
- Geology
- Environmental science
- History
- Economics

# Top subfields (level-1 concepts, ranks 1-10 by works count)
- Humanities
- Artificial intelligence
- Optics
- Internal medicine
- Law
- Biochemistry
- Quantum mechanics
- Remote sensing
- Organic chemistry
- Genetics

# Sampled subfields (level-1, every 5th rank from 15)
- Statistics
- Surgery
- Geometry
- Social psychology
- Information retrieval
- Metallurgy
- Endocrinology
- Neuroscience
- Theology
- Computer vision

# Low end (level-4 concepts by rank)
- Nucleofection
- Gestational period
- Diafiltration
- Emperipolesis
- Sclerodactyly
- Ocrelizumab
- Divertor
- Neutral beam injection
- Magnetic energy
- Langmuir probe
- Magnetohydrodynamic drive
- Cyclotron radiation
- Neutron emission
- Ballooning
- Transcriptome
- Superconducting Super Collider
- Cyclotron resonance
- Genomics
- Single-nucleotide polymorphism
- Gene expression profiling
- Repressor
- DNA methylation
- Acceleration voltage
- Magnetic particle inspection
- Clade

# Low end (level-5 concepts by rank)
- TSG101
- Articular cartilage damage
- Proteogenomics
- Protein isoform
- Durvalumab
- Pandemic
- Magnetic amplifier
- Edge-localized mode
- Central core disease
- Genetic association
- Coronavirus
- Genome-wide association study
- SNP
- DNA Transposable Elements
- KEGG
- Good agricultural practice
- Social distance
- Prompt neutron
- RNA-Seq
- Conservation-dependent species
- Neutron probe

# Original battery
- climate change
- the
- cancer
- protein folding
- machine learning
- quantum computing
- gene expression
- randomized controlled trial
- solar cell
- coral reef
- stochastic volatility
- nonlinear Schrodinger

The obvious answer to the headline question is that you would ask it a question in the same way you would ask a person, right? For some queries, I think that’s acceptable—(‘How big is the sun?’, ‘What’s 10+10?’, ‘Find a well-known mathematical conjecture and solve it. Make no mistakes.’)—, but often enough, it’s just not. In my experience, the key issues are obeisance, confidence, ‘nuance’, and tacit knowledge.

I’m going to consider three recent cases that I believe to be illustrative of these issues.

America’s High-Cost Healthcare

In 2014, (henceforth, “RCA”) produced an analysis that decisively debunked a number of popular talking points about American spending, healthcare, and so on, and he popularized Actual Individual Consumption and its importance to the blogosphere and the Twitterati. This blogpost has produced an inordinate amount of ire because left-leaning (= center-left and further) and far-right people interested in politics often hold strong, non-empirical views and corresponding positions on the topic of American healthcare.1

The main thing that RCA documented is that America is not a massive outlier in global healthcare spending once its vast wealth is accounted for. In other blogposts, he expanded on several potentially important points, debunking common ideas about the roles of inequality, prices (see also), administration2, and much more.

What do LLMs like Claude and Grok say when you ask them about RCA’s work and conclusions? If you ask naïvely, you will find that they deliver completely wrong answers. Not only that, but often enough, they’ll explicitly reference RCA and claim to contradict him, while citing evidence that he already dealt with. As an example, using a question from , Claude (on Max effort) appears to do just that.

When asked to evaluate arguments, Claude often begins by conceding that one side makes some points, then going and arguing against their case along some other lines, like faux-steelmanning. That’s typical of modern LLMs and that’s how this started off.

In its concession to RCA, Claude got numbers wrong, understating RCA’s case. For example, it said “AIC per capita runs something like 30-50% above the big Western European economies.” Is this the case? No. AIC was >50% higher than the EU28 economies, >50% higher than the 11 core EU economies, and 55-60% above the “big” economies (assuming these refer to Germany, France, the U.K., Italy, and Spain).

Claude then started to talk about how RCA’s argument was “fragile”, alleging that it was ‘circular’, meaning that high health spending is included in AIC, so its denominator is inflated. Is this true? Yes, but RCA dealt with this obvious objection and showed it was not explanatory. Therefore, Claude should not have brought it up; it was ‘prebunked’. The next point of fragility wasn’t even right: Claude said being “on trend” was sensitive to functional form—not true among valid forms—and the choice of small countries to include—part of the whole point of using AIC, to fix the issues with said small countries’ GDP numbers being off due to tax haven status. The next point of fragility is also not even a point, it’s a restatement of what RCA argued: that the trend line is evidence of common causes of health spending across countries. Claude, however, took issue with RCA’s arguments and said “Sitting on the trend line isn’t evidence of efficiency—it’s compatible with Baumol effects plus waste that scales with income everywhere.” How is restating the thesis a counterargument?

In the next paragraph, Claude made exaggerated claims about the role of administration and billing overhead—things RCA dealt with—, price dispersion across countries, branded drug costs—an indication that Claude knew that drug cost concerns were substantially overstated due to generic drug availability—, and then it cited an incorrect study that RCA already explicitly dealt with. The study—dealt with here—argued that utilization differences don’t explain America’s high health spending, but it did so by using select utilization measures that don’t relate much to health spending or have high U.S. residuals, which are things RCA, once again, explicitly explained. Yet, Claude cited this work when it was already shown to be in error, and then it went on to talk about another thing RCA already explained, before switching topics to talking about lifespans instead of health spending, seemingly just to say RCA was correct on that topic in the middle of a paragraph on spending—something I can only read as ‘attempting to sound like the model is being fair’.

Claude then ends its discussion by making a data-free inference that the U.S. is not normal for its wealth and bringing up three different things America’s alleged “dysfunction” is concentrated in, one of which wasn’t relevant, one of which made no sense, and another of which has some truth to it, but was already dealt with by RCA! Had Claude been seeking a proper answer, it would’ve gathered multiple years of the relevant, publicly available data and tested its proposition, or at the very least, said it was unsure.

Other LLMs reason similarly incoherently and also dodge the facts, the things RCA actually said, and they repeat what’s been published in papers, even when it’s irrelevant or wrong. If you read how the LLMs reason in response to MattY’s question, you’ll notice that they say a lot of things that seem to be relevant, but which really aren’t. As such, they appear to provide a lot of information, but they really just use a lot of words to say very little at all. They do not inform, they misinform.

But notice how the responses are written. We see LLMs that write in a way that, from a human, would be considered confident, even though you can press it on any number of these points, and it will concede. We see LLMs that stick to scholarly consensus, preferring debunked points published in JAMA to the debunkings published on a blog. We see LLMs that repeatedly says RCA makes some points, but that he’s overall wrong when everything is considered.3

Woke Madness

A few days ago, I saw posting his blogpost Does Therapy Culture Explain the Ideological Mental Health Gap? over on Twitter. I noted that his preferred theory was not viable since the measurement of this gap comes with evidence for measurement invariance when comparing liberals and conservatives.

He then asked me why the finding of measurement invariance rules it out, I explained, and he took to asking ChatGPT what it thought. The issue with using ChatGPT as a spot-check here is that it doesn’t have an understanding of the topic. In fact, in its response, it affirmed I was correct, but didn’t realize it!

Think about what the LLM should’ve known to say. There are many options consistent with how measurement invariance is defined and what it means—which, I’ll note, is specialist knowledge—, but my preferred response from an LLM would’ve involved interrogating the premise: how could therapy culture explain a liberal-conservative gap and be consistent with identical measurement parameters for both liberals and conservatives?

The implication of said identical parameters—that is, the state of measurement invariance—is that the causes of differences in the thing being measured must act the same between groups. But imagine that: how could they?

For one, presumably those who go to therapy would conceive of the questions used to measure mental illness in different ways, ways atypical of society-at-large but normal in the world of therapy. For two, is it plausible that this works for a trait like depression given what it’s measured by? Consider the most commonly used depression scale, the PHQ-9. Here’s a model of it:

Now, imagine where the variable “Therapy culture” would slot into that diagram. To retain measurement invariance, it would have to affect the PHQ9 latent variable or at least operate on all of the measured variables identically to how the PHQ9 latent variable appears to. Since depression is almost-certainly not a ‘real’ latent variable, but rather something more like a network, we can safely constrain our theories to be about the latter type of variable: a factor mimic.

Candidate explanations for filling this role are basically unimaginable: how would therapy culture cause a connection between anhedonia and sleep troubles, and also to every other symptom as it needs to? How could it act across a trait continuum where the symptoms can go either way, such as with too much or too little sleeping, or too much or too little eating? Therapy culture sounds quite magical! Add in that it must explain more conditions than depression—and those sometimes come with pathognomonic signs—and “therapy culture” as explanation really stretches the limits of what’s believable. It may have some role somewhere, but it’s no general explanation.

I’ll leave the theoretical speculation about how therapy culture could be consistent with measurement invariance to the people who believe in its role, since I don’t, and it’s their case to make. I’ll instead simply point out what the LLM does wrong here. For one, it doesn’t understand my objection. This is visible when the LLM mistakenly affirms what I’ve said while claiming to contradict it. It’s also visible when the LLM outlines at a high level what conditions have to hold for measurement invariance to be attainable, without elaborating on the complex and, I think, practically impossible theoretical constraints that would have to hold for the theory to make any sense.

If you prompt the LLM to think more deeply about how this theory could work and you keep poking holes, it eventually concedes the point and says it’s unlikely that therapy culture plays the all-important role it’s been ascribed. But it lacks the tacit knowledge, or the instinct to automatically ply tacit knowledge, to make these issues apparent from the outset at the asking of a seemingly simpler question. But this question is not really simple, and it takes asking a lot of questions to grasp that if you’re not already familiar with the ins-and-outs of psychometric methods. This is true of many questions, so if you ask any of them, for a correct answer, you have to be in a position where you likely already know the answer, making the LLM superfluous.

David Sinclair’s Sudden Blindness Epidemic

David Sinclair recently went on the Joe Rogan Experience podcast and lied egregiously about the side effects of GLP-1 receptor agonist drugs like Ozempic and tirzepatide. The lie that I called out was that “NAION… is now twice as more [sic] prevalent because of the GLP-1 drugs.”

Because David mentioned “sudden blindness” writ large in this section of his interview, I looked at all types of sudden/acute vision loss and found that their incidence hadn’t changed between 2016 and 2023 in national-level electronic health record data. We’re not well-powered for NAION specifically because it’s so rare, but the prevalence of that hadn’t even moved in the right direction. Nevertheless, I got a lot of comments like this one: an LLM was asked who was right, and the LLM suggested Sinclair might’ve been right.

The basis of the LLM’s confusion has to do with a simple mix-up. This is their logic:

1. There are cohort studies that suggest GLP-1RA users have a higher rate of NAION;

2. More people are using GLP-1RAs than used to;

3. Therefore, rates of NAION have increased.

Can you tell where the issue lies? Fundamentally, nothing in this chain of logic demonstrates that the rate of NAION has increased. To demonstrate that NAION has increased, this method just won’t do; instead, you’ll have to count the number of NAION cases and show that they increased.

But, you may ask, doesn’t that logic follow? And the answer is ‘no’, because it only follows under the restrictive condition that the identified effect size is causal, correct, and applied to an appropriate population that we can generalize from. If any of these is off, then you could find arbitrarily large effect sizes without any effect even existing.

I showed that the studies Sinclair was referring to were wrong in three ways:

1. The effect sizes have declined in newer, larger, better-controlled studies. The magnitudes he’s citing are thus unlikely to be correct;

2. The populations analyzed in his preferred studies were not representative enough. In fact, better representativeness has come with smaller and smaller effect sizes;

3. I looked at the measured population rate of NAION and a variety of other conditions and saw no increase. In fact, for NAION—which has such a low rate that changes are hard to determine—the direction of the effect wasn’t even right.

If you just ask the LLMs about Sinclair’s statements, they don’t think to bring up any of this. They may say that the base rate of NAION is low and thus the benefits of taking a GLP-1RA outweigh the risks for almost everyone, but they will not generally walk down the chain of logic and go ‘Hold up, something isn’t right. Where’s the doubling evidence?’ Instead, they make the same error as Sinclair and unfortunately many other scholars of similar ilk.4

Can you see the issues here? LLM answers sound right—the machine provides confident prose—, but to get the correct answer, you would need to know to press it in ways that are not obvious to most people. If you do not, then LLMs will state mainstream views and interpretations, default to what’s commonly understood on sites like Reddit (where much of the training data is derived) and in mainstream news sources like the Wall Street Journal or New York Times. They will incorrectly claim that open-and-shut questions require considering multiple angles to understand, and yet they won’t know when to probe deeper on questions because they lack the tacit knowledge to default to theoretical thinking about the things they’re asked about.

To make matters worse, LLMs will provide inconsistent answers depending on who they believe to be asking a question. I’ve seen this first-hand when, in my own sessions, I get something passable, and a friend asking a question similarly gets rubbish (and vice-versa!). In fact, this behavior has been documented systematically:

To answer the question ‘How do you ask an LLM a question?’, well, I’m afraid I don’t quite have an answer. I don’t think anyone does, as it’s too open-ended. For the question ‘Can you find a mathematical conjecture and disprove it?’ there’s likely enough detail there for it to go off on its own and return something substantive. For the question ‘What is the meaning of life?’ you’ll doubtless be disappointed no matter how many follow-up questions you manage to plow into the text box. Good luck.

This was a timed post. The way these work is that if it takes me more than an hour to complete the post, an applet that I made deletes everything I’ve written so far and I abandon the post. You can find my previous timed post here.

A message from my sponsor, Mechanize:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

Do you see the chart Trump is gesturing at on his Joe Rogan Show appearance?

It is perhaps the most misleading chart in modern health discussions: it’s widely circulated and it makes people think that America lags in life expectancy because it has a bad healthcare system. Here’s a better view of the chart:

The reasons Americans live shorter lives despite being much richer than their peers are well-understood at this point. The biggest contributor is that obesity is very bad for health, as countries get richer their citizens tend to get fatter, and America is the richest country. We can break this down explicitly by using methods developed here.1

Notice how the American life expectancy gap varies by the choice of comparison country not just in terms of the size of the gap, but also in terms of the causes and magnitudes of death that it can be attributed to. Notice too that the causes of the gap vary considerably by sex. I think we expect this. For example, consider automobile accident and homicide mortality. Those are attributable largely to extreme males, whereas women are generally less extreme in their behavior, and this is true globally, so we should see smaller gaps attributable to those causes. We also see that there’s not a difference in the cleanly health care-attributable share2 for men and women, which is odd if you put a lot of onus on that, given that women have more interactions with the health care system and they’re more reliable in their interactions with them (e.g., they take their medicines more consistently, make it to their appointments, etc.).

Take a look at the broader causes of the life expectancy gap:

One thing that’s interesting about this is that it will change over time. Two very critical ways that these things will change are that some things will shift towards being amenable to intervention by the health care system (e.g., we will one day be able to clear arterial plaques, reversing atherosclerosis, but we can’t yet), and some of these gaps will be reduced or reversed by technological innovation. Consider the recent advents of self-driving cars, GLP-1RAs (which you can acquire cheaply at this link), and gene therapies to permanently reduce cholesterol levels. What if we fixed just the gaps in automobile accidents, obesity, and heart disease? The impact would be huge:

To make the point that America’s gap is not about its health care system clearer, consider one of the most key—and controversial—explanatory variables that separates America from these peer countries: race. Take a look at those gaps. They’re really something:

All Americans regardless of race are, in principle, exposed to a health care system of the same quality understood broadly3, although there are gaps in access4, but those access gaps don’t fit the order of the longevity gaps—as a close proxy, in 2024, the insurance rates go Asian > White > Black > Hispanic. These race results and the sex ones combine to produce a picture that cannot be explained by a simple health care-centric story. In fact, when it comes to things the health care system can plausibly intervene on, America is usually the highest-performing country—at least cross-sectionally—or it’s very nearly so. When it comes to lifespan and America’s lag despite its riches, look elsewhere than health care.

This was a timed post. The way these work is that if it takes me more than an hour to complete the post, an applet that I made deletes everything I’ve written so far and I abandon the post. You can find my previous timed post here.

A message from my sponsor, Mechanize:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

1

The data used is the latest years of available data, matched to America in that year. Some of this is substantially impacted by COVID, unfortunately. Moreover, cause-of-death attribution like this is an accounting exercise as I’ve done it here, not dynamically modeled. So, for example, if you reduced heart disease deaths, a larger portion of the population would likely instead die from cancer (similar ideas hold for group comparisons). Keep the observational nature of the attributions here in mind. Also keep in mind incomparability in coding across countries, like that which explains what’s going on with Japan’s data.

For a combined instead of sex-split version, here you go:

2

Drug overdose, alcohol, lung cancer, diabetes, respiratory illness, and circulatory problems were grouped into Lifestyle & Behavior. Firearm-related, homicide, motor vehicle accidents, and suicide were grouped into Violence & Injury. Screenable/Treatable Cancers, Infectious Disease, and Perinatal/Congenital complications were grouped into Health Care-Amenable. Alzheimer’s Disease, Mental/Nervous, and Residual were grouped into Other & Residual.

Notice that in some cases—such as Perinatal/Congenital in Health Care-Amenable—the causes are not truly Health Care-Amenable in enough cases, that the allocation almost-certainly over-attributes American’s life expectancy gap to its health care system. This is because large portions of congenital issues are not remediable and large portions of all fatal perinatal issues have to do with things like maternal obesity. Classification also plays a role. (See this and this, too.)

Here’s an alternative view:

I often see people suggesting that there are countries that are richer than the U.S. This is true if we include wealthy micronations like Monaco or Liechtenstein; it is not true if we look at ‘real’ countries. This fact is sometimes obscured, however, by problematically inflated GDP figures for ‘real’ countries that are tax havens—such as Singapore, Switzerland, and Ireland—, and countries with extraordinary levels of inequality—like Qatar or the United Arab Emirates. But we can overcome these issues.

First, we have to figure out what to measure. The answer is not GDP per capita. Let’s ascend the ladder of comparisons!

1. GDP Per Capita: This is our familiar starting point.

2. GDP Per Capita, PPP: This is GDP per capita, but it also adjusts for price levels between countries. The adjustment is required because a dollar buys a different amount in different places. For America, rungs 1 and 2 are the same.

3. Net National Income Per Capita: This is like GDP Per Capita PPP, but it adjusts to account for the reality that capital wears out and some income belongs to foreigners.

4. Final Consumption Per Capita: This is actually what people consume, which is what their quality of life is based on in the immediate term. This penalizes thrifting.

5. Actual Individual Consumption:1 This is like final consumption per capita, except it accounts for the fact that some consumption is done collectively, and it uses consumption prices to promote comparability, too. This also counts out the things that we cannot attribute identifiably to people, like spending on defense and police, and it makes stuff like a publicly- and privately-funded hip replacement both get counted similarly, as the same sort of consumption.

And here’s where things get debatable:

• Median AIC: A mean is not a typical person; rather, it’s distorted by outliers.

• Low-Percentile AIC: A median is not a poor person, whose lives we’re interested in.

• High-Percentile AIC: A median is not a rich person, whose lives we’re interested in.

And there are also some things we definitely can’t do in international comparisons, which I’ll return to below. For now, let’s look at the rankings: how rich is America relative to other countries? The answer is: quite! It doesn’t have the richest poor people, but it does have the richest people at the 10th percentile of the consumption distribution and beyond.2 In other words, for most people, America is richer than anywhere else.

Common Failure Modes

You do not want to compare countries in terms of consumption per hour worked without making substantial adjustments. There are two main reasons for this.

The first and most important is that hours worked are endogenous, for multiple reasons. Firstly, there are income effects on labor supply such that richer countries buy leisure, meaning that hours per capita tend to fall as consumption rises, leading mechanically to inflation. Secondly, there are tax and transfer wedges that shift the labor-leisure margin independently of productivity.3 Third, non-employment isn’t leisure. For example, Americans and Western Europeans spend similar amounts of time on leisure, but Americans spend relatively more time doing market-remunerated work rather than doing work at home. Unemployment, being on disability, and being a discouraged worker are all things that cut into hours, and if a country boosted those, it would raise consumption per hour worked by putting people out of work, even though that’s not good for welfare. Finally, there’s the issue of selection: if the least-productive workers are the ones outside employment, the average worker is mechanically more productive.

The other big issue beyond endogeneity is that measures of hours worked are not comparable between countries, without putting in a lot of work. Consider the employment-to-population ratios for different countries in the OECD. For the U.S., it’s 0.5, for Germany it’s 0.545, for South Korea it’s 0.544, for Italy it’s 0.433, and for Luxembourg, it’s 0.765. Why is Luxembourg so extreme? Well, that’s because roughly half of its workforce are Belgian commuters who show up in the denominator but not the numerator. This is why I dropped it and Switzerland (0.591) when talking about consumption per hour. Similarly, it’s dubious to compare countries like South Korea and the Netherlands (0.565) given that their similar ratios come with very different hours per worker.

An even more major part of the comparability issue is culture. Turkey is an extreme example. Turkey has a low employment-to-population ratio for the OECD, in part because women’s market participation is low. Thus, because women—and in particular the most marginal laborers among women—don’t appear as represented there as in, say, Western Europe or the United States, consumption per hour worked at the household level can mechanically increase if it consumes at a decent rate.

The OECD is aware of these issues and provides an explicit warning not to compare countries in terms of its hours worked measure (levels). Rather, they would like the hours worked measure to be used to assess trends over time for particular countries or sensible groupings of countries. 4

It is also common to fail at doing international comparisons over time by using current PPPs instead of constant PPPs.5 With constant PPPs, you get a measure of real growth because it provides you with details on the change in the volume of a constant set of things which are being produced. However, accounting with constant prices can also overvalue things that are initially scarce after you start producing lots of them (e.g., software, a new fruit cultivar), so there’s room to err on either side with intertemporal comparisons. Know what you want to measure before measuring.

It is exceptionally unwise to chalk up a country’s consumption to mere trade deficits. Remember what a ‘trade deficit’ is: a current account surplus. If you have a trade deficit, that implies that your country takes in more from trade, investment, and transfers, than it puts out. This is not a bad thing, even though it’s widely perceived to be that way. Stated plainly, I actually think it’s hard to see America’s trade deficit as a bad thing, since what it means is that people are investing in America, which allows Americans to consume relatively more than people in other countries.

If we were to eliminate America’s deficit in isolation, then America’s relative position in terms of consumption is largely preserved, but its poorest citizens become notably worse off. America’s relative gap with other countries shifts from 13th-place for the 1st percentile to 15th place, from 3rd to 5th for the 5th percentile, from 1st to 4th for the 10th percentile, while the other percentiles explored above retain 1st-place positioning. The median gap with the population-weighted EU15 declines by almost 14%.

But is this meaningful? It clearly isn’t. If America closed its trade deficit overnight, it would radically disrupt the global economy and imperil the consumption not just of itself, but of everyone, everywhere. For America, the only ways to reduce the trade deficit are to cut down on investment—leading to sizable steady-state output loss—or to increase savings—mechanically lowering consumption for itself and indirectly lowering consumption for economies that export to it! For the rest of the world, since America absorbs 55% of all current account deficits, that saving has to be lent somewhere else at a lower rate, or simply not saved, and either way creditors end up earning less.

Chalking up America’s heightened consumption to a trade deficit is wrong on the numbers and it doesn’t even make a lot of sense. It is essentially arguing that America is ‘only so rich because its institutions are reliable enough that everyone wants to put their money there because there’s so much awesome stuff to invest in!’ OK! That seems like a great thing for America.

Shocks are bad! I’ve used 2022 data because it’s the most complete year. Unfortunately, this means that Norway’s consumption is inflated because of the gas shock that took place in that year, erroneously pushing it up considerably. This can happen, and is why the best picture incorporates multiple years of data.

Don’t randomly count out consumption you dislike. People often like to make arguments like ‘America only consumes more because it has a horribly expensive healthcare system.’ This is manifestly untrue, but people say it. Why? Ultimately, because America does spend a lot on healthcare. But let’s go over two things (which are maybe three things if you squint).

First, counting out healthcare, America still leads. Removing health spending in its entirety, America’s 1st percentile rank moves to 15th from 13th, its 5th percentile moves to 4th from 3rd, and its 10th percentile moves to 4th from 1st. Every other rank and most of the gap remain. Cutting out particular parts of healthcare like health administration, or making prices equivalent has smaller effects than just counting all of healthcare out, and the administration point is even being overly fair to comparator countries because America counts things like advertising in it and it doesn’t shift things like long-term care substantially into social spending or put scheme administration into collective consumption outside health measures. This is a comparability issue that’s slanted against America.6 (The conclusion also nearly persists if we also count away America’s trade deficit in isolation. America’s median goes to 2nd place behind Luxembourg by 0.2% if we do that.)

Second, this is just denigrating people’s spending choices. People want to spend on healthcare. It is a superior good for a society. As affluence grows, so does our ability and desire to spend, and so does the prevalence of diseases of affluence, so richer countries like the U.S. spend larger shares of their income on healthcare. This is true of other countries as well, and when there is an abrogation from this apparent reality, it is usually not lasting, as in the cases of the U.K.—which kept its spending out of line with its wealth by making healthcare practically unavailable—and Singapore—which had a high out-of-pocket share until people found it intolerable and moved it towards the norm for a country of its wealth.

Counting out healthcare spending is as arbitrary as counting out any form of spending, and it makes little sense to cut it out. Doing so simply unfairly penalizes richer countries. This is obviously the case, since everyone ‘wants’ to spend more on healthcare in the sense that as countries get richer, they will. Saying particular parts of the American healthcare system can and should be cut is more convincing, but rests on speculation about optimal governance, and the usual thing people suggest for America is single-payer, which means sacrificing consumer optionality, among other things. That is itself a notable tradeoff.

There are plenty of other common pitfalls in making cross-national comparisons, but I feel like I’ve listed enough here for now. If any notable additional examples come up, I’ll address them.7

Replication

420KB ∙ PDF file

Download

Download

This was a timed post. The way these work is that if it takes me more than an hour to complete the post, an applet that I made deletes everything I’ve written so far and I abandon the post. You can find my previous timed post here.

A message from my sponsor, Mechanize:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

Having a wealth of natural resources isn’t all upside; in fact, many of the countries with the greatest mineral bounties are among the world’s poorest and most dysfunctional. Economists have noticed this and dubbed it the “resource curse”, the paradoxical observation that countries with lots of natural resources have slower economic growth and worse living standards compared to countries with fewer such resources.

This has been explained in a few ways. Primarily:

• Dutch Disease: Money going to one industry or resource drives up the value of local money, so other industries like farming and manufacturing become too costly to produce exports.

• Price Changes: World prices for resources like oil and gold go up and down all the time, with high volatility, which makes the budgets of governments reliant on those prices unstable.

• Bad Leadership: Easy money from natural resources allows leaders to ignore expanding the tax base and going out and collecting taxes, it allows for lower public accountability because the leader’s power comes from a few easily secured resources, similarly, it allows for greater levels of corruption, and a fast track to dictatorship.

To escape the curse, economists recommend putting aside money from resource booms to buoy the country’s economy in future down times; using the money from resource booms to build up other industries; opening up the government and sharing its financial data publicly so as to prevent corruption; and, most important of all, using the money to educate the citizenry so they have the skills to diversify the economy and compete on the global stage.

But what actually happens when a country does all these things? Botswana did.

Growth in Sub-Saharan Africa is anemic. Between 1990 and 2023, the region’s mean real GDP per capita PPP rose by less than $1,000. The outlier was Botswana: its GDP rose by more than $8,000—an amount that’s small by first-world standards, but stands out as more than twice the average GDP in its region and puts the country a tier above.

Botswana has been exceptional since shortly after geologists with the De Beers company discovered some of the world’s largest, most accessible diamond stores there in 1967. This led to staggering growth multiples. But there’s no need to be vague; we can make this comparison much more precise by placing Botswana next to peer countries that started off similarly poor and plotting exactly how it’s grown:

During Botswana’s boom, they took economists’ advice seriously: they were a world-leader in saving resource rents (#1 among peers), building foreign reserve buffers (#1 among peers), investing in education (6 of 166 globally), maintained high national savings (16 of 183 globally), invested heavily in gross capital formation (#7 among peers), and—partly because of their Southern African Customs Union membership—they had the 4th-lowest tariff rate in the world. What Botswana did was not yet ‘textbook’, but would become it: they established a sovereign wealth fund to overcome resource volatility issues before Norway was the reference; they bound recurrent spending to non-mineral resource revenues so their mineral wealth could only fund investment; through rigorous anti-corruption efforts, they became far and away the least corrupt country in Sub-Saharan Africa, ranked #2 in regulatory quality, #3 in rule of law and government effectiveness, and ranked #4 in voice and accountability, with globally respectable rankings in each category.

Botswana’s governance success has allowed it to avoid any coups since 1966, to avoid ever being an IMF structural adjustment case, and has given it the leverage needed to renegotiate its agreements with the De Beers company, moving the state towards capturing ever larger portions of its diamond rents.

Nobel laureate economist Daron Acemoglu dubbed Botswana “An African Success Story” and chalked up its success to a combination of “relatively inclusive pre-colonial institutions, placing constraints on political elites”, “maintaining and strengthening institutions of private property [being] in the economic interests of the elite”, “Botswana is very rich in diamonds, which created enough rents that no group wanted to challenge the status quo at the expense of ‘rocking the boat’”, and “a number of important and farsighted decisions by the post-independence political leaders, in particular Seretse Khama and Quett Masire.”

But let’s be real: the story is that Botswana sold a lot of diamonds, and little more than that. There’s no meaningful evidence of precolonial institutional hand-offs that explain Botswana’s current success and most policy successes economists credit with keeping Botswana’s economy above those of its neighbors are icing on the cake rather than the sponge itself.

Botswana has invested substantially in education, but world-class educational investment hasn’t made their population any more cognitively capable. Their spending is longitudinally unrelated to the measured cognitive skills of their students, and it doesn’t set them apart from peer countries.

Consider Botswana’s lousy TIMSS performances:

Or look at Botswana’s pitiful SACMEQ improvements:

Botswana may nominally be doing things right when it comes to education, but they don’t have a lot to show for it. In fact, they may nominally also be doing things right when it comes to economic diversification, but it doesn’t look like they’ve managed to follow the typical route taken by developed countries, where manufacturing increases to a high level then falls back, to be replaced by services:

The chart may make it look like Botswana has rapidly moved towards advanced economy status, with a focus on services, but this is misleading. The kinds of services it’s funding are not that productive, and are in no small part driven by previous spending: a lot of this is the public sector, with a large share just being teachers’ pay, mandated by the state’s wasteful educational investments. Those sorts of unproductive jobs are alright to a degree, but without leading to other, more productive jobs being created, they’re not self-supporting, as in Botswana’s case.

But Botswana has diversified their economy. If you look at the diamond-specific share of Botswana’s GDP, you’ll notice it’s absurdly high, and indeed, the world’s highest there is. But, they have reduced that share from being nearly a third of everything to being just under 10% now. Success?

Unfortunately, Botswana’s diversification has not been beneficial; their economy has not grown as a result of moving away from diamonds. In fact, they’ve just had their first pair of consecutive years of GDP decline, even though many of their neighbors and mineral export-reliant peers are doing just fine in that time.

The issue for Botswana is that they are extraordinarily reliant on diamonds and the demand for diamonds is down. Specifically, the demand for natural diamonds is down. There are two things going on here. Firstly, the technology to create artificial diamonds that are large, clear, and cheap has advanced substantially. Lab diamonds were in 1-in-8 American diamonds purchased in 2019, and that’s risen to 61% by 2025:1

To make matters worse for natural diamond sellers like Botswana, marriage rates are down, and those account for the vast majority of diamond sales. 90% of U.S. engagement rings had natural diamonds in 2020 versus just 42% in 2025, and in that time, fewer Americans are getting married. The decline in marriage in China alone is enough to explain about 6% of the roughly 42% global decline in natural diamond demand between 2022 and 2025. An additional trend might explain another up to 14%, and that’s China switching towards gold jewelry and away from diamonds.

Through competition, lab diamonds have also pushed down the price of natural diamonds. As a result of all of these trends, the value of Botswana’s diamond production fell from $4.7 billion in 2022 to $1.98 billion in 2025. Accordingly, the Botswanan economic miracle has likely ended:

Despite decades of institution-building and population-education efforts, Botswana has only wound up a bit richer2 than peer countries, and it’s both unclear how long that can last and how it can continue growing from here. Despite institutional reform proponents’ suggestion that you can effectively set up quality institutions and let things ride, Botswana seems like the bull case, and it’s an example of good institutions allowing extensive resource extraction with much less in the way of lasting good.3

You can find everything needed to reproduce my results and more here. Happy coding!

This was a timed post. The way these work is that if it takes me more than two hours to complete the post, an applet that I made deletes everything I’ve written so far and I abandon the post. You can find my previous timed post here.

A message from my sponsor, Mechanize:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

1

An alternative view:

If you talk to a member of the public about data centers, chances are, they’re opposed. The most common arguments in opposition also generally have a left-wing character, focused on the environment or the displacement of our precious farmlands and natural resources. For a quick overview, see my plot of Gallup’s latest polling:

The other common arguments are usually uninformed or silly, but also nonpolitical: opposition to noisy buildings, disdain for having to pay higher taxes, a fear of AI replacing human workers, and so on. Few of the arguments are solidly right-wing by any measure. But one argument did recently come across my feed, and while I don’t think it’s going to become a nationally prominent one, I think it could become prominent online, and from there, it would generate headaches and acrimony for everyone involved in the discourse: the claim is that data centers “attract a segment of the tech workforce that is overwhelmingly foreign, mostly Indian.”

This is an interesting theory, because it’s very obviously untrue if you’ve ever been in a data center. Data centers employ few people (~50 for a 250,000 square-foot data center is normal), about half of the employees are contractors, and basically none of them are H-1B workers. In fact, if we look at Census data on the demographics of the people working in data centers, the whole idea looks preposterous:1

For most people, this is probably enough evidence to conclude that the claim about data centers attracting “a segment of the tech workforce that is overwhelmingly foreign, mostly Indian” is false, but there’s a way to defend the claim. One could argue that the alleged influx of Indians is due to the emergence of “the secondary ecosystem of cloud architects, IT engineers, consultants, and tech vendors that crop up in data center clusters.”

The evidence I’ve presented so far isn’t inconsistent with that idea per se, but if you know about the history of “data center clusters”, you’ll know that this claim is also false. If you understand data centers and thus that they do not create said secondary ecosystem, you’ll also realize the error in this idea. What’s more, we can rigorously demonstrate its inaccuracy with more data from the Census.

The Ecosystem Story

Does the construction of data centers spawn a secondary ecosystem composed of cloud architects, IT engineers, consultants, and tech vendors in the local area?

The first way to look at this is to treat any data center opening as the trigger, and the second is to look at places with >1 data center opening, so that the thing we’re looking at is definitely a “cluster” of data centers. The choice of operationalization doesn’t matter, however, because either one (and at any scale, hyper or not) fails to support the idea that these things cause a tech cluster—Indian-filled or otherwise—nearby.

That claim can be considered debunked.2

The Indians and Foreigners Story

Now, the whole idea that data centers attract “cloud architects, IT engineers, consultants, and tech vendors” would seem to contradict the idea that they also attract bunches of Indians. Unless we focus in on IT engineers—which Indians are quite overrepresented among—, then we’re at a loss with how to make this story about Indians, because Indians are not massively overrepresented among the other listed occupations. But, let’s test the claim anyway by looking at data center establishment effects on numbers of Indians.

This claim can be considered debunked.3

Indians do not show up in meaningful numbers in the areas around data centers. In fact, they tend to show up further away from them, within tech metro exurbs. The more plausible theory, with real evidence behind it, is that Indian tech professionals go where there are suitable jobs, where there’s demand and the possibility for people to fill tech jobs, as professionals of any race or ethnic background tend to do.

Who Cares, Anyway?

Well, for one, many people within the current Trump administration. There’s a lot of anti-Indian sentiment right now among appointees and pundits, and this mostly has to do with H-1B employment being seen as a means of taking away American jobs. H-1Bs don’t do that in aggregate, and in fact they stimulate employment and innovation in America, but I digress. The point I want to make is that the people against Indians are the people who care, and some of them have the power to turn this disdain into an actual cudgel against data centers, even if they might like data centers in the abstract.4

So let’s reassure those people: there’s no merit to this, one of the few right-wing anti-data center arguments. It can only be supported through ignorance of what data centers are and who works in and around them, and ultimately, it is simply false.

You can find everything needed to reproduce my results and more5 in this zip. Happy coding!

A message from my sponsor, Mechanize:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

If you just want to skip the data and get down to the part on how to source HGH, click here.

It’s true: people are paying to make their kids taller than they otherwise would’ve been.

And here’s the kicker: it works! By paying a little money up-front, parents are able to make their kids a few inches taller in normal cases and sometimes more than a foot taller in cases where kids are suffering from pathologically low growth hormone levels.

But this raises a nasty question: are these parents murderers? By making their kids taller, are they also raising their risk of different ailments, from cancers to lung tears? Are they giving their kids scoliosis and diabetes in exchange for a handful of precious, socially-valuable inches? After all, it’s common knowledge that taller people—socially advantaged as they are—live shorter lives because they get cancer more often, right?

The answer to that isn’t so clear-cut. Take a look all-cause mortality vs. height:1

But this diagram is based on observational data, it’s basically just a correlation, and we know that what it depicts isn’t likely to be a description of real cause-and-effect.

Consider why people might be very short. Why might a man be 4’10” for example? One prominent reason is dwarfism, which also tends to come with a number of other skeletal and metabolic problems that dispose people to shorter lifespans, above and beyond the effect of just being short. Likewise, consider why someone might be incredibly tall. A person who’s 7’3” is much more likely to be someone who has gigantism, which does indeed make you tall, but also puts you at risk soft tissue swelling, circulatory problems, muscular weakness, and so on.

A more familiar reason for very short stature is osteopenia—age-related bone loss. But if you’re experiencing that, then chances are you have other frailty-disposing comorbidities, and you’re likely to be quite old, with all that entails. To really get at the effect of height on lifespans, we have to go beyond confounded cross-sectional data. But first, we have to ask an even more complicated question:

Are the height gains people experience from taking HGH in their childhood like height in the general population? What if HGH users are gangly, with disproportionate limbs, freakishly widened jaws, and other signs of abnormal growth? Basically: Is it even valid to use data on the heights of normal people to infer something about the eventual heights of HGH users?

First, we need a prior. That prior is going to be “Size matters.” That’s what we’re all going to agree to believe, right now. Size matters. This is a widely accepted view on the risk for cancer: the more cells in the body, the greater the risk of cancer. In fact, this view was articulated in a well-known 2018 paper entitled “Size matters: height, cell number and a person’s risk of cancer”. The key idea from that paper, however, is much older; papers supportive of this case have been published as far back as 1988, more recently in 2013, contemporaneously with “Size matters” in 2018, and plenty of times after the advent of COVID as well. This has been investigated with merely observational and with causally informative methods, and they both agree: more cells, more cancer in humans and in other animals.2

Second, we need to know if HGH-driven growth is isometric or allometric. What this means—and I am slightly misusing these terms—is we need to know if it involves disproportionate growth in some areas compared to others (allometry) or proportionate growth (isometry) when administered throughout regular development.

As it turns out, most literature agrees that the growth is isometric, and that ‘HGH height’ is basically like normal ‘genetic height’. Some data indicates slightly greater renal growth relative to height, and some other data indicates slightly trunk-favoring growth, meaning that sitting height—which is generally a tad more heritable than standing height!—is impacted more by HGH administration. More importantly, compared to their similar-height peers who did not receive HGH, those who did are usually a little lighter, which I’ll return to.

We also have data on what would happen if a person went extremely overboard, because some people are genetically disposed to their bodies doing basically that. In cases with pathologically elevated HGH levels, as in acromegaly, there’s increased epithelial cell proliferation in the colon, leading to an excess of the small, bulgy pouches that happen when intestinal lining pushes through weak parts of the muscular wall when under pressure, and perhaps explaining their higher rates of colorectal and gastric cancer (among a few other elevated types). But even this elevation is slight, and it would be very odd for parents to push their kids this far.

So, to clearly and quickly summarize on the nature of HGH-driven growth as opposed to ‘genetic growth’: given approximate isometry, ‘HGH height’ carries ~70-100% of the cell number-cancer relationship prior, with a mild discount because HGH-treated kids—especially ones starting from a state of HGH deficiency—tend to end up somewhat lighter in terms of both bone and muscle mass. This relative lightness3 has to do with the fact that the HGH-added height is delivered to kids via stimulation of the IGF-1 axis, and for a person at a given ‘genetic height’, that stimulus remains, whereas for ‘HGH height’, it doesn’t. The ‘HGH tall’ kids just keep the height without the rest.

If you throw all of this information together, we end up with a prior that HGH-delivered height will be a little less cancerous than the same ‘genetic height’. But this doesn’t mean adding height is all bad: remember, taller people also end up healthier in other ways. For example, height improves lung function and lipid profiles. On the other side for the heart, it has some unfavorable effects on hemodynamics because it enlarges atria and creates longer venous columns. On net, the causal effect of height on the risk of various conditions and death is… not bad. And we know this because of the wonders of Mendelian Randomization—a method that exploits subtle differences in genetic endowments to estimate the causal effects of things as diverse as having higher or lower cholesterol or being taller or shorter. Take a look at how height affects different conditions:

And if we just want to get at causal curves, well, that’s easy enough! We can that if we sum to get the curves from our Mendelian Randomization! That looks like this:

In short, the expected effect of height is on net somewhat protective, with larger benefits for males than for females.4 We can’t extrapolate out to, say, the level of professional basketball players, but we can extrapolate out to every level that’s reasonable, and we can conclude that more height ultimately doesn’t look to be bad, and at the point where it might be bad, it’s only slightly so, and thus is likely ‘worth it’ from the child’s perspective.

But let’s add a final note here: the main source of risk from greater height is cancer, and we’re getting much better at treating cancer all the time. By the time children come to the rather advanced age where cancer is a real risk to them, we’ll likely have made significant progress, and possibly abolished some of the types of cancer they might face, leaving us with significantly greater upside.

Or, in other words, the case for making your kids taller is getting better all the time, and the case that you’re shortening their life by doing so keeps getting weaker.

The Parents’ Section: How to Source Cheap HGH

Read more

It sucks having a cold. For me, the worst part is the stuffed nose. When I have colds, I often reach for the one nasal spray that always works: Afrin, active ingredient, oxymetazoline.1 One puff per nostril and I’m clear for hours, saving me a lot of grief and making it possible to go about my day pretty much normally.

Afrin works, but I can’t recommend it. If your cold drags on for more than a couple of days, Afrin’s benefits turn to downsides. Oxymetazoline starts off by stimulating alpha-adrenergic receptors in the blood vessels in your nose, causing vasoconstriction: blood vessels tighten up and the blood flowing into your nose’s lining is pushed away and swelling goes down. That’s how Afrin opens your nose up, but after prolonged use, your nose adapts to this vasoconstrictor effect.

After a few days with Afrin, the nasal lining shifts to a more inflamed, leaky, and swollen state with local ischemia from repeated vasoconstriction causing reactive edema, increased vascular permeability, and chronic inflammatory changes.1 As a result, using Afrin can actually leave you more congested and you can become addicted; you can develop rhinitis medicamentosa, otherwise known as rebound congestion.

Suddenly, you need Afrin to be clear at all.

Hundreds of thousands and perhaps low millions of people come down with this condition every year by overusing Afrin and similar nasal sprays.2 There are whole online communities dedicated to quitting it. Those places are filled to the brim with people cycling in and out of addiction because they’re unable to put down the Afrin when it gets hard to breathe.

This isn’t some benign thing where the only cost is having to keep a spray bottle nearby. For one, a bottle of Afrin usually runs $10 to $20, and people who are severely addicted can run through a bottle every 2-3 days. Over a year, that adds up to about $1,200 to $3,700 in annual spend on a drug that should only be used when you’re sick.

Even if cost is no object, living with rebound congestion isn’t convenient. If you read the anecdotes on the various Afrin quitting forums, you’ll find stories about waking up every 2-3 hours at night, having to carry conspicuous bottles to use at functions, not to mention headaches, nervousness, dryness, sneezing, nasal discharge, dry mouth, mouth breathing, snoring, burning, stinging, and rarer side effects like turbinate hyperplasia and atrophic rhinitis to boot! Some people have even had their addiction result in high blood pressure, insomnia, and eventually surgeries.

Millions of people have turned a temporary stuffy nose cure into a self-sustaining chronic nasal disease, that brings with it poor sleep and generalized low quality of life.

Given how many people are affected by overuse of Afrin and Afrin-like drugs, you’d think there would be a standardized, high-quality protocol to get off it. Unfortunately, despite recognition by the medical profession, no one has gone through the effort to make that happen. So, to that end, I leveraged my subscribers for some testing.

I randomized convenience-sampled subscribers, Discord users, and friends to different strategies in groups of 20. Two strategies involved continued use of Afrin and the others involved either mechanical nasal dilation or the use of different nasal sprays. Here’s what we did:

• Single-Nostril Comedown: This group was tasked with stopping their wholesale use of Afrin and instead just using it in one nostril, until they felt they could breathe in the other one, whereafter they would try to stop using Afrin entirely. 19 completed the trial.

• Saline Wean: This group was to keep using Afrin, but was instructed to refill the bottle to the top with saline every night. That way, they would end up using a more and more diluted product over time. 16 completed the trial.

• Saline Rinsing: This group was asked to rinse their nostrils as-needed with saline. They were also able to use a standard saline nasal spray. 32 of 39 completed the trial and I treat this as a sort of baseline or placebo group for everything else.

• Flonase: This group used a nasal spray containing the corticosteroid fluticasone propionate every day. 20 completed the trial.

• Flonase+: This group used as nasal spray containing the corticosteroid fluticasone furoate every day. 19 completed the trial.

• TCA: This group used a nasal spray containing the corticosteroid triamcinolone acetonide every day. 14 completed the trial.

• Nasonex: This group used a nasal spray containing the corticosteroid mometasone furoate every day. 15 completed the trial.

• Xylitol: This group used a xylitol nasal spray every day. 16 completed the trial.

• Xylitol+: This group used a xylitol and capsaicin nasal spray every day. 11 completed the trial.

• Azelastine: This group used a nasal spray containing azelastine HCL every day. 13 completed the trial.

• PO: This group used a nasal spray with 0.6% or 1% povidone iodine every day. 19 completed the trial.

• IC: This group used a nasal spray with iota carrageenan every day. 17 completed the trial.

• Intra-Nose: This group was asked to use an Air Max brand in-nose nasal dilator and, if they needed it, saline. 14 completed the trial.

• On-Nose: This group was asked to use an Intake brand magnetic nasal strip and, if they needed it, saline. This group started and ended with 18.

Everyone was told to use standard dosing but was able to do more if they needed, they just had to use the correct compound and avoid failing to Afrin.

What Worked?

Here’s the main result in one picture. This is affirmative responses to ‘Are you now able to breathe clearly unassisted?’ Volunteers were asked if they were able to wake up and go throughout their day without using nasal sprays or mechanical assistance.

The bar coloring has to do with steroids (blue) vs. Afrin strategies (red) vs. mechanical methods (teal) vs. the rest (purple) vs. the placebo (soft periwinkle). This is because these are similar enough and I wanted to get statistically meaningful results so I had to group different interventions together. Given the variation within these groupings, you can probably tell that I mainly did that to get at the least varied and most successful group: the steroids.

Relative to the ‘placebo’ saline-only group, steroid spray-using completers were 48.9 percentage points (pp) more likely to end up with a clear nose at the end of a month. Assuming the people who didn’t make it to follow-up or who told me they couldn’t make it were all failures, then steroids had a success delta of 42.9pp, with 78.7% succeeding vs. 35.9% of those in the placebo group. This is the most robust result:

If you want to get off Afrin, use a steroid nasal spray if you can.

The other results are much less clear—for now (see below). For example, Xylitol+ is a xylitol nasal spray with capsicum oleoresin (capsaicin) plus aloe and grapefruit seed extract. It stings a little bit for some people and it achieved a success rate of 64%. But, it also had the highest attrition rate. If you assume the non-completers were failures, then Xylitol+ actually looks slightly—nonsignificantly—worse than basic Xylitol.

Anyway, the correlation among completers between rated feelings of clearness and rated feelings of comfort was a modest but statistically significant ρ = 0.57.

What’s the Catch?

In my opinion, there’s one catch, but it might not even be real: If you use everything as directed, the costs of these different treatments are nominally a bit higher than the costs of Afrin.3

In general, the treatments for Afrin are more expensive than using Afrin 1-2 times a day. But that comparison is biased in favor of Afrin.

For that graphic, I used name-brand prices, but you can get off-brand prices that are much lower for several of these options. Additionally, people who are addicted to Afrin frequently use more than they should, so this understates Afrin-related spend. Plus, if you break an Afrin addiction, you don’t have to keep using the alternatives, you can just stop using them altogether after recovery. And, the cost of Afrin is not just measured in dollars! Rhinitis medicamentosa is a real medical condition that is no fun to have and it comes with tons of common side effects.

The way I’d think about the Afrin alternative side effects is like this:

• Saline rinsing, plain xylitol, iota-carrageenan, and the mechanical devices have rare, mild, and local side effects—temporary burning, maybe causing a minor nose bleed if you use them really incorrectly, and possibly some ear fullness if your sinuses are particularly messed up. Saline is also just obviously safe and in the broader literature, both xylitol and iota-carrageenan are well-tolerated to the point where it’s hard to take side effect worries seriously.

• The intranasal steroids can cause some local irritation: nosebleeds, burning, headache, sore throat, cough, and other nuisance effects that will mainly affect you if you’re using the products incorrectly, if you’re particularly messed up, or if you’re new to the products and having a reaction because you’re one of the minority of people who needs to get used to them. The most serious concerns for this intervention class mostly exist on their labels—eye effects, infection risk, systemic steroid effects, etc.—but not in the real world with typical or even double or triple the recommended OTC nasal use.

• Azelastine has moderate costs. It helps with symptoms, but compared with intranasal steroids or saline-like approaches, it has a more noticeable tolerability ‘tax’ in the form of its bitter and unusual taste—which you can avoid by using the kids’ product, which has the same dose and efficacy, but tastes better—, nasal discomfort, fatigue, and sleepiness. But azelastine is still quite safe.

• Xylitol with capsaicin may have moderate-to-high costs in the form of discomfort. Capsaicin has plausible therapeutic properties, some data to back them up, and it feels good to me and most of my friends, but if you’re not used to spice, you might find your nose burning, you might end up sneezing and coughing, and possibly feeling some short-term irritation. I love this stuff and some people I know hate it. It definitely produced the most polarizing reactions among people I talked to, but it is quite safe.

• Povidone iodine has very formulation-dependent effects. Low-concentration products can be reasonably well-tolerated with mild transient nasopharyngeal discomfort as a common side effect, while higher concentration intranasal formulations frequently cause temporary burning and nasal pain. It’s basically more painful saline and I don’t think it’s up to the task for helping people get off Afrin, so maybe ignore it as an option (but research its other benefits!).

Compare all of these things to Afrin and there’s just no question! Quitting wins!

Afrin addiction—rhinitis medicamentosa—is a serious, self-imposed chronic disease; it means prolonged congestion, dependence, and risk of other nasal diseases far beyond anything these alternatives cause. The alternatives sometimes cause transient, localized annoyance, but they don’t generate dependence, they don’t lead to other serious medical conditions, and they don’t leave you with the incessant worries that being stuck on Afrin usually does.

What’s really remarkable is that to the extent people get off Afrin, they seemed to agree. General side effect reports correlated at ρ = -0.89 with success: as much as one of these options helped to get people off of Afrin, they felt it had lower side effects.4 This doesn’t match up with the literature on these drugs’ effects in normal use, but fixing Afrin addiction is decidedly not normal use, so I think this finding makes sense:

Afrin withdrawal is so bad that whatever side effects alternative drugs had, the relief from being able to come off Afrin was so great that effective treatments were experienced as having fewer side effects overall.

The Extended Writeup

I plan to keep updating this article as more data comes out, and that includes putting future trial results down below, too. To keep up to date with it and to see everything else that I’ve written on this subject, keep scrolling!

• Statistical Analyses

• Outside Data For Every Option

• Untested Possibilities and Things I Didn’t Have Big Enough Groups For

• Supplements: Do They Help?

• Bees! Bees! Bees!

• Combination Treatments

• The Cold Turkey Group

• Can We Design an Effective Mechanical Dilator?

  • Smelling Sticks?

• Cocaine, Ketamine, and MDMA

• The Guaranteed-Success Option

• Another Trial?

• Testimony

• Analyzing Forum Sentiment

  • Reddit

  • Amazon Reviews

• Does This Work For Really Long-Term Addiction?

Read more

Coffee causes cancer. Or at least, we used to think it did.

Back in 1981, MacMahon et al. observed that there was a strong dose-response relationship between coffee consumption and pancreatic cancer. This association survived controls for cigarette use and appeared internally robust, but in the wider literature, the association collapsed almost immediately due to a textbook problem: the control group was drawn from patients of physicians who diagnosed cases, and those patients had been advised to cut down on coffee intake.1

Substantive and direct replication failures started rolling out and large-scale meta-analytic data failed to confirm any link among cohorts of people who’d never smoked before.

Shortly after the idea of coffee harms was crushed, the literature flipped to the idea that coffee is not just not associated with cancer, but instead, that it’s actually protective against it. (You would think the null results would temper expectations; alas, no.)

In 2012, Freedman et al. found a ~10-15% all-cause mortality reduction for people who drank two or more cups of coffee a day. In fact, coffee seemed to be protective for most things, save for cancer. Gunter et al. then reported that more coffee drinking was associated with lower mortality, with particular benefits for digestive and circulatory diseases. An umbrella review by Poole et al. corroborated these findings.

But protective findings were based on the same sorts of problems the initial harmful ones were: they built on confounding artefacts, whereby sick people quit drinking coffee and healthy people opted into drinking coffee. To get around this sort of problem researchers would have to do be a bit more clever. To that end, they used Mendelian Randomization!2

Leveraging the fact that there are genetically controlled differences in caffeine metabolism—and thus in the reason to even drink coffee in the first place—researchers were able to show that coffee consumption does not seem to cause any risk elevation or reduction for mortality, cancers, etc. And to close the loop, it doesn’t seem to cause pancreatic cancer.

Coffee was viewed as harmful, then it was considered helpful, and now it’s just nice to drink. This is a story that’s been repeated many times. There’s a mountain of proposed food items, drinks, habits, and supplements that have been proposed to hurt or help in broad ways but which have later been found to have essentially null effects outside of the areas we know they definitely help in: Coffee is great for energy thanks to the caffeine; it doesn’t make you live longer. Viagra is great for erectile dysfunction; it doesn’t make you live longer. Statins are great for preventing heart disease; they don’t cause dementia.

One proposed target for study is “hot drinks”—tea, coffee, maté, etc. Several recent papers have claimed that consumption of drinks like these, above a certain temperature, causes cancer. One of the proposed mechanisms is thermal injury: every time you drink something hot, you burn the mucosal lining of the esophagus and what gets built back is more likely to be something with an error that leads to cancer.

The strongest evidence for this theory comes out of Iran, and specifically from the Golestan Cohort Study.

The Golestan Cohort Study shows that the consumption of drinks at 60°C or greater is associated with a ~40% greater risk of esophageal cancer. Consumption of ‘hot tea’ is associated with a 60% greater risk of esophageal cancer, and in another permutation of the analysis, a 107% greater risk, with ‘very hot’ tea being associated with 716% higher risk of esophageal cancer.

So, are hot drinks bad? I wouldn’t say that on the basis of the Golestan Cohort Study.

There are a lot of issues with this study. To get to the meat of the matter, hot drink consumption is heavily confounded with factors like sex, age, drug consumption3, and so on, and virtually everyone in this cohort drinks a lot of tea every day. If you believe the results of this study, you can’t just believe them piecemeal: the result that says hot drinks have a large effect also says that cigarette smoking has no effect! As it turns out, in Golestan, those things are heavily confounded, so the obviously cancer-causing effects of cigarettes can’t be distinguished from the effects of hot drinks.

Cigarettes obviously cause cancer, so we cannot accept the conclusions of this study unless we qualify them in a really convoluted and odd way. But why do that? There’s not even a consistent dose-response effect to temperature, as you’d expect under the thermal damage theory. I say it’s better to move on. To that end, I recommend looking at this with the National Health and Nutrition Examination Survey (NHANES), which I’ve used for several other blogposts.

The first thing I did was I looked to see if hot drink consumption being linked to cancer replicated in the NHANES4. It did. But, even using basic controls eliminated any signal. As a positive control, this was not true for smoking. As it turns out, the hot drink → cancer link in the U.S. looks to reflect the fact that older people drink more hot drinks and also get more cancer, whereas smoking just seems bad.

We can verify this confounding signal by scrutinizing the specific types of cancer that hot drinks seem to ‘increase’. Several cancers cannot be increased by hot drink consumption because there’s no way for there to be thermal damage effects and the substances in coffee and tea are not carcinogenic in their own right. Even switching around the comparison group materially changes the results. For example, comparing hot drinkers to non-drinkers, lung cancer seems elevated for the drinkers, but looking among hot drinkers, consumption is robustly linked to lower risk of lung cancer!

If you look on the right-hand panel, you’ll see that some things even look to be carcinogenic in some analyses, when we know from causal analyses that they are not, providing us with more assurance that the pattern observed for hot drinks in some cohorts can arise even without causality being present.

To be sure, I’m not the first person to notice these sorts of odd findings. Some people have been motivated to look into oddities around the claim that maté causes cancer. For example, Bates et al. found that it was related to cancer among ever-smokers, but not never-smokers, and that the cancer in question didn’t make sense, as it was bladder cancer. De Stefani et al. also conceded that temperature couldn’t explain a bladder cancer risk bump and Weng et al. failed to affirm the link with tea, with largely culturally disjoint samples—a likelier explanatory factor.

It’s unlikely that hot drinks cause cancer. Where we have causally informative evidence for a particular type of hot drink, there’s nothing to see.5 Where the evidence is more limited, suspect, and seems to be wholly dubious, there looks like there’s something, but it’s clearly confounded to such an extreme extent that if we accept the results, we have to also accept ideas like that smoking doesn’t cause cancer. Where the quality of non-causal evidence is very good, we can make out that the signal is clearly confounding-driven and it disappears with adequate controls.

Frankly, I think we can just put this idea to bed and say that hot drinks probably don’t cause cancer. To the extent they do, it has to be limited, below the level detectable with samples reaching into the tens of thousands.

This was a timed post. The way these work is that if it takes me more than one hour to complete the post, an applet that I made deletes everything I’ve written so far and I abandon the post. You can find my previous timed post here. I wrote this because a friend wanted me to assess whether hot drinks were likely to cause some of his health problems.

A message from my sponsor, Mechanize:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

Unless you’ve been living under a rock, you’ve probably heard by now that Anthropic’s latest model, Fable 5, has been taken down after the U.S. government issued an export control directive to suspend access to all foreign nationals. Whether or not you believe this action was justified de jure, it seems fairly clear that, to some extent, Anthropic brought this upon itself; for example, in an essay published shortly after the Fable 5 release, Dario Amodei compared advanced AI to nuclear weapons and called for proactive government regulation:

The government should have the power to block or deter deployment of the model if it is determined, in light of third-party assessment, to present unacceptable risks.

[...]

[I]t is my very strong belief that AI is something much more profound, something that resets the whole game board and around which all future geopolitical strategy must be shaped—like nuclear weapons, but potentially even more so.

To be sure, Dario did not call specifically for the imposition of export controls on strong models. Nor did he say specifically that access to Fable 5 itself is akin to possession of a nuclear warhead. But history has shown time after time that the Leviathan, once awakened, defies control. There is a deep irony in the fact that the founder of Anthropic, a company originally focused on AI safety and interpretability, failed to appreciate this historical observation.

Whether or not you agree with Anthropic’s characterizations of Fable’s cyber capabilities, what is done is done, and enough ink has been spilled about this topic. I am concerned, however, by the approach that Anthropic is taking to biology. It is a confused mixture of ungrounded rhetoric, poor execution, and fear-mongering. There is a real possibility that just as Anthropic has essentially ‘shat in the soup’ as far as models’ cybersecurity capabilities (including defensive ones) are concerned, it may end up doing the same for biology.

Precisely—Anthropic is now also in the habit of making strong public claims about biological risk while demonstrating poor public evidence of biological competence; there is a real risk of producing onerous government overreach and hindering beneficial scientific progress. The result of immaturely posturing about their models’ capabilities and openly inviting regulatory action may halt development of model capabilities in the laboratory sciences—potentially a major accelerant to the discovery and validation of novel therapies over the course of the next several decades—and lead to harsh regulations on the proliferation of existing biotechnology. Imagine, for example, if every model release from now on were required to adhere to the same biology-relevant safety guardrails as Fable 5 was upon release. This would mean that you would be blocked from asking questions about personal health problems or undergraduate coursework; in fact, merely having a preexisting history of interest in biology would plausibly lead to a blanket refusal on all queries.

Technological development is inherently dual-use

To begin at the highest level of disagreement—Anthropic has argued multiple times that AI is, at minimum, a technology with dual-use capability which must be regulated. Indeed, in Policy on the AI Exponential, Dario writes:

I believe the best analogy, at least at the current stage of the exponential, is to cars, airplanes, or drugs—powerful technologies essential to the modern economy, but capable of killing large numbers of people if designed or operated poorly. I therefore believe we should model AI regulation on agencies like the Federal Aviation Administration (FAA).

This argument is only compelling in the most superficial reading. Take the example of drugs. Nations such as India or Turkey have notoriously looser controls on pharmaceuticals than the U.S., yet inappropriate use of pharmaceuticals is not a major cause of death in those countries. It is likely good to have some degree of federal certification for aircraft, but the 1978 Airline Deregulation Act is broadly considered a major success, with cutthroat market competition afterward driving prices down while maintaining an excellent safety record; in fact, if there is danger associated with flying, it may now originate predominantly from the FAA itself due to its failure to properly hire and qualify new air traffic controllers. And similarly, while it is believable that drivers’ licenses should exist in some form, it is likely that people take some care to not mow over pedestrians not because they will lose said licensure but because they would simply be charged for the crime of murder.

Importantly, there is no coherent, empirically founded argument made as to why access to current frontier models enables mass murder to a greater degree than other technologies. It is easy enough to claim that Fable 5 “make[s] cyberattacks substantially easier and cheaper to commit,” but this is a generic claim that is true of essentially any advance in computer technology and discounts the defensive capabilities provided by models. More generally, all technology is dual-use, and the burden of proof lies rightfully upon the claimant to demonstrate that the differential benefit versus harm supplied by a particular nascent technology is uniquely imbalanced so as to warrant strong government regulation. Moreover, Dario falls prey to a particularly subtle form of survivor bias when he brings up these examples of regulation—they are familiar to us precisely because they are the ones which have proven durable enough to survive into the modern day. How many people today remember, for example, that cryptography techniques were previously included in the United States Munitions List and subject to draconian controls?

Of course, we should prevent access to nuclear weapons. But let us not forget that nuclear weapons were purposefully built to kill people. In contrast, technology developed by private companies is generally built to satisfy customer demand, and history teaches us that we ought to have a strong prior that such products are not so inherently dangerous as to warrant preemptive regulation. Just because I can imagine some hypothetical, catastrophic dual-use scenario in a written essay does not by itself justify anything; the real world should be governed not by ability to transform a vivid fantasy into tedious verbal argumentation but rather by actual empirical data and careful study. (Also, given the existence of Chinese models, any such regulations would have no practical effect beyond delaying access to frontier capabilities for under a year, but that’s a different conversation…)

Anthropic fundamentally misunderstands biological progress

Another reason why Anthropic should exercise caution in opining about biosecurity is that, despite considerable financial investment in the field, they simply do not seem to understand basic aspects of scientific progress or the biomedical sciences. It ought to be a basic principle that if you fail entirely to understand some specialized field, your attempts to advocate for special regulation in that field should be entirely ignored, because it is rather unlikely that someone failing to grasp basic facts would possess anything resembling a coherent model of the threats and risks in said field.

The ignorance of basic biology displays itself at multiple levels. In Dario’s recent essay, Policy on the AI Exponential, he writes:

Many of the steps in the clinical process that previously required expensive and slow experiments may soon be done via AI simulation or analysis. [...] Areas where this could apply include:

• AI-based pharmacodynamics and pharmacokinetics (PD/PK) modeling;

• Prediction of toxicology to avoid the need for multiple species animal toxicology;

• More accurate dose selection, to reduce to the need for large dose ranges in trials;

• Biomarker validation via analysis of large datasets;

• Synthetic control arms in clinical trials, to reduce the need to recruit more participants;

• Developing surrogate endpoints (particularly important in aging and neurodegeneration).

The problem with criticizing this passage is that it is so profoundly misguided that one hardly knows where to begin. Does Dario truly believe, for example, that pharmaceutical companies are simply too inept to consider the idea of synthetic control groups, or is it more likely that synthetic controls are genuinely difficult to formulate and ultimately less convincing than proper randomization? Does he believe that ‘more accurate dose selection’ is an actual limiting factor in the proper design of a clinical trial, or is it more likely that given the $100M price tag of a Phase III trial, the dose selection is probably optimized enough such that remaining margin for improvement in this direction is relatively low even if you were willing to pour more cognitive labor into it?

More generally—and here is where the relevance to biosecurity comes in—Dario’s writing is all premised on some tacit assumption that performance on these tasks can in some sense be dramatically improved with sufficient supply of cognitive labor. This is wrong for the simple fact that even with a true abundance of model intelligence, this work is fundamentally bottlenecked by the inherent need to sample, slowly, from the complex distributions of real-world biology, which cannot be so easily disentangled through sheer might of pure reason. To take a concrete example, holding the disease fixed, one can posit the existence of some perfect drug for that disease which exists somewhere in the 10⁶⁰ space of all druglike small molecules. However, the mere existence of such an optimal compound does not mean that one can, even with superhuman intelligence, efficiently reason your way anywhere close to that point.

Indeed, one of the core lessons from human history is that complementarities matter. Consider the following example from an essay by Mechanize, Medical AI isn’t the bottleneck to medical progress:

The case of MRI machines is instructive. While they’re now a cornerstone of modern healthcare, MRI machines would have been impossible to develop and mass manufacture in 1925, no matter how much people spent on medical R&D. At its core, a contemporary MRI machine consists of a large donut-shaped magnet built from superconducting coils. To achieve superconductivity, these coils are cooled with liquid helium to temperatures near absolute zero. Yet until the mid 20th century, reliable helium liquefaction plants and widespread distribution networks simply didn’t exist. Moreover, MRI machines rely on extremely sensitive radiofrequency electronics to detect faint proton signals, a technology that was highly underdeveloped until the 1940s. Affordable computers capable of quickly performing Fast Fourier Transforms, without which MRI machines would be incapable of producing coherent images, didn’t come out until the 1970s. All of this means that, from the standpoint of someone in 1925, MRI machines required far more than laboratory research to build: they needed a larger and more advanced economy, with new supply chains, supporting industries, and specialized technologies.

In this particular case, just because a model with sufficient intelligence could in principle reason about the existence of MRI machines does not mean that there is actually a ‘search path’ leading to the development of MRI machines that does not involve broad economic growth and the confluence of multiple unrelated technological developments. In a separate essay, the Mechanize founders argue—quite convincingly—that technology not progress in sharp leaps and bounds, but instead in a relatively predictable sequence where novel discoveries occur shortly after all of the prerequisite ‘input technologies’ are developed:

Technological progress occurs in a logical sequence. Each innovation rests on a foundation of prior discoveries, forming a dependency tree that constrains what we can develop, and when. You can’t invent the telescope before discovering how to grind optical lenses, or develop electric lighting before learning how to generate electricity.

We did not design this tech tree; it arose from forces outside of our control. The evidence for this lies in two observations: first, technologies routinely emerge soon after they become possible, often discovered simultaneously by independent researchers who never heard of each other. Second, isolated societies converge on the same fundamental technologies when facing similar problems and resource constraints.

Simultaneous discovery is common. The Hall–Héroult process now used to smelt the world’s supply of aluminum was discovered in the same year, 1886, by different people, separated by an ocean: Charles Martin Hall in the United States, and Paul Héroult in France. Both worked independently, using the same method of dissolving aluminum oxide in molten cryolite and using electrolysis. Neither knew of the other’s work.

This is not merely a theoretical argument. The development of new technology, including that of model capabilities, is, as we have argued, inherently dual-use, and the threat model which Anthropic is utilizing originates from a worldview which predicts that we might see extremely sharp, unpredictable developments, hence motivating preemptive government action, rather than a gradual unfolding of events allowing a more measured, responsive approach.

To be sure, if one day models were incapable of high school chemistry and then the next day able to order precursors, autonomously direct laborers, and synthesize sarin gas for you, that would indeed be some cause for concern. And if one day a model is suddenly able to cure a disease or invent a revolutionary technique, that same model may very well have the capabilities to formulate deadly toxins or methods of chemical genocide yet unseen. But that is not the world we live in. Instead, we live in a world where the growth of both model capabilities and fundamental science and technology progress gradually, even if at a rapid clip. Intelligence alone can accomplish a great deal, but their concerns here appear more theological than empirical.

We should not assume competence by default

When it comes to Anthropic’s actual biosecurity content, it is relatively difficult to evaluate the quality of their thinking or the rigor of their benchmarks for the simple reason that we are not able to actually inspect the contents of their biosecurity evals. Take, for example, the Fable 5 system card, which contains Section 2.2.4, Biological risk results: automated evaluations. Some of these evaluations already appear quite suspicious, such as a “virology-specific multimodal multiple-choice evaluation.” It is unclear to me exactly how the ability to answer a multiple-choice examination well, even if it may at times include images, videos, or other files, is coherently related to a notion of biological risk.

More problematic, however, are benchmarks such as the “task-based agentic evaluation” of “virology tasks,” the “DNA Synthesis Screening Evasion” benchmark, or the “Black-box RNA sequence design” task. What are the input data to these benchmarks? What are the precise computerized settings in which these evaluations are run? What are the actual model outputs and how are they evaluated? Because none of this information is known to us, we cannot actually inspect the data and judge for ourselves whether or not Anthropic is conducting them in a sensible way. Perhaps it is fine to play somewhat loose with these details if you are measuring, say, your model’s ability to answer math questions. However, if the entire premise of conducting these evaluations is to justify the imposition of onerous safety blocks and government regulation, then we should expect them to be conducted well and to reflect genuine measurements of biological risk.

It is plausible that others have seen the contents of these evaluations. I have not. For all I know, they could be perfectly fine. Indeed, Anthropic has many intelligent employees, and I am sure they take pride in their work; how bad could it be? Perhaps we ought to simply give them the benefit of the doubt.

Recall that the definition of Gell-Mann Amnesia is:

The phenomenon of a person trusting newspapers for topics which that person is not knowledgeable about, despite recognizing the newspaper as being extremely inaccurate on certain topics which that person is knowledgeable about.

I would be far more inclined to yield them that benefit if they had any high-quality public output in the domain of biology that was not obviously sloppy, misguided, or erroneous. This is not a high bar to meet; I am not asking for a novel scientific contribution here but merely some indication of ordinary levels of competence in execution. However, such evidence is not forthcoming.

Take, for example, their two most recent blog posts with scientific content. The first post is an evaluation of Claude on twenty NMR-based problems, which are essentially undergraduate-level content. They proudly present favorable comparisons against ChemDraw and MestReNova; when I asked an expert in the field about these tools without supplying the context of the post, they immediately commented that they had previously tried to use them and were disappointed by the quality. The second post is an overly effusive diatribe glazing up their implementation of a command-line tool, gget virus, which allows agents to more easily access a virology database. They claim that usage of this tool allows agents to achieve 100% performance on their benchmark, VirBench. This is certainly not a bad thing; however, the entire premise of this post is that AI agents are bottlenecked by their ability to navigate complex, legacy user interfaces for biological data. However, from their own graphic, progress on VirBench has advanced from 16.9% one year ago (Claude Sonnet 4) to 91.3% (GPT-5.5). If the benchmark is essentially going to fully saturate within the next half-year through general capabilities improvements alone, I’m not sure that this is exactly the highest-value contribution they could have made to agentic AI in biology, not to say anything of the rather specious premise that agentic AI is actually bottlenecked because it cannot correctly download the requisite data.

A more pertinent comparison would be the quality of one of their biology benchmarks which is publicly accessible. Thankfully, such a benchmark does exist. Two months ago, they released BioMysteryBench, a benchmark which is meant to measure model capabilities in bioinformatics “while tackling some of the challenges inherent in evaluating complex and noisy biological systems.” If this benchmark is well-constructed and shows signs of care, then perhaps that would serve as a favorable datapoint indicative of a similar level of care in their biosecurity work.

I had the chance to examine this benchmark in detail. (Anyone can do so following a short application on HuggingFace.) It is, unfortunately, chock-full of comically sloppy work; the results of a brief examination are as follows:

• One problem asks for the “human organ” which serves as the source of a particular dataset. Upon Googling several of the genes in the dataset, it is clearly a dataset of mouse genes.

• One problem asks to identify which sample is contaminated with A. fabrum. However, when you actually examine the attached genomic data for the supposedly correct sample, the sequence matches the fungus S. pararoseus nearly perfectly, and has no evidence of contamination by A. fabrum.

• One problem asks the model to identify a genomic breakpoint and specifies the correct answer as an insertion at position 4,144,431. When specifying an insertion, you can insert to the right of position i or the left of position i+1, and both of these obviously refer to the same insertion. The prompt does not fix the reporting convention, leading to grading problems.

• One problem asks for the “complete species name from NCBI Taxonomy.” The supplied answer is “Norovirus GII.4,” but in actuality, NCBI Taxonomy clearly states: “Norovirus Hu/GII.4 … belongs to the species Norovirus norwalkense.”

• Some problems have rather suspicious numbering issues. For example, one problem asks the model to correctly classify and order samples 1 through 15 according to developmental stages. Oddly, the rubric specifies that the correct answer consists of stage 0 (sample 1, 2, 3), followed by stage 1 (sample 4, 5, 6), followed by stage 2 (sample 7, 8, 9), and so on… I asked Codex to solve this problem, and it responded that the earliest stage is actually represented by samples 10–12, which seem to be in the L5 (5th larval instar) stage. Similarly, another problem asks the model to identify patients diagnosed with ulcerative colitis out of a series of numbered samples. The rubric tediously enumerates the correct answer to be: Sample_1, Sample_2, …, Sample_74. I asked Codex this problem across three different sessions and it yielded a much different and noncontiguous set of sample IDs ranging from 5 to 24.

• In general, every problem is formulated as a question paired with a single free-text rubric item. The rubric item is typically of the form, e.g., “Expected answer is [so and so]. Score 1.0 if the answer meets the criteria above.” This is extremely odd, because essentially every question in this benchmark can be graded with 100% accuracy by specifying a fixed output format, rather than relying on unreliable model-based grading. There is actually simply no benefit to using model-based grading for problems with such simple answers, especially if there is only a single, simplistic rubric item being specified rather than a variety of different criteria.

To be clear, I did not have the time to exhaustively verify each question manually, so it’s plausible that some of these critiques may be mistaken. However, if a nonprofessional in the field is able to identify multiple points of concern within the hour after opening your benchmark dataset for the first time, that is not a good sign as regards the overall level of quality of other biology benchmarks for which you are publishing results. If anything, I am rather concerned that there may be quite a few more subtle errors which I did not notice.

Closing thoughts

The purpose of this post is not to lambast Anthropic’s biology team; instead it is to urge them to do better. Extraordinary claims require extraordinary evidence, and if you repeatedly shout from the rooftops that you are building dangerous technology which merits immediate regulation, those claims merit scrutiny, especially when the U.S. government is actually listening to and taking your claims seriously.

What is Anthropic’s impact in biology so far? To date, they have:

• Released a sloppy benchmark along with several blog posts of little interest

• Acquired an unknown company with fewer than 10 people for $300 million dollars

• Claimed that their models are approaching thresholds of severe biological risk

• Implemented extremely stringent safety guardrails that prevent even the simplest biological questions from reaching Fable 5

This is all quite sad, because the potential impact of frontier models on biology really is tremendous. It has the potential to save millions of lives. But combining fearmongering with a fantastically nonsensical mental model of how biology—and science in general—works is not the path there.

A message from my sponsor, Mechanize:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

This was a timed post. The way these work is that if it takes me more than one hour to complete the post, an applet that I made deletes everything I’ve written so far and I abandon the post. You can find my previous timed post here.

If you’re reading this, you’ve probably seen some permutation of the meme that ‘I didn’t move right, society moved left’. In the above cartoon, the idea is that today’s ‘far right’ are the normal people of yesteryear. But, is that true? Does it matter?

To figure out the answers to my questions, I opened up the General Social Survey and had a look around. To get started, I defined a few sets of political views: Institutional Confidence, Criminal Justice & Guns, Political Tolerance, Economic/Pro-Government, Racial Liberalism/Civil Rights, Gender-Role Egalitarianism, and Sexual & Moral Liberalism, and then I outlined a set of important social views. We’ll go in order.

People have become less confident in America’s institutions over time.

When it comes to criminal justice and firearms, more people think courts are too harsh on criminals, more people oppose the death penalty, and fewer people think we ought to requires permits to buy guns. The last of these didn’t change that much.

Political tolerance has somewhat increased since the 1970s, but it’s somewhat down since the 1990s. This general trend masks something interesting: more acceptance of gays and atheists, less acceptance of racists and militarists.

Movement on economically left-wing views has been generally pretty flat, which is roughly what we also see for economically right-wing views.

When it comes to racial liberalism, people have shifted far to the left. People have become more likely to ascribe Black-White gaps to discrimination and less to a lack of effort, among other things.

When it comes to gender roles, egalitarianism has greatly increased.

Finally, when it comes to sexual and moral liberalism, views on sex education, divorce, marijuana legality, and so on have, in some cases, quite radically shifted towards left-ward positions. Abortion legality for serious defects is the only exception among the bunch.

All of this sums up to saying that the nation has moved considerably, but not overwhelmingly, to the left.

Does this support the thesis that a normal person from 30 years ago would be far-right today? Not on its own. Many of the views that society has shifted decisively on remain well within the Overton Window and are certainly not so generally proscribed that they’re rejected out of hand.1 The sort of position that is nowadays rejected firmly is like what the average person from a little over 30 years ago believed on interracial marriage: they didn’t approve.

What can we say about the normal person from 1995? In many ways, he was much like us. In other ways, he was what most people would now regard as kind of a dick. He didn’t like interracial marriage and he wasn’t too keen on gays either. These social views aren’t alone: plenty of things now considered taboo were, at the time, wholly acceptable, even in polite society, and sometimes these were the majority view.

Right or wrong, I don’t think appealing to normal people in the world 30 years ago is likely to make many friends after realizing the sorts of things people used to believe.

A message from my sponsor, Mechanize:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

1

Another interesting finding is that the correlations among different dimensions have changed as people and their views have politically shifted. For example, economic and racial views are now correlated about twice as highly as they were in the 1970s.

GLP-1 drugs are so effective at helping people safely lose weight that, if they were made widely available, the obesity rate could be pushed below 1980 levels in a single year. The issue is availability.

Retatrutide, CagriSema, Amycretin, Zenagamtide, Survodutide, MariTide, VK2735, Ribupatide, Berobenatide, Enicepatide, Elecoglipron, Pemvidutide and several other GLP-1-based drugs are all poised to enter the U.S. market within the next two years. If they all manage to make it through the FDA wringer, then there’ll be ten different companies pushing GLP-1 drugs, plus whoever else makes it through in that time. This is a dramatic juxtaposition with the current situation where the only competitors on the U.S. market are Eli Lilly and Novo Nordisk.

Competition improves products and lowers prices, benefitting consumers, right? Perhaps in theory, but in reality, these products are barely differentiated. The main factor setting any of them apart is how long doses last—a week or a month. But more importantly, I don’t think it’s likely that competition is going to meaningfully lower prices!

The general trend in medicine is for the introduction of new branded drugs to barely affect the prices of existing ones.1 The introduction of generic versions, on the other hand, leads to massive cost savings. Take a look:

The difference has to do with the fact that price competition only occurs when the buyer of the drug is treating the options available to them as interchangeable and the buyer is price-sensitive, with the ability to pick between options. By law, generics satisfy these requirements, but differentiated brands don’t necessarily satisfy any of them.2

With this in mind, what I expect from more GLP-1s is that the prices will remain exorbitantly high. They might go down a bit, but not so much that they’ll become affordable for average Americans or worthwhile for most health plans. That won’t happen until there’s a generic, and due to their abuse of the patent system, Novo Nordisk has managed to stall the entry of a generic in the U.S. by five whole years.

If we want to realize a world where America is skinny, we’ll have to figure something out. Unfortunately, there aren’t many viable options.

Use the Section 804 Importation Program. America has a program to import drugs from Canada, at the prices they’re sold at up there. This program needs reformed to be able to allow the mass importation of GLP-1s at Canadian generic prices, but it can be done. I’ve described how to do that, here.3 Meanwhile, this program can be used to at least match Canada on branded prices, which are quite a bit lower for them.

Protect Foreign Drug Import Optionality. The FDA has historically allowed people to bring in up to a 90-day supply of their prescriptions from abroad. They could, theoretically, allow people to import prescription drugs from a country with vastly lower prices like Brazil, India, or China, with the importing party being the one to take on all the risk. This is a very hands-off approach that’s unlikely to get much use, but it would allow some consumers to obtain licit, non-gray market, GMP drugs from other countries, at remarkably fair prices, with reasonable safety guarantees.

Protect Compounders. Compounding pharmacies have been the greatest force for getting down American GLP-1 prices to date. They’ve managed to deliver the drugs to millions of patients with almost no problems as a result. They’ve saved thousands of lives and billions of dollars. Novo Nordisk and Eli Lilly hate them for doing this, since they’ve undermined their sales. They’ve been trying non-stop to get the FDA to step in and stop the compounders from selling what is basically their drugs, for less. But the compounders are operating legally—if dubiously, and in the role of regulatory arbitrageurs—, and they should continue to be supported to keep prices low, so long as their quality remains sufficiently high.

Negotiate Down the Prices. The Inflation Reduction Act created a Medicare price negotiation program that has selected to negotiate down the price for the semaglutide products Ozempic, Rybelsus, and Wegovy. This takes effect on January 1, 2027 and constitutes a 71% discount from the list price for the CMS, meaning they’ll just pay $274 for a 30-day supply. This could be expanded on or pursued more aggressively.4

Protect the Gray Market. Gray market sales have allowed hundreds of thousands and perhaps millions to obtain very cheap GLP-1s without much issue. This option forces customers to take risks on their own and to do quasi-licit orders that might leave them scammed out of their money, but patients should be able to take these risks into their own hands. If this were a bigger thing, it could materially damage sales for branded drugs, but because patients are naturally averse to paying out-of-pocket, they want quality guarantees, and so on, it’s not a legitimate threat and it never will be. It’s also not worth thinking about in this case, since the branded drugmakers have extracted massively excessive profits by abusing America’s intellectual property laws.

Buy Off a Patent. Congress could make an offer to one of the companies with an existing, approved GLP-1, paying them to give up their patent and allow generics into the market. This is unprecedented, but the authority is there, and if Congress wanted to make it happen, it could. It could do so at a relatively meager cost, too. Think about it:

What would happen to the value of, say, Novo Nordisk’s patent on semaglutide if the Presidential administration and the Congress pursued all of the options that I listed above? What if they also…

• Put Medicare, Medicaid, ACA, FEHB, VA/DoD, and willing employer plans into a pooled purchasing process? This was done for hepatitis C drugs.

• Made coverage expansion conditional on a low net price?

• Attempted to tie public program prices to peer-country prices or to require rebates when prices exceed a benchmark level? They could even force manufacturers to rebate amounts above a benchmark price for drugs with a high budget impact.

• Challenged improper Orange Book listings, limited 30-month-stay gamesmanship, and made it easier for follow-on versions to launch when core patents expired?

• Authorized contract manufacturers to supply the government while paying the patent holder their “reasonable and entire” compensation per 28 U.S.C. §1498?

• Let states pool GLP-1 purchasing and offer preferred placement in exchange for deeper supplemental rebates?

• Force negotiated rebates to reduce what patients pay at pharmacies?

• Required PBMs to pass through rebates and fees and choose drugs based on lowest net cost rather than on the basis of the largest retained rebate?

• Prioritized review of follow-on GLP-1s, generic liraglutide and semaglutide pathways, alternative delivery forms, and competing incretins?

• Paid full price only for patients who achieved agreed-upon outcomes like weight loss thresholds, A1c improvement amounts, low discontinuation, etc.?

• Paid different net prices for diabetes, obesity, cardiovascular risk reduction, sleep apnea, MASLD, etc.?

• Conditioned public coverage or preferred status on reliable supply, vial options, and transparent production?

Basically, the idea is to lower the margins associated with the on-patent drug the government wants to buy by as much as possible. Then, Congress could approach, say, Novo Nordisk with an offer that’s not just reasonable, but cheap given all the value low-cost GLP-1s would produce. This is possible, if only we have the will.5

This was a timed post. The way these work is that if it takes me more than one hour to complete the post, an applet that I made deletes everything I’ve written so far and I abandon the post. You can find my previous timed post here.

A message from my sponsor, Mechanize:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

2

A generic is bioequivalent and AB-rated to a brand. Legally and clinically, it’s indistinguishable, and in some states, pharmacists must substitute the generic without calling the prescribing doctor to tell them. A ‘me-too’ brand, by comparison, is a different molecule, so they are not automatically substitutable and buyers can’t treat them as one commodity, so they fail to compete as one commodity. This is basically the difference between Bertrand competition and monopolistic competition.

In a branded drug class—meaning one without generics—, there’s an agency problem because the prescriber picks the drug, the patient pays a copay, and the insurer pays the balance but doesn’t make the prescribing choice. Thus, nobody is shopping on price and demand is price-inelastic, without any pressure to make cuts. A second branded option just gives prescribers another high-priced option. With a generic, on the other hand, the decisions moves towards the pharmacy or the PBM, which is price-sensitive, either capturing the spread on acquisition or having a payer mandate to do the lowest-cost dispensing.

A ‘me-too’ branded drug also holds its own patents and exclusivity, so it never has to compete on price because it competes on differentiation, opinion, and PBM rebates for formulary placement. Rebates help drugs to obtain position, and they frequently end up pocketed by middlemen. Generics only play a role when exclusivity has lapsed, and they don’t have marketing, so they can only compete on price, which they do. Thus:

But note where competition does matter: it matters on the launch of new branded drugs. More branded drugs in a category means lower initial prices for the new ones.

We can also redo the earlier chart, as such:

I’ve been asked about fifty times in the last few weeks if taking {Viagra or Cialis}1 every day will make people live longer. The bull case for these drugs is simple: both drugs are PDE5 inhibitors, which raise cGMP signaling, downstream of nitric oxide, improving vasodilation and vascular smooth muscle relaxation. This is why they give people erections and how they help with pulmonary arterial hypertension. It’s also plausible that this benefits endothelial function, lowers vascular tone, improves perfusion, and reduces rates of ischemic injury, explaining their known benefits.

Or even simpler: they’re good for your heart, and your heart is often the organ that ends up killing you, so if you take the boner pills, you’ll live longer. Right?

There is some evidence these drugs might help. For example, one study found that men with erectile dysfunction (ED) were 90% more likely to have a cardiovascular event. A review of cohort studies found that people who took these drugs had a 30% lower risk of dying over the follow-up period. Another review even found that Cialis was associated with 32% lower risk of dementia, with Viagra trailing it somewhat, with a still-considerable 25% risk reduction.

But is that convincing evidence to anyone? Is ‘guys with ED get more heart attacks’ sufficient to conclude ‘If we treated their ED with Viagra, they would therefore have fewer heart attacks’? Is ‘people who take Viagra die less often’ enough to infer that ‘If we gave more people Viagra, then they would die less often’? Is ‘people who take Viagra are less likely to develop dementia’ really what you’d conclude from that association, or would you think ‘Guys who are still having sex with their wives are probably a group that’s less likely to become demented, regardless’?

Selection issues in this domain are absolutely enormous. The people using Viagra are mostly guys who are seeking it so that they can have sex. They also have enough wherewithal to pursue treatment and fill a prescription, and they’re physically healthy enough to withstand a vasodilator. They are not random people; the facts that they live longer, look healthier, and seem to be less demented are not surprising even if Viagra doesn’t help them one bit.

Researchers are not unaware of this issue; their studies frequently do include attempts to address it, but mostly in ways that are far from sufficient. One of the better examples comes from a cohort study of Alzheimer’s dementia where the researchers included a three-year lag to account for protopathic bias, meaning that they tried to account for the fact that people in the early stages of Alzheimer’s are still symptomatic in ways that might make them less likely to go and seek out Viagra.

In their lagged model, the initially significant ‘protective’ effect, whereby taking Viagra seemed to reduce Alzheimer’s risk by 18%, became a statistically nonsignificant 7% reduction.2 There may be room for protection, but I’m still doubtful: given that the preclinical and prodromal—meaning early, pre-diagnosis—period for Alzheimer’s is more than three years3, this lag certainly doesn’t go far enough.

The gold standard test that would answer the question of whether Viagra is life-extending is a large, long-term randomized, controlled trial (RCT). The only problem is that there aren’t any of those and the outcomes that have been assessed in RCTs so far aren’t the outcomes we want.

For dementia, we have two trials: OxHARP and ETLAS-2. These trials didn’t look at dementia per se, but they did look upstream, to cerebral vasculature—the blood vessels in the brain. They found that patients with cerebral small vessel disease—a condition characterized by reduced blood flow to the cerebrum—there was some suggestive evidence of improved perfusion in one of the trials, but there was nothing for cognition. In another small trial, there was some evidence of blood flow improvement, but all of this is very indirect and besides the point. There’s not a lot of good evidence for Alzheimer’s and dementia.

For cardiovascular outcomes, we have several trials, but their results are very indication-specific. For pulmonary arterial hypertension, these drugs are clearly beneficial. For heart failure, there don’t seem to be improvements or harms. In some cases, there’s even evidence for harms, like for patients with persistent pulmonary hypertension after corrected valvular heart disease or for another type of heart failure patient. For more normal people, like men with suspected coronary artery disease, Viagra at least appears to be safe, but it doesn’t seem to be helpful.

For life extension itself, this is where the RCT evidence is weakest. There’s practically nothing out there for all-cause mortality. The best evidence we have is in this review of Viagra for ED and diabetes, where there appeared to be a 65% mortality reduction, but despite such a massive effect, it was nonsignificant. The combination of a huge effect and statistical nonsignificance suggests this literature is just too imprecise to infer anything from.

The only other trial-based evidence we have on life extension is only tangentially relevant. In this trial, people with pulmonary arterial hypertension who were prescribed higher Viagra doses lived longer: over 32 months, 26.4% died in the 5mg group, 19.5% died in the 20mg group, and 14.8% died in the 80mg group. It’s hard to generalize from a group with this high a death rate, especially given the condition they needed to be in to enter the trial and that fact that there was no placebo group. The PASSION trial for heart failure patients was also uninformative, because in it, they had to stop early due to a medication supply problem. At the end of the trial, all-cause mortality was five-times higher in the treated group, but the confidence intervals were very wide, so the result isn’t useful (p = 0.04).

That’s the gold-standard evidence and it’s bad! It’s certainly not enough to actively recommend everyone start popping boner pills like we’ve been seeing a lot of podcasters tell their audiences recently. But we shouldn’t stop there. We technically don’t need RCT-based causally informative evidence to figure out if these pills are likely to be good. We can also update by leveraging causal-ish evidence.

For example, we have a large-scale active comparator study to go off of. An active comparator study means that we’re observing people with the same conditions being prescribed different drugs, one of which is the drug we’re interested in. After some additional matching on observables, these groups should, in theory, be comparable, allowing us to make causal inferences about the drugs. Alas, among patients with pulmonary hypertension, there’s no protection against dementia for patients prescribed Viagra versus those prescribed endothelin receptor antagonists—drugs that do not help with ED, while still helping with blood flow.

The causal-ish evidence that’s informative and well-conducted almost singularly points to null effects, and the causal-ish evidence that’s poorly conducted and obviously open to massive residual confounding just can’t be taken seriously. But we’re not out of evidence yet. There’s one more method to use: Mendelian randomization.

As it turns out, some people are born with genes that basically make it so they’re living life on Viagra. Much like how there are people who basically live life on statins and there are people who basically live life on GLP-1RAs due to the genes they were born with, these people are just lucky. But we’re also lucky, because we get to look at them and see how they’re doing.

If people with genetically-proxied Viagra are doing well, that provides us some positive indication that the drugs, taken over a very long period of time, will also make the people taking them better off too. So, how are they doing? Great!4 Look:

People who have basically been on dosing Viagra every day since birth have lower risk of erectile dysfunction and much lower risk of pulmonary arterial hypertension, both of which the drugs are used to treat. This constitutes positive control evidence that the genetic proxy works as expected.

Looking down further, we get to the juicier dementia outcomes: considerably lower risk of vascular dementia—as expected given its cardiovascular origin—and no increase in the risk of Lewy body dementia. However, there is a warning sign that contradicts some of the cohort evidence. It looks like these people have very slightly increased risk of Alzheimer’s dementia. But we’re lucky because that signal does not withstand scrutiny and the other signals do hold up, among men and women.5

My verdict, if I had to guess, is that {Viagra or Cialis} is good for cardiovascular health quite broadly and mildly, and if you’re at high risk of bad cardiovascular outcomes, then it will probably help you to live longer and to live healthier. Moreover, it’s good to take anyway for the treatment and prevention of ED if you’re a man.

But the key question for this post isn’t about a few particular risks. The question we really want an answer to is if preventing vascular and Lewy body dementia and slightly lowering the risk of cardiovascular problems translates to greater life expectancy.6 Unfortunately, I don’t think so. The mechanism behind these drugs does not tend to show up as a target for life extension.7 When we explicitly test if it helps with lifespans like I did in this earlier article,8 we also get a disappointing result:

• Lifespan, overall: +0.20 years (p = 0.576)

  • Men: +0.38 (p = 0.412)

  • Women: -0.08 (p = 0.881)

• Parental lifespan: -0.11 (p = 0.540)

• Healthspan: HR = 0.988 (p = 0.623)

• 90ᵗʰ-percentile longevity: OR = 0.824 (p = 0.043; marginally significant harm)

• Both parents 90ᵗʰ-percentile longevity: OR = 1.004 (p = 0.668)

Viagra has many benefits, but it does not seem like the drug that will extend lifespans.

A message from my sponsor, Mechanize:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

When you think of the people interested in longevity—people who ‘want to live forever’—, what comes to mind? Early bedtimes? Rigid dieting? A 60-item supplement stack? Peptides? Five hours a day in the gym? Never eating out or drinking alcohol? Testing your girlfriend’s vaginal microbiome before having sex?

All of that is fair—it certainly has become part of the image—, and, to varying degrees, those things can be fine; they’re even useful for some. But much of that defeats the point of trying to live forever. I don’t know if I even want to live forever if it means eking out a regimented, overly-engineered existence. That can scarcely count as living.

Take a look at this tweet from longevity advocate David Sinclair. Try not to be aghast.

Despite all the things David does for his health, all the advice he gives to others for their health, and all the research he does in the hopes of improving human health, David can’t handle a single snack bar before bed.

Is this longevity? Is this how you wants to live your life? Do you want to be cowed by carbs? Paralyzed by protein? Fearful of fat? Hoping for the right nutrient timing, lest your glucose spike a teensy tiny bit and you wake up having made yourself—and let’s be clear, this is psychosomatic—ill?

This is not the longevity movement I want. Longevity should be grander: immunity to the environment, life without the need for concessions, eternal life rather than a few piddly years of extension. It is possible.

Think about the torrent of significant medical innovations coming out, one after another, right now. Just in the past few years, we’ve gotten to witness the rise of GLP-1s, the advent of PCSK9 inhibitors, cancer after cancer transitioning from a death sentence to a few months of treatment followed by near-complete recovery. And biotech is just getting started. It can deliver a grander longevity movement.

Consider a recent phase-1 result from Eli Lilly for their new gene therapy VERVE-102:

VERVE-102—or more likely, something similar to it—will make it so people can eat whatever and live however they wish and still have low LDL cholesterol, allowing them to avoid most heart disease. When this happens, we’ll have achieved a one-and-done solution to a problem that prematurely kills millions of people each year.

Vertex Pharmaceuticals is working on implantable islet cells to treat type-1 diabetes. If those work, the condition might be curable, and maybe one day even non-diabetics could benefit, reclaiming or improving their glycemic control. There are a handful of stem cell companies working on fibrosis reversal, and it seems to be working—that could be another major killer down. Gene and antibody therapies for muscle preservation could help to defeat age-related sarcopenia and make weight loss more healthful. It’s possible to edit people for resistance to numerous diseases like HIV and malaria, and there are even experimental anti-drug vaccines that bind nicotine, cocaine, opioids, methamphetamine, etc., making it so less reaches the brain.

The list can go on and on and on. There are people working on therapeutics that allow you to feel just as rested on less sleep. There are companies working on genetically-modified versions of common gut bacteria to prevent hangovers. There are companies pursuing vaccines for tons of common diseases from chlamydia to gonorrhea to Epstein-Barr. Some companies are working on ways to permanently lower inflammation. Alnylam seems to be close to a once-every-six-months injection to control blood pressure. Cyclarity may be poised to actually reverse atherosclerosis.

No one can seriously doubt that we’re on the cusp of having a much better world through biotechnology. But everything I’ve listed so far only gets us to where proper longevity work starts; it’s all longevity normalizing: it brings up the left tail of the lifespan distribution.

Normalizing is how lifespans have increased across countries historically. Take Sweden and compare the distribution of ages at death in the 1750s to the distribution in the 2020s and you’ll see what I mean:

Another way to look at this is to look at the percentages surviving to given ages across different years. This chart hopefully makes the problem clear: despite massive increases in average life expectancies, the right tail barely moves. We’re bringing the left tail up, but not actually seeing extension of maximal lifespans.

If we have a revolution in everything related to what I’ve listed so far, we conquer the travails of living. We become effectively immune to our environments: an end to infections, an end to degradation from plaque accumulation and the stress of glycemic spikes, a practical end to withering. It also means an end to wellness culture—no regimented dieting required, no extra benefits from structured exercise programs and retreats. You’ll be able to drink and party and you’ll be no worse off for having done it.

But that revolution does not conquer death; it lets us live, and it makes room for what’s next. After reaching that state and finding that maximal lifespans still aren’t ticking up and healthspans still degrade, the focus should naturally shift towards promoting more radical approaches: facilitating immunity to mutation, regeneration of limbs, tissues, and organs, cell programming, targeted DNA repair, and eventually, slow, methodical, and hopefully unnoticeable full-body replacement. That is the goal.

A message from my sponsor, Mechanize:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

Spend enough time talking about heart disease online and you will inevitably be confronted with the claim that lowering your cholesterol increases your risk of getting dementia. Why? The reasoning can look like this:

‘Your brain is full of cholesterol. Statins inhibit the cholesterol synthesis enzyme HMG-CoA reductase, and sometimes they enter the brain. Low cholesterol has been associated with dementia in a handful of observational studies. Therefore, LDL lowering with statins or other forms of HMGCR inhibition must deprive the brain of cholesterol and cause dementia.’

And, if someone is reasoning about mechanisms, it can look like this:

‘Statins lower cholesterol levels, leading to a cholesterol shortage in the brain, causing dementia. This happens because cholesterol is needed to support cell membranes in the brain, because it is crucial for myelin, because it’s needed for synapses and synaptic vesicles, and if you lower your cholesterol, all of these things become broken.’

The problem for these and similar theories is that, when you check the data, it doesn’t line up. Mendelian randomization lets us get at the effect of taking statins or other LDL-lowering medications for a person’s whole life—decades of exposure!—and we see basically nothing when it comes to signals of harm. If anything, statins seem good!1

I think these results are only shocking because of semantics. People hear that something called ‘cholesterol’ is involved in the brain and assume that anything that lowers cholesterol must therefore lower it in the brain, depriving the brain of something like a critical nutrient, a requirement to function. But ‘cholesterol’ does not just mean one thing; you do not just have one cholesterol level.

The way statins work is by lowering liver-based cholesterol synthesis, resulting in relatively more hepatic LDL receptor activity, pulling more LDL from the blood, resulting in lower risk of atherosclerosis because of reduced risk of plaque buildup. But the brain, while cholesterol-rich and cholesterol-dependent, does not have its cholesterol supply affected by cholesterol-lowering medication. The main reason for this is the blood-brain barrier (BBB) and the fact that lipoproteins in the blood do not freely enter the brain. Accordingly, to get its cholesterol, the brain has to make and recycle it locally.

Reviews are unequivocal about this:

[C]holesterol in the [central nervous system (CNS)] is almost entirely synthesized within the brain since the BBB prevents any direct transfer of sterols from the blood to the brain, especially when they are contained in lipoprotein particles.

Peripheral (blood) cholesterol and CNS (brain) cholesterol are practically distinct. The brain produces its required cholesterol locally in glial cells, astrocytes, and oligodendrocytes, and it transports within the brain using ApoE-containing lipoprotein-like particles. Where there’s a relationship, it’s primarily through the brain disposing of cholesterol by converting it into 24S-hydroxycholesterol, which crosses the BBB and deposits cholesterol in the bloodstream to bring it to the liver for elimination.2

The National Lipid Association’s (NLA) advice is to lower your cholesterol, for as long as possible, to achieve better heart health.3 This is good advice because, for most people,4 LDL is the key causal factor involved in causing heart disease—specifically, atherosclerotic cardiovascular disease. This is completely unambiguous; with very low LDL, most heart disease goes away.5 And, if you followed this evidence and took the NLA’s advice to heart, you would likely not encounter elevation of dementia risk.

Consider the FOURIER trial. People who attained very low LDL with a PCSK9 inhibitor and a statin didn’t show any worse cognition after 2.2 years. We saw similar results in IMPROVE-IT after six years. Results like these are so often repeated, and so strongly supported by the bulk of trials, prospective cohort studies, and Mendelian randomization evidence that the American College of Cardiology simply maintains that reducing “LDL-C is not associated with cognitive decline or dementia.”6 Full stop.

With that said, LDL lowering does—as we’ve seen—seem to reduce the risk of vascular dementia, and it certainly reduces the risk of stroke. Both of those things either do or can involve serious cognitive impairment, and should for that reason be avoided. If they really cared about cognitive impairment, opponents of statins would be jumping for joy about this. But, revealingly, they remain opposed.

I suspect that the biggest reinforcement for confusion comes from failing to think causally about this issue.

Consider this distinction: someone with very low LDL due to prescribed lipid-lowering drugs is not being cholesterol-starved in the brain. However, someone who has near-zero cholesterol without a medical indication is more likely to have a rare genetic disorder like abetalipoproteinemia or hypobetalipoproteinemia, which comes with impaired fat-soluble vitamin handling and subsequent neurological and eye problems. If you looked at these people in an observational study and didn’t realize their disease state, you might think low LDL causes risks of its own. But, these people are high-risk less due to LDL, and more due to their other issues, which also cause low LDL.7

Additionally, in late life, low cholesterol sometimes indicates illness, frailty, weight loss, inflammation, malnutrition, cancer, or preclinical dementia, even though it’s not its cause. This sort of reverse causation has tricked tons of people into thinking that there are harms. We are lucky, then, that most of the observational evidence seems to be unaffected by these things. Accordingly, if we look at all of it, then we won’t make a big mistake when it comes to LDL, but, two things: first, that’s not always the case; and, second, it’s tempting to look at outlier studies. The best we can do is continue thinking causally, lest we wind up confused.

A message from my sponsor, Mechanize:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

5

Because of the lack of other mechanisms for plaque buildup.

Back in 1970, this would be 91% of all heart disease deaths, and as of 2022, it’s 53%. This leaves behind arrhythmias, heart failures, valve disease, cardiomyopathy, myocarditis, congenital heart disease, hypertensive heart disease, and other conditions, and the main reason for the decline in heart disease is that we’ve substantially tamed the LDL-driven afflictions, even though several factors involved in aggravating it—such as obesity and glycemia—have gotten worse. These other factors are less under the control of the medical system.

This is considerable prima facie evidence that statins have had an amazing impact on cardiovascular health.

If population blood pressure (hypertensive heart disease; heart failure with preserved ejection fracture; atrial fibrillation and many arrhythmias), glycemia, and weight were better controlled, most of the remaining heart disease deaths would likely also disappear. However, this would still not eliminate many cardiomyopathies, myocarditis and pericarditis, congenital heart disease, or many cases of pulmonary heart disease, since they are driven by genetics, viruses and inflammatory disease states, alcohol and drug abuse, chemotherapy, pregnancy, thyroid disease, autoimmune disease, toxins, maternal pregnancy factors, etc.

If you achieve low LDL and are otherwise fit, without major genetic defects, developmental insults, drugs, or serious infection, you are very unlikely to get any form of heart disease.

Today’s post is brought to you by my sponsor, Mechanize. They’re hiring junior software engineers at $300K/year base salary.

Always kvetch. That’s how anything ever gets done. In this case, that seems to be how the major electronic health records (EHR) software company Epic was convinced to finally provide quarterly updates on obesity and GLP-1 usage for over 100,000,000 (one hundred million) Americans.1 Now, we can finally resolve a little debate started by The Economist last year: Has America’s obesity rate been declining? The answer? Yes!

Epic has also been generous enough to supply us with quarterly data for 2021 onwards. Adjusting for seasonality in that period, we can see that the fall is continuing into 2026. This might inform us about the causes. Think about it: if the trend was being driven by a mass die-off from COVID, as some have proposed, for example, then that moment should’ve passed, but the trend is continuing. In fact, every quarter lately is lower than the corresponding one from the previous year:

We still don’t know what’s driving this, but we have hints.2 For example, the population is getting older—a potentially countervailing influence—and, perhaps more crucially, the population is increasingly using GLP-1RAs. In fact, as of mid-2024, some 1-in-5 Americans had tried them. Judging based on Epic’s dashboard, that number has likely continued to rise, and the EHR-recorded number has now become ~25% for diabetics and ~13% for the obese.3 Adding plausibility, we know that GLP-1RAs are capable of wiping out obesity.

While not totally satisfying on answers, this data release from Epic is a good start.

A message from my sponsor:

We’re hiring software engineers to build environments and evals that frontier AI labs use to train coding agents.

To get a better sense of the work we do, you can check out GBA Eval, where we had models build Game Boy Advance emulators from scratch and scored their performance.

Base pay starts at $300K/year for junior software engineers, with more for senior roles, plus equity and performance bonuses. Apply here.

This was a timed post. The way these work is that if it takes me more than one hour to complete the post, an applet that I made deletes everything I’ve written so far and I abandon the post. You can find my previous timed post here. I finished this post and then waited to obtain the ad section to add after finishing. In that time, I added the Epic tweet.

When looking at population cohorts that are (1) measured with consistent methods1, (2) measured with consistent sampling2, and (3) measured when they’re sampled3, there seems to be a recent trend towards increasing testosterone levels. As far as I’m aware, there’s only one cohort that meets all three criteria, and that is what it shows.

The big question this raises is Why? People are getting fatter and older, so why—and, more importantly, how?—are they also getting manlier? Let’s check some possibilities.

Is the rise race-specific? No, the rise appears to similarly affect Asians, Whites, and Hispanics, with Blacks showing a somewhat different trend that’s consistent with stability in the last two periods—no rise, no fall.

Is the rise age-specific? No, in unweighted analyses, there’s a rise at every age range, and in weighted analyses, it’s at almost every one. Every age has hormonally masculinized or stayed roughly the same.

Did everything rise? By this I mean, did all of the lab tests go up? If so, then the rise could represent a general lab work artifact. Alas, this is not the explanation:

What about changes in measurement instruments or the timing of sampling? Those also don’t have an effect, as the measurement methodology is constant, and controlling for the time of day of sampling leaves the rise intact.

What about free testosterone levels? Those also seem to have risen. We know this because we have measures of sex hormone-binding globulin in some waves, and levels have fallen, so the free androgen index has gone up for two reasons: the increase in total testosterone and an increase in testosterone availability.

Did estradiol rise in lockstep? Curiously, estrogen levels did not also rise. They were basically flat, so the testosterone/estrogen ratio increased majorly.

Is this a birth cohort comparison problem? We see a rise recently in most available birth cohorts in terms of decade of birth, with the exception of those born in the 1940s. So the trend is largely not about that issue.

Increased…? Decreased…? Based on dietary recall, protein consumption isn’t explanatory: grams per day are down 9.7% 2011-12 to 2021-23, per kilogram of body weight. Carbs are also down 14.5% in the same time span, as are total calories, by 9.2%. And all of this is demographically consistent. Moreover, physical activity doesn’t seem to have increased, although that’s hard to say for sure since the physical activity questionnaire was redesigned during this time series. However, the sedentary time measure was consistent and changed from 382 to 371 minutes per day, which isn’t a lot of change. Triclosan levels fell; BPA, parabens, benzophenone-3, all fourteen of the phthalate metabolites in the data, PFAS/PFCs, urinary arsenic, mercury, and iodine, urinary VOCs, perchlorate/nitrate/thiocyanate—all of these things were also irrelevant in the waves they were measured, so it’s unlikely they mattered between waves. PBCDs and blood VOCs, blood lead and cadmium, all of that was a bit odd as well, since those went down and were positively related to testosterone within waves.

Marriage! Married men tend to have somewhat lower testosterone levels than unmarried men and marriage rates have fallen, but the trend is present in both sets, so that’s unlikely to explain the full trend.

A related possibility is the decline of fatherhood, as childlessness rates are up and having kids is related to lower testosterone. But alas, this also isn’t likely to be explanatory:

Could it be the rise of testosterone replacement therapy? (TRT) This actually seems like a real possibility, until you notice that, for example, there’s a rise at the median and the mean, across most of the distribution, lopping off extremes doesn’t do much (and that’s where TRT-related gains would be concentrated), and the gains are age-stratified in the wrong way, being smaller at ages where you would expect more TRT usage.

What about the decline of drinking and smoking? Oddly, drinking and smoking are related to higher testosterone. If I had to guess, experimentally, cutting those things would be good for testosterone, but at least cross-sectionally, their fall doesn’t seem to be a relevant factor.

Diets? Dining habits? GLP-1s? Ultraprocessed food? None of that seems relevant. Nor does marijuana usage, cholesterol levels, waist circumference, sex, PCOS, DHT conversion, blood pressure, sleep, or anything else that I could control for (and that list is long).

Ultimately, I don’t think I have a good explanation for the increase in testosterone recently. It’s clear that levels have gone up in recent years, but it’s far from clear why. Personally, my top explanation—which I cannot directly support—is that there’s sampling error, and something broke in the COVID era that made it so that lots of survey results have changed in odd ways.

The Testosterone Validation Loop

People often seek somatic proof for social anxieties—they want something that validates their perceptions of trends in society. One major example is feminization.

Plenty of people believe that today’s men are softer than they used to be. And they might be, but that doesn’t mean testosterone is to blame. Nonetheless, to support the idea that men today are feminized, people will point to testosterone levels, allege they’ve fallen—and, due to common methodological issues, it’s easy to make it appear as though they have—and use that decline to support their belief that men are more feminine than they used to be. Moral, political, and aesthetic judgments have been validated by something measurable, natural, and scientific.

This is a powerful thing. Biology offers things social claims lack: objectivity, causality, and inevitability. ‘People seem less masculine these days’ is contestable. ‘Testosterone is collapsing’ is something that sounds measurable (and it is, in theory). ‘Modernity is making men biologically weaker’ sounds causal. ‘This is what bodies are actually doing’ sounds inevitable.

People perceive a social change, they search for a bodily marker of it, the marker becomes the proof for the harder to support social claims, the social perception is used to interpret the marker, and any disconfirmation is absorbed as denial, corruption, or hidden decline. Never mind that there’s missing data to support the notion that society has been feminized on the basis of observed testosterone levels, biological legitimation is just too intuitively plausible, so the arguments seem to work. But biology shouldn’t be used this way, because it does not work that way: a biomarker like testosterone does not necessarily co-move with masculine behaviors—look across cultures and it clearly doesn’t! Look to all the available supplementation experiments and it also fails there!

More importantly, in keeping with the focus of this article, since testosterone levels are actually up, what happens to the notion of feminization anyway?

May 20, 2026 Update: A New Meta-Analysis!

Rodríguez et al. conducted a meta-analysis on the relationship between testosterone levels and risk-taking behaviors in economic games across several fields including behavioral economics and neuroendocrinology. Their meta-analytic sample size ended up being 17,340 participants across 52 studies, and their overall result was what you might now expect if you read the rest of this article: nothing.

The meta-analytic effect estimate was r = -0.0021 (p = 0.92), basically zero. Moreover, there was no evidence of any sort of publication bias.

The type of risk measure was not a significant moderator of this relationship, and there were seven different measures, they just all produced similar results.4 Studies that measured testosterone from blood, fingers, and saliva all produced similar effect estimates, and those from administration found somewhat larger effect, around r = 0.1—a quite small effect size, all things considered. Sex also wasn’t a meaningful or significant moderator of differences in the effect of testosterone, so it’s not as though high or low testosterone mattered more or less for men or women. Excuse the colors:

Per the authors:

[T]estosterone is unlikely to provide a general explanation to human risk preferences. Rather than a single-hormone account, the evidence points to a biopsychosocial view: what we label “risk taking” emerges from the interaction of task demands, cognitive and affective processes, and situational context, with endocrine influences (when detectable) appearing narrow and contingent rather than broad and trait-like. […]

The most defensible position at present is modest: testosterone is not a universal driver of risk preference. Where associations surface, they are best understood as products of specific constellations of task demands and endocrine measurement rather than a single hormonal lever on decision making. Research programs that embrace multiplicity (i.e., neural computation, learning history, affective state, social context, and carefully specified endocrine pathways) offer a more credible route forward than continued searches for a unitary biochemical signature.

The intuitive understanding of testosterone as masculinizing is true in a physical sense; the evidence that it reliably behaviorally masculinizes people is tenuous.