AI is doing math

In the past few weeks, OpenAI announced 10 new (apparently impressive, according to mathematicians) results achieved by an internal model; Anthropic followed with a Claude-discovered Riemann-related result (not a proof, but a substantial improvement on the lower bound).

In light of these results, Fernando Borretti refutes some common arguments for why AI might not transform mathematics.

“Computers are superhuman chess players, yet we don’t care, and continue playing as normal. Why should mathematics be different? The main reason, I think, is that chess is self-contained: results from chess don’t help us understand the orbits of the planets or the binding of drugs to protein surfaces. But mathematics, famously, is the great dynamo of science, the best language and method for understanding the world. A machine that can replace a human mathematician, but better and faster and cheaper, is materially useful; a better chess engine is not.”

Noah Smith, while maintaining that the role of the mathematician will continue, argues that a specific form of heroic individual contribution is likely on its way out.

“We normally think of AI as something that replaces a single human worker, but I think that’s the wrong way to think of it. AI is really a new world-mind — a technology that takes the accumulated findings of individual humans (or robots, or other sensors) and integrates them into a general picture of the world… [E]ach time one of us reports a new scientific finding or expresses a new perspective, that adds to AI’s knowledge and understanding of the world… But unlike human society, AI doesn’t need heroic human geniuses. Its ability to understand complex ideas is not limited by the ability of a single human brain to apprehend, intuit, or communicate those ideas.”

Paata Ivanisvili recounts the pace of mathematical discovery from the front lines.

A small episode from the AI–math battlefield:

Aug 4: a paper appears on arXiv https://t.co/eJyFl3cfAG claiming partial progress on the Bourgain–Brezis Sobolev conjecture.

Same day: I ask AI, “can you try solving the full conjecture?”

Aug 6: AI says, “done.” (The proof attempt and all iterations are in my Overleaf history.) I start asking experts if they would like to check the argument.

Aug 8: another paper appears on arXiv https://t.co/stCwYd7ik7 claiming a full resolution of the conjecture (also with help from AI).

The pace is getting wild.

Tim Gowers explores the question of what aspects of mathematics, specifically, LLMs are good at, and the implications of potential answers.

“If it is true that current models are particularly good at finding examples, that is probably not because they have a particular affinity for existential statements, but more because the proof-discovery methods that are appropriate for finding certain kinds of examples play to the obvious strengths of LLMs: wide knowledge and the ability to explore many paths of the search tree that humans would judge to have a low probability of success.”

Terry Tao gives a lecture on mathematics in the age of AI, and the need for the field to take a more active role in examining its own role.

“What are the precise goals, objectives, and values of our mathematical community, and the enterprise of mathematical research? Not just the explicit goals that we communicate to the public (or to funding agencies), but also the implicit goals that we actually seek in practice?”

And Iris Shi writes a beautiful, difficult-to-summarize-but-very-worth reading meditation on the present and future of math as her PhD comes to an end.

“My physicist friend once asked me what the point of doing research was if someone like Terence Tao could have figured out everything in my dissertation in a tenth of the time. I answered by pointing out that Terence Tao didn’t. Terence Tao did not find a small open problem posited by my advisor and publish a bite sized result making incremental progress. He has only so much time and so many other fish to fry. And in his absence, I was given the chance to touch the edge of knowing and experience the unmatched exhilaration of discovering structure in the labyrinth of everything.”

Is AI doing other science?

There’s also been a flurry of activity in AI for non-math-science, with the first AI-designed virus, breakthroughs in cyclone forecasting,1 a shake-up at DeepMind, and a new Jeff Dean-led AI for science company.

To figure out what we can actually say about AI’s capabilities in drug discovery, an all-star team takes a careful look at the evidence and finds it wanting (if promising).

“The focus of AI in drug discovery must shift from doing what can be done - such as modelling data that is readily available, but that is unlikely to move the needle - to doing what should be done, even if this requires, for example, substantial data generation.”

“‘AI in drug discovery’ can mean two very different things: advanced, rapidly developed methods that lack thorough validation, as well as mature, established methods that have only just become practical to use at scale in production environments. What is common to both situations though is that the translation from ‘AI’ to ‘drug discovery’ has not yet been observed in an efficient and sufficiently durable manner to lead to clinical impact.”

Noah Olsman posits that many AI-for-bio efforts implicitly view biology as chess, when it is actually more like Kriegspiel.

“There is a variant of chess, called Kriegspiel, where each player can see only their own pieces. The game is played with three boards: two players sit back-to-back and move their own pieces while a referee keeps track of both players’ movements and mirrors them on a third board. The players go back and forth taking turns and, each time, the referee tells them whether that move is legal, and if they took a piece. The identity of the taken piece is only revealed if it’s a pawn.”

“Recently I have come to think that the differences between chess and Kriegspiel serve as a good metaphor for the gap between theory and practice in the life sciences. I entered biological research as a theorist who believed that enough mathematics could turn study of life into a game of chess: difficult, but ultimately legible. After spending years as an experimentalist, though, I have come to think that biology is much closer to Kriegspiel. We make moves (experiments) with only partial information, infer the hidden state of the board from sparse feedback (data), and often learn what mattered only after the consequences have unfolded, sometimes weeks or months later.”

Meanwhile Carlos Outeiral lays out the case that the gap in performance between AlphaFold 2 and AlphaFold 3 is because both rest on modeling evolution rather than physics, which works better for shape (AF2) than binding (AF3).

“The thesis of this essay is that the root cause of the slowdown is reliance on coevolution. The fundamental breakthrough in AlphaFold 2 was its ability to mine the evolutionary history of a protein to rapidly explore conformational space. Within a single protein, residues that touch in the folded structure (“contacts”) are constrained to evolve in concert… So if you align a protein’s sequence against millions of its relatives in genomic databases, these correlated changes capture a faint statistical shadow, left by evolution, of which residues sit close in space. … Put another way: the model appears to have learned the physics of biomolecular interactions, when it has largely learned protein genealogy.”

A team at DeepMind discusses both the promise of AI agents for speeding up scientific ideation, and the yet-unsolved gap in speeding up the validation of promising ideas.

“Most scientists have no shortage of ideas. The challenge is knowing which ones to pursue. This is where agents are starting to help. An agent can digest a field’s accessible literature, making connections across disciplines that no single researcher would have time to trace.”

“At the moment, however, the validation gap in most disciplines is widening, not closing. An agent can propose a novel genetic lead to reverse cellular ageing, but cannot say definitively whether it actually works. This explains why companies like Google DeepMind, Ginkgo Bioworks and Lila Sciences are investing in automated labs. But they only suit some fields, are expensive to build and are still early in development. And even automation cannot rush nature’s clock.”

Daphne Koller breaks down the misalignment between where AI-for-drug-discovery efforts focus and where the actual bottlenecks in drug discovery are:

“AI will certainly generate new and better molecules at an unprecedented rate, but will it generate drugs that unlock diseases for which there is currently no meaningful treatment? … [T]he biggest step functions in our ability to drug the undruggable have historically come not from better molecular design tools, but from expanding our repertoire of therapeutic modalities: first biologics, then siRNA and antisense oligonucleotides, then gene editing. Each new modality opened a class of targets that was simply inaccessible before.”

And the CRUX team, in a study led by Peter Kirgis, Sayash Kapoor, and Arvind Narayanan, look beyond AI agents’ ability to make progress on clean, verifiable tasks, and test whether they could conduct open-ended, independent research. For now, no.2

“Our key finding was that while agents could solve the engineering problems necessary to do the research, they failed to produce original research at the caliber of a top ML conference. Our main takeaways:

1. The agents lacked the judgment to identify when a problem was adequately solved….

2. The agents lacked awareness about the resources available to them and the timeline for the project…

3. The agents did not creatively respond to feedback about poor research design…

4. The agents did not effectively backtrack from unpromising approaches….

5. The agents did not follow concrete instructions....”

AI is reviewing (and submitting) science

Some aspects of the validation gap (at least in computational or theoretical fields) may be taken on by AI review. This possibility has caused two stirs on twitter in the past week or so, one when Refine announced its partnership with AEA and Econometric Society (discussed here), and another today based on this post from Itai Yanai.

Efforts to evaluate AI review tools often face the issue of mediocre ground truth (i.e. did a human or an LLM think the review was good overall? Did they think the individual comments were substantive?). Paul Litvak instead inserts errors and checks if tools can spot them (spoiler: they seem to do better than humans).

“Most systems work by ensembling, having the LLMs judge each other’s output. That, as we’ll see, is a good strategy, since different LLMs are surprisingly uncorrelated as far as the errors they can catch. Nonetheless, this isn’t satisfying since we don’t know whether even pooled models miss some categories of errors. Using human peer review as the gold standard, as any professor will tell you, is also unsatisfying, since human peer review varies widely in quality. Of Claude’s suggestions, one stood out — inserting errors into papers and seeing whether a given AI system can spot them.”

Beyond reviewing papers, advances in AI-based reproduction and replication of data+code continue apace, this time with a discipline-agnostic tool from Liu, Tjiaranata, and Tan.

“We present VERITAS, the first end-to-end domain-agnostic replication framework built around CLI coding agents. VERITAS supports flexible input modes, extracts the paper’s claims independently of the original authors, and withholds those claims from the replicating agent to mitigate leakage. The pipeline spans claim extraction, replication, and verification, and aggregates the per-claim verdicts into an importance-weighted Replication Score for the paper.”

But what about replication of findings in new data? From an ambitious new benchmarking effort:

“Several benchmarks have been proposed to develop and evaluate LLM agents on research reproduction tasks, including CORE-Bench, PaperBench, and REPRO-Bench… Existing benchmarks operate under the assumption that the new data sample is readily available to the agent. There are no benchmarks designed to assess agents’ capability to replicate research claims in a setting where a new data sample must be retrieved in advance... [W]e introduce ReplicatorBench, a benchmark for replicating published research claims in social and behavioral sciences.”

Let’s close on a would you rather: (a) have an LLM review your conference submissions, or (b) review dozens of LLM-authored conference submissions? I don’t know for sure which they’d pick, but I can say that Caleb Robinson and Isaac Corley did not enjoy the conference review circuit this summer.

“Between the two of us, we reviewed 22 paper submissions this summer, spread across NeurIPS, WACV, and TerraBytes (a geospatial workshop at ECCV). Fifteen of the 22 (68%) contained entirely fabricated citations, fabricated author lists for existing papers, and/or were clearly LLM-generated (e.g. hallucinated technical jargon, nonsensical writing, irrelevant citations). This is called being in the ‘slop trenches.’”

1. The Growth Effects of Energy Shortages - Willow Latham-Proenca

2. America Makes The Wrong Steel For Its Transformer Supply Chain — Dylan Matthews

3. Building New Homes - Alex Armlovich

4. GDP makes you better off, but it’s not the meaning of life - Matt Clancy

Nisha and Saloni are off this week. There will be more from Jordan later this week!

The Growth Effects of Energy Shortages - Willow Latham-Proenca

Energy-rich countries may not be all alike, but energy-constrained countries are also often each unhappy in their own way. Taiwo Hassan Odugbemi, Meron Tesfaye and the Energy for Growth Hub break down the monolith of unreliable power supply in Africa, showing how the grid can struggle in a variety of ways with distinct economic consequences.

Nigeria is the poster child for grid dysfunction, with chronic outages leaving the grid so unreliable that self-generation (usually with diesel generators) exceeds the total capacity of Nigeria’s grid. Transmission is a big part of the problem, likely a main culprit in keeping nearly 70% of the grid’s theoretical capacity offline. Beyond the direct cost of outages, the cost of buying and maintaining expensive backup generators increases firms’ production costs (potentially to 4x that of industrialized peers), and probably keeps firms that can’t afford the cost out of the market. As the authors flag, the “missing growth” cost here is hard to estimate (although apparently easy to invent?), but almost certainly in the billions.

In Ghana, in addition to outages and self-generation, long-dated power contracts rushed through in response to an energy crisis continue to undermine utilities’ financial stability. South Africa illustrates the continuing economic drag of past energy outages, even after (partial) recovery, with an estimated 1.5% GDP loss in 2023 yet to be recovered. Egypt’s case is perhaps the most interesting analog for the US right now – (barely) sufficient grid capacity and limited self-generation for the moment, but an uncertain grid expansion path when staring down the barrel of rapid demand growth.

America Makes The Wrong Steel For Its Transformer Supply Chain — Dylan Matthews

One of my favorite subgenres of business journalism / analysis are articles deeply exploring one particular, hyper-specialized product or resource that turns out to be strangely crucial to the whole economy. A great example is this piece in The Ecomodernist by Ryan Alimento on the production of “grain-oriented electrical steel” (GOES).

A specific alloy of iron and silicon, GOES is used to make cores for transformers. Transformers, an essential component in the grid that allow voltage to be adjusted for different types of transmission, are in extreme shortage, with order waiting times stretching to three years or more as the recent surge in electrical demand in the US forces more and more transmission buildout.

The US only has one steel plant making GOES, however, and its GOES is generally lower-quality and produces less efficient transformer cores than the GOES from Japanese and Korean mills. To make matters worse, the Trump administration briefly imposed steep tariffs on foreign GOES last year until outcry by the power sector forced it to reverse course.

Alimento gives a great breakdown of the problem, and an appealing suggestion for solving it: rather than rely on protectionism and import barriers, offer financing to enable US steel firms to develop new GOES-producing plants. He makes a compelling case that this is one specific area where industrial policy might be highly effective.

Building New Homes - Alex Armlovich

Brian Potter’s reading list has a good housing subsection this week:

First, there’s a striking divergence in existing home price data versus transaction prices at the largest homebuilders: Lennar has cut average selling prices 25% since 2022, overall median sale prices on FRED are down, but the Case-Shiller repeat-sale index for existing homes hasn’t followed. I see one unintended implication of this: Even the largest homebuilders can rapidly pivot to cheaper starter homes wherever they’re in demand, as soon as such homes are legalized in a jurisdiction.

Second, a new CalMatters piece documents that California cities have approved their Regional Housing Needs Allocation (RHNA) housing plans but are building almost nowhere near their targets. “Fair Share”-style housing frameworks like RHNA place a massive implementation burden on localities to proactively plan for growth from scratch, leaving the details wholly under local control. Decades of zero or weak results in these frameworks are beginning to draw more expert skepticism. State laws (like SB79) that prescriptively modify local zoning by default, subject to a local equivalent plan if a jurisdiction has staff capacity to make such a plan, increasingly look like a promising way to bypass the pain, paperwork, and massive staff time losses to RHNA-like processes.

GDP makes you better off, but it’s not the meaning of life - Matt Clancy

GDP per capita is positively correlated with a lot of desirable things, like health and subjective well-being (even though self-reported well-being in the United States has been falling for the last decade, it remains roughly at the level one would expect given our GDP per capita!). But GDP is not everything - it turns out reports of meaning in life actually decline with GDP per capita! Is this because “The meaning of life is a problem for people who are not desperate, people who can count on survival, comfort, security, and some measure of pleasure”, as social psychologist Roy Baumeister once wondered?1

Maybe not. A 2014 paper by Oishi and Diener argues this effect is actually driven by religiosity. Religiosity is positively correlated with meaning; it’s just that religiosity is also negatively correlated with the GDP per capita. If you try to predict meaning with two variables - religiosity and GDP per capita, religiosity carries the day and GDP per capita no longer carries any additional predictive value. Indeed, when the data was collected, the US was actually an unusually religious high income country, and also a leader among high income countries in terms of how meaningful its residents find their lives. US religiosity has dropped substantially since 2007 though, and so I suspect a repeat of the study may not find the same results.

Here are a few other highlights and announcements from our team and grantees:

• Last Friday, Alex wrote about the vast potential for transit-oriented development to end America’s housing problems.

• You can now watch Saloni Dattani’s TED talk!

• California YIMBY’s newsletter, The Homework, gave an update on the legislative session in all things California housing.

• Renaissance Philanthropy has opened applications for the Biosecurity version of its Big if True Science Accelerator

Watch the TED talk

I’ve also adapted it into a longer, companion article for Works in Progress.

I think we’re at a point of divergence when it comes to medical innovation. Our tools are improving rapidly, but our ambitions to use them have fallen behind. Science funding has slowed down, aid budgets have been slashed, and immigration restrictions mean we’re driving away expertise. Even when things have come back online, the disruption and loss of expertise will make it hard to return to our previous trajectory. We could be doing so much more. Medical innovation might not seem urgent in the abstract, but most of us know someone whose life it could change, and eventually I think it affects us all. We can change that.

That’s part of the motivation for my work on clinical trial reform on the Abundance and Growth Fund, but the talk is also larger than that: it’s a call to arms for more people to work on medical innovation, and explains why you don’t have to be a scientist to do it.

I hope you enjoy it!

Reading list

The talk covers lots of topics I’ve previously written about or covered in podcast episodes. Here’s a list if you’d like to read more:

• The Golden Age of vaccine development The first vaccine was a lucky accident. Now we can design new vaccines in weeks, atom by atom. [Works in Progress]

• Death rates from cardiovascular disease have fallen dramatically Over a century of progress in surgery, drugs, prevention, and emergency response has driven down death rates from heart disease and stroke. [Our World in Data]

• Childhood leukemia: how a deadly cancer became treatable Before the 1970s, most children affected by leukemia would quickly die from it. Now, most children in rich countries are cured. [Our World in Data]

• Why we didn’t get a malaria vaccine sooner (written with Rachel Glennerster & Siddhartha Haria) Hundreds of thousands of people die from malaria each year, but it took 141 years to develop a vaccine for it. [Works in Progress]

• Lenacapavir: the miracle drug that could end AIDS (episode co-hosted with Jacob Trefethen) Saloni and Jacob talk about the development and prospects for this new drug, as well as the history of HIV, the initial discovery of retroviruses, and how HIV was transformed from a death sentence to a manageable condition. [Hard Drugs]

• Should everyone be taking statins? (episode co-hosted with Jacob Trefethen) Heart disease is the leading cause of death worldwide, but it’s also one of medicine’s biggest success stories. [Hard Drugs]

• Medical breakthroughs in 2025 An end-of-year round up of the biggest news in biotech and medicine that I wrote last year. [Scientific Discovery]

Robot cars are coming

Waymo last reported a service area now growing to more than 1,400 square miles across 11 cities. We have finally reached the point where “self-driving cars are five years away” is no longer a joke.

To be clear, this is still good on balance. Robot drivers that never drive drunk, respect cyclists, and crash less often than a human driver would be a major public-health improvement. Not having to park is a real urban amenity, especially given the difficult political economy of charging market prices for curb access. Beyond the safety benefits, autonomous vehicles (AVs) could, with sound policy choices like congestion pricing in urban cores and High-Occupancy/Toll (HOT) lanes on highways, complement transit for trips and riders that fixed-route service handles poorly.

But AVs will probably also compete with transit. Ride-hailing offered a premium-priced but convenient door-to-door substitute for discretionary bus trips during the 2010s, while rising incomes and increased household car ownership diverted even more transit ridership in some markets. (Uber became the focal point for the wider phenomenon of urban cars outcompeting urban mass transit during the 2010s, even though it’s surprisingly hard to prove to a peer-reviewable standard that Uber has actually been an important culprit in the transit decline.)

Figure: National Transit Database ridership file rendered by system and mode. 80 out of 115 US rail systems in the NTD file peaked in Unlinked Passenger Trips in 2018 or earlier, long before COVID.

Now, a transit expert might balk that this concern about competition with mass transit concedes too much to both existing ride-hailing firms and to future autonomous vehicles. The first objection is empirical: the stagnation of American mass transit systems in the 2010s is not mostly attributable to ride-hailing alone.4

The second objection is conceptual and longer-term: the densest urban cores cannot realistically be served by single-occupant vehicles. As the well-worn “vehicle geometry” problem frames it, cities just cannot squeeze enough single occupant vehicles on the roads for the volume of people moving through the highest-density urban cores. But in-motion vehicle geometry is not the only problem; there is also a bottleneck in the mechanics of loading and unloading. If you’ve been to a concert or sporting event in San Francisco, you’ll note the vast pool of Waymos trapped in clumsy, low-throughput pickup and dropoff lines while people fumble with the doors and slowly clamber inside. High-throughput mass transit stations with level boarding and multiple wide large-vehicle doors are necessary at the human geometry level to be able to embark and disembark quickly on limited urban core curb frontage.5

But AVs do not need to eliminate rail to create a problem. They only need to skim enough trips to weaken fare revenue and political support, producing service cuts that drive away still more riders — driving yet another wave of higher taxes for worse service. In potential cases where voters balk at higher taxes for worse service, urban rail systems will collapse on a scale unseen since the 1970s, unleashing congestion and serious particulate matter pollution even if the cars are all-electric.6

A physically realistic benchmark goal for national TOD: 40 homes per gross acre

Density alone is not the only predictor of travel behavior: A dense neighborhood with no sidewalks and zero transit service definitionally cannot have any transit ridership.7

Academic caveats aside, residential density is an important and nonlinear predictor of transit ridership. The Transit Capacity and Quality of Service Manual, a policy guidebook published by America’s national academy for transportation science, advises transit and land use planners that the ridership returns to density near transit are steeply increasing with rising density. See the figure below, a copy of Exhibit 3-9 in the Manual:

Figure: TCQSM nonlinear relationship between density and transit demand

The national manual for transit service policy and planning is not omniscient, but it does propose a convenient and long-accepted national density threshold at which transit demand inflects to its most transit-friendly level: 40 dwelling units per gross acre. For federal policymaking purposes, a person living in an urban-density neighborhood above 39 units per acre is more than 41.5x more likely to use mass transit than a person living in an exurban, roughly half-acre zoning neighborhood at 2.35 units per acre.8

This 40 units per gross acre density threshold will form the analytical backbone of our TOD scoping and benchmarking exercise. There is no single correct answer about how many transit trips any given new apartment will produce, because service frequency, mode travel time competitiveness, tolling and parking policy, and other variables can also affect mode choices. But a reasonable lower bound estimate of this hypothetically built-out housing exercise still yields enough new transit trips to fully reverse the national post-COVID transit ridership losses (see below).

A deliberately mechanical benchmark

We’ve established that transit-oriented development matters for saving America’s rail systems from a nasty mix of continued operational decline, lower social returns on higher transit subsidies, and some lower-probability tail risks of tax revolts and total transit system collapse. But is the TOD opportunity space big enough to matter for the housing crisis?

If America is “short” somewhere between 4 million homes and 20 million metropolitan homes, the unused growth potential on the land near our rail systems has to be in the millions to be relevant. And it can’t just be any old land near any old train; it has to be where people are voting with their wallets to live: high-wage, high-amenity places where housing prices are above marginal housing supply costs (meaning that growth-control regulations are already the only thing stopping more people from building and living there).

The short answer is yes: physically and financially realistic TOD housing potential is a big deal, and is underrated as a reform opportunity on a surprisingly massive national scale. To be sure, one should treat this particular rail analysis as an order of magnitude rather than an exact count. But America’s bus networks are much vaster than our rail networks, and cover many more cities. Beyond this rail exercise, TOD in reality includes opportunities around buses and all the neighborhoods within walking and biking (or other micromobility) distance of downtowns. Without including those opportunities, this rail-only figure still yields strong-market capacity to build homes on a scale that is large relative to America’s short-term housing needs.

View Live Analysis & Maps

Finally, I want to push back on any hyper-literal or “overly engineering-brained” interpretations of these rule-of-thumb forecasts and planning heuristics for complex social and urban economic systems with endogenous or even self-regulating feedback loops that often do not behave like linear engineering problems. The elevation of rigidly simple traffic engineering and growth control planning rules of thumb, prioritized over social science and observable preferences revealed in market prices and behavior, created the very transportation and housing bottlenecks this analysis hopes to help solve. The point of this exercise is just to show that the mix of physically realistic growth opportunities currently banned by land use regulations, combined with revealed consumer preferences for transit-adjacent living observable in existing rents and home prices, together yield a reasonable expectation that relaxing the worst forms of central planning near our transit systems would plausibly yield voluntary actions on a large enough scale that is likely to solve our country’s most pressing housing undersupply and transit ridership sustainability problems.

Desirability and first-cut feasibility: Filtering physically realistic capacity from 31 million to financially plausible capacity for 13.5 million homes

America’s Census block-level capacity for allowing up to 40 homes per gross Census acre within a half mile of America’s rail systems is 31.3 million additional homes.9 This exercise thus answers one narrow question: How many more homes would these station areas contain at a half-mile-area average of 40 homes per Census land acre? First, some caveats:

• This is an illustrative gross-density gap representing the physical realism of development potential (it is not zoned capacity or a production forecast).

• A straight-line half mile is not always a half-mile walk.

• Some land should not or cannot be residentially developed.

• The calculation assigns the same benchmark to tidy residential blocks and to blocks dominated by office buildings or other nonresidential uses.

  • However: I excluded all Census blocks that now contain zero housing units or population, eliminating almost a quarter of the housing potential from this forecast to substantially account for airports, parks, and heavy industrial areas (see methodology appendix).

Even with those caveats, 31 million usefully establishes the physically realistic scale. It is larger than any conventional estimate of the national housing shortage. And remember, it’s still just a rail-only exercise: America’s bus networks cover much more land area than our urban rail networks, and high quality car-competitive bus networks can be scaled up in less than a decade.

The next question is where additional market-rate construction is at least plausible, financially. In much of America, especially outside high-wage metro areas, existing homes sell for less than it costs to replace them with new apartments. Upzoning those places is harmless, and could facilitate sparks of subsidized community development, but it will not unlock much construction in the current market.

I divided the station-area blocks into broad market tiers tied to my own rough estimates of inflection points in housing construction costs and first-cut market feasibility. These are binned using American Community Survey (ACS) median home values, with ACS contract rent as a fallback for blocks that didn’t have home value data.10 The “cool” category ($150k to $300k) approximates the roughly $250,000 “minimum profitable production cost” per single family home in Glaeser and Gyourko (2018).11 For the next bin, the “hard” construction cost to stack an extra 1,000 square foot home onto a Type I high-rise tower in NYC, America’s highest-hard-cost construction market, averages roughly $534,000, and San Francisco is the only other US city over $500,000, making $500,000 a relevant rough-cut threshold for the leading edge of high-rise multifamily construction feasibility. $800,000 is the threshold where even a $500,000 marginal hard cost of multifamily construction, with a soft cost multiplier around 30%, could together yield a ~20% gross margin to the builder on those hard and soft total marginal development costs.12 In other words, $800,000 per home is a plausible threshold for where incremental units added to (even high-rise Type I) skyscrapers in very difficult construction environments, at heights triggering all of the most hostile building code mandates beyond 7 stories, should be widely feasible if allowed cleanly by-right.

I label two tiers as high-demand:

• Hot: median home values from $500,000 to $800,000, or monthly rents from $2,360 to $3,780. These blocks account for 7.8 million homes in the gross-density gap.

• Red-hot: median values above $800,000, or rents above $3,780. These blocks account for another 5.7 million.

Together, the hot and red-hot blocks account for physically realistic capacity for 13.5 million additional homes under the 40-per-gross-acre benchmark in the highest-demand station areas where home prices are very likely to be above minimum profitable construction costs. This price screen still does not directly convert theoretical headroom into truly “market-feasible capacity.” I have complained at some length about planners making related mistakes. A serious soft-site analysis also needs parcel characteristics, the value and condition of existing structures, land assembly, local construction costs, financing, taxes, fees, and the actual density increment allowed by zoning. My screen tells us where demand conditions indicate zoning is most likely to bind. It does not identify the specific buildings likely to be demolished next year.

To be sure, one would also need to zone much more liberally than a 40 units-per-acre cap to achieve a realized production outcome averaging 40 units per acre within the next few business cycles. SB79’s zoning for 120 units per acre on high-quality rail sites (allowing seven-story buildings) is a good way to think about what should be allowed near rail, even in places where only the cheapest IRC rowhouses or three-story IBC Type V-B walkup apartments will pencil in weaker land markets.

Figure: State-by-state counts of current housing units and physically feasible capacity. Try other cuts of the data for yourself, and view every covered Census block on this live map.

What 40 units per gross acre actually looks like

“Forty units per acre” may sound more visually dramatic than it often is because density gets discussed through the most noticeable building rather than the land average around it. New York City offers a useful example. Among New York’s Census blocks near rail, I find the median block has about 29 homes per acre, and 62% of the blocks fall below 40. Yet the median home near rail sits on a block with about 63 homes per acre.

Most blocks are relatively low-density, but most people live in the smaller number of blocks containing apartment buildings.13 In many blocks, reaching the benchmark could mean one small or medium apartment building: a 24-home single-stair walkup, perhaps, or a 100-home mid-rise on a larger parcel.

Figure: Forty gross units per acre doesn’t entail Manhattanization on average, just rowhouses and the occasional apartment building. Even in New York City, 92% of residential structures are still three stories or fewer.

The largest gross-density gaps appear in the New York, Chicago, Philadelphia, Boston, Los Angeles, and Washington metropolitan areas. Commuter rail looms especially large: The station areas of Metra, SEPTA, NJ Transit, and the MBTA each contain more than two million homes of “gap” at the 40-per-gross-acre benchmark. Much of the opportunity is not in downtown skyscraper districts. It is along mature rail corridors where the public sector owns the expensive infrastructure and local zoning still treats an ordinary apartment building as an exotic industrial use.

A benchmark ridership scenario:

Having warned above about the dangers of taking mechanical “traffic engineering-brained” forecasts too seriously as a tool to justify coercively mandating or banning innocuous choices in housing and transportation policy, let’s indulge in one last mechanical forecast to estimate a lower bound of how many riders our 40 units per acre gross building envelope would yield.14

Figure: Calculated annual unlinked passenger trips from our land use exercise if new units are occupied at 2.2 people per home, with those people taking mass transit only as frequently as the average regional resident in earch rail system’s home region. 2024 has fewer trips per capita than 2019; calculations from both years are shown.

At 2024 transit ridership rates per person in each of the urbanized areas with rail systems, the nearly 6 million red-hot homes in our exercise would yield 1 billion new unlinked trips per year, increasing nationwide transit ridership by about 13 percent over current levels. Adding the nearly 8 million “hot” homes bumps that up to 30 percent. (There were roughly 9.9 billion unlinked transit trips on all modes in the US in 2019, and about 7.8 billion unlinked trips in 2024.)

The upshot: Physically and financially plausible rail-only TOD would be sufficient to completely reverse the national-level decline in transit ridership after COVID. If we expanded this analysis to high-demand blocks around frequent bus routes and neighborhoods walkable to job centers, the results would transcend mere pandemic recovery, and venture into outright reversal of America’s long-term declining relationship with mass transit. The benefits to economic growth, congestion, climate, land conservation, and the amenity and livability benefits of allowing most of America’s next 30 million homes to be built in the most extremely desirable neighborhoods in the country near mass transit (if people in the aggregate choose to do so voluntarily, without eminent domain) are immense.

Legalize a rider base

Will allowing homes near transit in high-demand places actually work? The closest thing American housing policy has to a natural experiment is Long Island City. LIC was rezoned to allow housing in 2001, with development taking off sharply after 2010 as LIC went from roughly 35,000 to 63,000 residents — 78% neighborhood-level growth in a decade, most of it in market-rate apartments a subway ride from Midtown. That is what clean transit-oriented upzoning does when the zoning constraint is actually released. Large reform case studies like NoMA in Washington, DC; Hudson Yards in Manhattan; Downtown Brooklyn; Gowanus, Brooklyn; and Austin’s University Neighborhood Overlay are all-star examples of rapid financially feasible growth commencing immediately after “clean” by-right upzoning.15 Multiply that fast-acting neighborhood level land use reform pattern across the hottest neighborhoods on America’s rail systems and we get 6 million potential red-hot homes.

Again, this is just a draft analysis looking at the development potential of rail systems only. Widening the analytical scope to alternative-transportation development opportunities beyond rail would include cheaply scalable bus rapid transit corridors, high-frequency local bus corridors, and any land areas within walking distance of downtowns but not yet served by a frequent fixed-route transit line. Someone should do that wider analysis but this rail-focused minimum viable product shows rail alone is big enough to matter. The rail analysis also conveniently centers the most above-cost large metro areas and the overpriced, undersupplied neighborhoods targeted by the Abundance and Growth fund (AGF)’s urban economics methodology.

This is not to say that the only place new housing should be legal is next to America’s trains: AGF supports broad upzoning to allow freedom of location choice. The goal of this analysis is only to demonstrate that TOD is not a marginal or boutique idea and should be a much bigger part of the near-term housing reform conversation.

Climate funders in particular should take note: Even if housing abundance is not an organization’s main goal, anyone who cares about climate has an opportunity to work with AGF grantees and the whole YIMBY ecosystem to prioritize land use reform wins on the scale of California’s SB79. That means cheaply decarbonizing transportation for millions of new homes simply by rescuing existing transit systems while reducing per-capita energy consumption all while increasing economic growth. American voters and volunteers are already solving the housing crisis with all-of-the-above solutions; climate funders now have the opportunity to make sure TOD is at the top of the housing reform conversation in every state with a big rail system.

America’s large rail systems are probably not at mortal risk from autonomous vehicles: The prewar urban core corridors of the biggest transit systems move too many simultaneous people for low-throughput, car-based vehicle geometries to fully shut down their transit systems. (Multiple parallel autonomous Bus Rapid Transit lines could combine enough corridor throughput to be able to replace most rail US systems, but autonomous cars alone cannot). Voters are more likely to vote for higher subsidies to keep the systems running even at lower utilization levels, unless we take this opportunity to solve both our housing and our transit problems. We already built the trains! Just let the riders live near them.16

Subscribe now

Methods appendix

To start, I overlaid 2020 Census blocks with the Bureau of Transportation Statistics’ National Transit Map. After consolidating duplicate platforms and station entrances, the data contain 5,672 rail stations and 152,105 populated or developed Census blocks whose centroids lie within a straight-line half mile of one. Those blocks already contain about 9.2 million homes.

Then I asked how many additional homes these blocks would contain if each reached 40 homes per full Census land acre. I chose 40 homes per gross acre because it is the floor of the top transit-supportive density tier in the TCQSM’s density-demand relationship, which is measured — conveniently for a Census-block analysis — in households per gross acre.17 I also excluded Census blocks that currently have zero households or residents in the Census data, reflecting the fact that many transit-adjacent Census blocks are airports, parks, heavy industry, or other places where residential redevelopment may be unlikely.18 Those excluded Census blocks are 2,041 water-only blocks covering 34,309 acres, and 46,026 completely vacant blocks with 267,406 acres of land. This is 24 percent of all Census blocks and 22% of all land near rail stations, enough to hold 10.7 million homes at 40 units per gross acre. This creates some risk of an underestimate: Converting unoccupied industrial land to residential land is the source of some of the nation’s most successful recent upzonings like Gowanus in NYC and NoMA in DC. But on the other hand, some of the land I did count includes commercial central business districts with small but nonzero resident populations, like Midtown Manhattan, where it’s not obviously likely that office conversions and new infill will reasonably be able to achieve 40 gross units per acre amid massive office towers in a land market where office usually remains the highest and best-yielding use. On balance, excluding 22% of all station-adjacent blocks for vacancy reasons gives me high confidence that this exercise is at least not an underestimate.

Finally, the gross versus net land area distinction is an important difference in what a “units per acre” number means. Forty homes on an acre of buildable residential land is approximately one attached home per 1,100 square feet. Forty homes per whole-block gross acre must fit the homes and the streets and other nonresidential uses on the same acre. Unless the street has become one of the homes, the latter requires smaller lots, apartment buildings, or a mix of the two. So 40 at the Census-block scale is more demanding than either 40 on residential parcels or Colorado’s 40-times-non-exempt-acre capacity goal.

Full replication files are available on request (and if there’s wider interest, we will edit this post and host them openly for public use, localization vetting, and extensions to further analysis suggested herein).

Predictably, there have been a lot of takes.

On the pro side, the arguments I find most compelling are: (a) Peer review has been buckling under its own weight for years, with editors even at top journals struggling to find reviewers. That problem gets much worse if AI drives a substantial increase in submissions, as it seems to be doing. Something’s got to give, and AI is likely to be part of the answer, so I would much rather it be via well-designed and publicly-validated tools than reviewers pasting manuscripts into a chat (note: we covered Refine’s own evaluation a few weeks back, and Paul Litvak has a new AI review eval that I will discuss more in a future post).

And (b) There are several subtasks that human reviewers have never been especially good at — like catching math or software errors, checking consistency between tables and text, and verifying that citations accurately represent cited work — but that we’ve fully relied on them for nonetheless. Offloading those subtasks could let reviewers concentrate on the tacit and taste-based components of review, like assessing the novelty of results, their impact on the field, how they fit (or don’t) with existing literature, etc. Peter Hull, Paul Novosad, and Kevin Bryan each proposed related versions of what the future of combined AI and human review could look like.

On the con side: (a) At equilibrium, there will need to be ways to avoid creating capture for for-profit companies – i.e. if journals pay subscription fees for AI review on one end, that heavily incentivizes authors planning to submit to those journals to pay for those services up front as well. That is less of an issue for this pilot, where Refine’s use does not affect publication decisions and is being offered to journals for free, but if it ever shifts upstream then this concern would need to be resolved.

And (b) There is also a possibility that offloading technical checks could make human reviewers less careful/thorough, which could reduce overall review quality on net. This is something Refine and the journals could formally study, and I hope they do.

Overall, props to Ben Golub for his work on Refine, and to the AEA and the Econometric Society – two of the publishers who have also been on the front lines of replication packages and data checks – for being forward-looking and experimental about research evaluation.

1. China tightens the reins on clinical trials — Dylan Matthews

2. A keyhole view of the next pandemic — Saloni Dattani

3. Europe’s scaling problem — Matt Clancy

4. PJM’s large load experiment — Willow Latham-Proenca

5. ROAD Act implementation comes to the states — Alex Armlovich

There will be more from Jordan later this week!

China tightens the reins on clinical trials — Dylan Matthews

One of the big biotech stories of the 2020s has been the rise of China. A recent survey of industry leaders found broad agreement that the US now lags behind China in clinical development of new drugs, and two years ago China oversaw more clinical trials than the US for the first time. This has led a lot of observers, like Congress’s National Security Commission on Emerging Biotechnology, to urge emulation of some of China’s regulatory approaches to drug development, like letting investigators initiate phase I trials without needing regulator approval first.

But Stat’s Brian Yang reports that China may be moving in the opposite direction, toward stricter regulation. Essentially, two deaths in gene therapy trials, and the bungled attempt by their investigators to cover up the deaths, pushed the Chinese government last October to advance Order 818. The order represents an across-the-board tightening of Chinese trial regulation. Now, investigator-initiated trials are restricted to the highest-ranked subset of hospitals, limited to certain types of drug designs, and required to meet “good manufacturing practices” standards akin to those the FDA requires of commercially available drugs in the US.

Yang finds that the rules may already be slowing down the famously fast-paced Chinese biotech center: “Since May, the China National Center for Biotechnology Development, an organization under the National Health Commission, has approved 80 or so [investigator-initiated] projects, but a STAT tally showed nearly all of them were retroactive registrations of work that had already been underway, and all of them are testing cell therapies. Only one is genuinely new.”

A keyhole view of the next pandemic — Saloni Dattani

I’ve been following the Ebola outbreak in central Africa, which has just surpassed 2,000 reported deaths and 4,381 confirmed cases. (The real numbers are likely higher due to underdetection.) What’s surprising about the Ebola outbreak is that we actually have an Ebola vaccine, but it was developed and tested against the Ebola Zaire virus during the 2014–2016 outbreak, while this outbreak is caused by Ebola Bundibugyo virus. The two viruses have similarities and some laboratory and animal data suggested there may be cross-protection, although the evidence is still unclear.1 But until last week, the WHO hadn’t decided to test that vaccine in this outbreak; the focus was on developing and testing new Bundibugyo-specific vaccines and antivirals instead.

In a recent STAT opinion piece, Krutika Kuppalli and Placide Mbala argue that pandemic preparedness shouldn’t be so narrow. Rather than focusing so narrowly on each virus, we should invest in broad-spectrum diagnostics, vaccines, and treatments, and should have tested the effects of the Zaire vaccine on Bundibugyo earlier; I agree.

Interestingly, the Ebola Zaire vaccine came about through innovative trials and coordination. It was developed through funding by BARDA and the DoD, tested using ring vaccination trials led by the WHO, incentivized through an advance market commitment, and stockpiled by Gavi, WHO, UNICEF and the Red Cross for future outbreaks. That stockpile has been useful in the years since then to contain outbreaks of Ebola Zaire quickly. But now that a larger, harder to contain outbreak has arrived, we’ve lost expertise and funding to tackle it, and we don’t know if the existing Ebola vaccine protects against this virus too. We could be doing a lot better.

Europe’s scaling problem — Matt Clancy

In 2008, the total market capitalization of US-listed firms was about a third larger than the market capitalization of European-listed firms. In 2023, it’s about triple. What happened? A new NBER working paper by Becker, Benmelech, and Monteiro tries to answer that, first by ruling out some potential explanations: it’s not a few lucky superstar firms; it’s not driven by differences in the relative size of different sectors; it’s not about how many firms get listed; it’s not about successful European firms re-listing in the states; and so on.

Instead, they argue Europe’s smaller market cap is downstream of the fact that successful European firms struggle to scale. Part of the explanation here is about financing. While the difference in the leverage of European and American firms did not change much between 2008 and 2023, the venture capital supply gap widened by a substantial level over the same time period. Since VC money supports new high-growth firms, this difference shows up most acutely there. While the largest European and American firms face similar costs of capital, the terms are much less favorable for small firms in Europe than in America.

The other problem is European firms have a harder time tapping into continent-wide markets than American firms do. The size of European firms is more closely related to the size of the national market they are based in, than the size of an American company is to the size of the state-wide market it is based in. The greater integration of American markets means it doesn’t much matter whether a company is based out of Idaho or Texas for its ultimate size, but it does matter in Europe if you’re based in, say, Germany versus Denmark.

PJM’s large load experiment — Willow Latham-Proenca

PJM – the country’s largest regional transmission organization by population – recently filed plans for two big steps on large load integration. As covered by Maeve Allsup in Latitude, the plan is two-part. First, PJM will try to procure enough new generation to cover the gap left at its latest capacity auction (where PJM procures enough power to ensure a reliability margin to cover unexpected peaks or outages) through a generator-load matchmaking process and an auction for the remaining shortfall. Second, PJM will require data centers that don’t bring or purchase their own supply to accept “curtailment” requirements – essentially, cutting the power they draw from the grid during emergency conditions.

PJM sidestepped the limits of its own authority (and the political fallout) by delegating – states and utilities will be responsible for implementing and enforcing curtailment, as well as the final cost allocation for generation capacity procured through the auction.

This is explicitly not a long-term solution – first of all, the capacity procurement is targeting a single year’s 7 GW capacity auction gap, a fraction of the 50-60 GW shortfall PJM itself expects over the next decade (others have widely varying projections here). Second, “matchmaking” pairs generators with the wealthiest suitors (probably the data centers), crowding out access for other types of new load. A long-term version of this type of program could undercut the viability of the capacity market itself, making it harder for cost-sensitive electricity users (like industrial loads) to come online.

ROAD Act implementation comes to the states — Alex Armlovich

IFP and Pew have a fresh state and local guide to ROAD to Housing implementation, complementing the NAAHL federal guide featured in our July posts. Will Poff-Webster, Alex Horowitz, and coauthors provide at least one sentence on every key provision in the 300+ page ROAD Act that requires or rewards action by states, localities, and tribes. The jurisdictions that move first on zoning, building codes, and permitting will capture the new federal dollars, and the ones that don’t will eventually lose some CDBG funding under the Build Now Act’s conditionality mechanism. The most immediate action item is manufactured housing: states need to certify to HUD that their laws treat chassis and chassis-free homes equally, which means most will need to amend statutory definitions that currently specify a permanent steel chassis. The piece also flags preapproved building plans, single-stair grants, and the HOME program overhaul that finally replaces per-unit cost caps with area-adjusted limits. Pair this with the NAAHL guide for a full introduction to what has to happen next.

Here are a few other highlights and announcements from our team and grantees:

• Adam Kroetsch joined IFP to help advance their work on FDA policy and clinical trial abundance.

• Our Coefficient Giving colleague Rafa Latham-Proenca published The Staggering Welfare Consequences of Catch-Up Growth After AI, estimating how much faster AI-enabled growth could benefit developing countries that adopt AI later, and finding the welfare gains are enormous even for latecomers.

• Center for Building, The Kelsey, and Sightline Institute were all quoted in a Washington Post piece on why the US doesn't have the small, affordable elevators common in Europe and China, a building code barrier that limits both accessibility and density.

• Michigan adopted the four-story single-stair rule statewide, building on Center for Building’s ICC win earlier this year. Jurisdictions with their own fire departments can go up to six stories.

1. Can beauty solve the housing crisis? - Alex Armlovich

2. Can cheaper inputs solve the housing crisis? - Alex Armlovich

3. How much of the housing crisis can we blame on endangered species? — Dylan Matthews

4. A new kind of problematic regulation — Willow Latham-Proenca

5. Cutting the Future STEM Workforce — Matt Clancy

6. It doesn’t take two trials — Saloni Dattani

7. What we talk about when we talk about institutional design — Jordan Dworkin

Can beauty solve the housing crisis? - Alex Armlovich

Samuel Hughes has a new Works in Progress essay asking again whether beauty can solve the housing crisis. He distinguishes NIMBYs (locals who oppose development of all kinds nearby) from NITBYs (”Not In That Backyard”): people who live nowhere near a project but casually oppose it because it looks ugly to them. Hughes argues that NIMBYism predates modernist architecture by a generation (first-wave restrictive zoning swept Europe and the US while new buildings were still classically ornamented), so beauty alone won’t dissolve local opposition. But the historical conservation movement, which froze European city centers and large swaths of American downtowns starting in the 1960s, could plausibly be a NITBY response to modernism: thousands of years of demolition-and-rebuild stopped almost overnight once the replacements got “ugly”. For infill development, it appears true that sometime after WWII, American city residents lost faith in the future of what their cities would look and feel like.

But competing stylized facts remain: cinderblock strip malls and vinyl-sided “mcmansion” sprawl continues to be the central feature of the most rapidly-building metros in the US (a sharp difference from the UK’s strict century-old ban on “ribbon development” sprawl). Beautiful projects like I’On in Charleston and the Duchy of Cornwall’s Faversham still get NIMBY’d to a standstill, and Robert AM Stern’s traditionalist neoclassical limestone revivals brought favorable reviews but few permitting wins. Finally: beyond the uncontroversial option of preapproved “pattern books”, nobody has an obvious policy lever to make new buildings beautiful at scale.

Can cheaper inputs solve the housing crisis? - Alex Armlovich

One of Brian Potter’s latest Construction Physics installments asks whether cheaper labor or materials can meaningfully cut housing costs: Looking at national average data in the US, he mostly answers no. Construction can’t chase cheap labor the way garment manufacturing did because most work happens on-site; modular builders that ship from low-cost locations (Idaho, even China) typically claim schedule certainty, not hard-cost savings.

From AGF’s perspective, however, it remains a puzzle how HUD Code manufactured homes have achieved huge hard cost savings over site-built homes, while modular construction has not done so. (Potter, too, has cited >50% lower costs per square foot for HUD Code manufactured homes). Some of the cost advantage for traditional HUD Code homes is related to entry-level finishing and materials: more upscale HUD Code CrossMods, built to middle class tastes and finishes, are “only” 20% to 30% cheaper per square foot than site-built homes. If multifamily modular could match the 20% savings of CrossMods, that would be a meaningful improvement. It may be unwise to expect miracles from multifamily modular policy or innovation, particularly in the national average data that covers wide swathes of the US where construction costs are already low, regardless of any code reform or innovation. But I also don’t think it’s crazy to explore whether some mix of code reform and innovation could help bring, say, NYC’s current Type I high-rise hard costs down closer to Chicago’s current Type I high-rise hard costs.

How much of the housing crisis can we blame on endangered species? — Dylan Matthews

You may have heard about the effects of the Endangered Species Act on big construction projects — like the famous case of the tiny “snail darter” fish and its role in delaying the Tellico Dam in Tennessee. But what about the effects on more everyday construction?

Harvard econ grad student Maxwell Tabarrok has a nice Journal of Public Economics article (official, free PDF, blog summary) estimating how housing permits change in areas that see a nearby species listed as endangered under the law. Listing of animal species leads to a large drop in permitting rates (0.5 permits per thousand residents, or 10 percent of an average area’s permitting rate); de-listing a previously endangered species leads to a similar-scale increase in permitting.

Tabarrok’s methodology is to treat each listing/delisting as a policy change and then do a “difference in differences” comparison of permitting rates in places that have the species nearby (the treatment group) and those that don’t (control group). This is about the best you can do in the absence of an actual experiment or some source of truly random variation in species listing, though it leaves some room to question whether the Endangered Species Act is causing this slowdown in permitting.

Tabarrok’s accompanying blog post suggests that the Trump administration’s new interpretation of the ESA could greatly weaken its effects on housing construction. Specifically, the Trump team is junking the current interpretation that includes damage to habitats as a form of harm to endangered species, and embracing a new interpretation where the law only applies to literally, directly killing or injuring animals. Under the now-final rule, set to take effect in September unless courts block it, the ESA would go from a law that bans construction that might harm endangered species to one that bans hunting and poaching endangered species. (The exception is projects on federal land or with a “federal nexus,” like the famous Tellico Dam.)

I like housing construction, but conserving endangered species is also valuable. Tabarrok has nicely illustrated the costs of the law, and I’d love to hear more about the benefits and how to balance the two optimally.

A new kind of problematic regulation — Willow Latham-Proenca

A forthcoming paper lends some structure to the intuition around regulated utilities (these are the vertically-integrated kind, with a local monopoly and their own generation facilities – they’re “regulated” because a state Public Utility Commissions, rather than the market, sets their rates). Gowrisankaran, Langer, and Reguant look to explain why coal has hung on so much longer in regulated markets than deregulated ones during the start of the natural gas boom – 26% of coal capacity in deregulated states was retired between 2006-2018, versus 17% in regulated areas.

They find that the standards Public Utility Commissions use to set rates are likely part of the explanation – to be included in a utility’s rate base (the pool of capital investments on which a utility can earn a rate of return), generation facilities must be “used and useful.” This creates a perverse incentive for utilities to keep expensive legacy facilities operating even if a cheaper option (in this case, natural gas) becomes available.

Interestingly, emissions don’t really factor into the decision over the long term; the authors find the cost advantage alone would be enough for a cost-minimizing utility to retire practically all coal generation over a 30-year horizon. However, the authors make an important caveat on reliability – reducing the rate base too far, without other compensation, could hamstring utilities’ ability to invest enough in the infrastructure that keeps the lights on.

Cutting the Future STEM Workforce — Matt Clancy

Last week, the Wall Street Journal reported that the Trump administration is considering implementing a $100,000 fee on international students who want to work in the United States after graduation via the optional practical training visa, one of the most common pathways to immigration to the United States. This policy, if implemented, would be the second recent blow to US legal high skilled immigration: earlier this month, the administration changed visa duration rules to add new burdens and uncertainty for international students. Previously, common student visas lasted for the duration of a course of study. Now the visas last up to a maximum of 4 years, with extensions possibly but uncertain and requiring application.

(The trouble is that four years is too short for a PhD. I did mine in five years, which put me on the fast end of my Economics PhD cohort, and the typical STEM PhD takes between 5 and 6 years.)

Collectively, these policies (along with the heightened uncertainty around funding for research) are likely to reduce high skilled immigration from international students. A new report from Amy Nice, Michael Clemons, and Jeremy Neufeld tries to quantify the size of the impact, looking at survey evidence and the historical share of the STEM workforce that follows affected pathways. They estimate that, if these policies had been in place for the twenty years between 2003 and 2023, the US STEM PhD workforce would be 11.5% smaller than it is today. That would mean lower US economic growth (and probably no significant benefits to native-born STEM workers; more here).

It doesn’t take two trials — Saloni Dattani

In June, the FDA issued new draft guidance stating that drug makers no longer needed to run two trials to demonstrate the efficacy of a new drug before it could be approved; rather, “one adequate and well-controlled clinical investigation with confirmatory evidence” could satisfy the evidentiary standard. It was a controversial move, although others pushed back saying the two trial guidance isn’t commonly followed in the first place and was an unnecessary burden.

In a recent post, the biostatistician Kaspar Rufibach shared a characteristically thoughtful take. When drug developers want to maximize their chances of succeeding in the second trial, they often try to design it identically. But, for a drug that “just” showed significance in the first trial, its chance of succeeding in a second, identical trial is just a coin toss: 50%. Better, he suggests, would be to run a single, larger trial, or run a second trial that’s different and gathers evidence for the drug’s effect on another disease, in another region, or a different route of administration, for example.

What we talk about when we talk about institutional design — Jordan Dworkin

If you’ve been following the conversation around the design of scientific institutions and funding strategies, you may have come across a table breaking down how different characteristics map onto different mechanisms.

Last week Hiya Jain, Stuart Buck, and Aishwarya Khanduja published “A Taxonomy of R&D Orgs”, an excellent longform contribution to this tradition that distills the multidimensional differences between research organizations into nine key parameters: (1) optimal project timeline, (2) level of user input, (3) size of discretionary research budgets, (4) career incentives, attracting talent, and personnel models, (5) revenue strategy, (6) degree of focus on a single mission, (7) size of organization and/or research teams, (8) governance and project selection mechanisms, and (9) openness and IP strategy.

They go deep on each, exploring the range of possibilities and laying out case studies of institutions that sit at different points on the spectrum. It serves as both taxonomy and history (e.g. I did not know that staff scientist/engineer roles in part emerged as a result of early ARPA block grants to CMU, MIT, and Stanford), and will be a useful resource as the scientific ecosystem tests out a wide range of new (and old) models.

Here are a couple of highlights from our grantees and team:

• 1Day Sooner is still looking for a clinical trial abundance policy director

• Saloni and Jacob Trefethen released a new episode of Hard Drugs on drug repurposing: Are miracle drugs hiding in plain sight?

• Dylan Matthews wrote about our chance to fix the Jones Act

• A paper about the Atlas of Innovation was published as an NBER Working Paper

Here in the US, at least one 2026 policy change has been a clear winner for economic growth. Starting on March 17, 2026, the Trump administration has waived the Jones Act, America’s cabotage law barring non-US built vessels from conducting trade between US ports, for certain oil, gas, and fertilizer shipments. The waiver was extended in April; it’s currently set to expire on August 16.

This move shouldn’t have been overly surprising. Kevin Hassett, Trump’s Council of Economic Advisers director in his first term and National Economic Council head today, is a vocal Jones Act detractor, and wrote in a 2021 book that Trump hates the law too (“This Jones Act, it doesn’t make any sense,” he quotes the president saying). Casey Mulligan, another economist who served during Trump I and is now chief economist at the Department of Health and Human Services, reported the same in his memoir You’re Hired!: “President Trump hates the Jones Act. It is the type of harmful regulation that he has succeeded in ending in health insurance, telecommunications, farming, and many other industries.”

Of course, Trump World contains multitudes and other advisers (notably Peter Navarro, Trump’s “senior counselor for trade and manufacturing”) are vocal Jones Act defenders. Mulligan and Hassett, in their books, recall being defeated by Navarro and allies in Congress in term one when they fought for a Jones Act waiver for liquid natural gas. But that a waiver actually went through this year, and has continued to be renewed, suggests the law’s critics are winning the intra-administration fight this time around.

This makes 2026 perhaps the single most promising moment for meaningful Jones Act reform in decades. Durable reform, though, has to go through Congress, and will have to be bipartisan. It will also have to grapple seriously with what transitioning to a new regime looks like.

The case against cabotage

Defenders of the Jones Act like to call it “America’s oldest law,” noting that rules limiting which ships could use US ports are as old as the first Congress. But the “Jones Act” generally refers to section 27 of the Merchant Marine Act of 1920, a provision authored by a Seattle shipping lobbyist and added to the law by Washington Senator Wesley Jones. This is still the most important single statute regulating shipments between US ports.

As it exists today, the law requires that ships be…

• Captained by a US citizen, with all US citizen officers, and US citizens making up at least 75 percent of the rest of the crew

• If owned by a corporation, owned by a corporation with at least 75 percent US citizen ownership

• Assembled in the US, with the “major components” of their hull, main deck, and other “superstructure” all built in the US

Cabotage laws like this are common internationally, but they often only restrict trade to vessels flagged in the country in question, or owned by nationals of that country, or crewed by staff in that country. The Jones Act is unusual in requiring that ships moving between US ports be built in the US. Most other countries with large cabotage trade, like China, Japan, and India, don’t require domestically-built vessels, despite the former two having massive shipbuilding industries much more successful than that of the US. As a result, only the 92 Jones Act-compliant US-built ships currently in operation can move between US ports when the Act is in effect.

The economic case against the Jones Act is simple: there are many useful trips to be taken by ships within the United States; the Jones Act, plus the tiny size of the Jones-compliant fleet, means many trips are not taken when the Act is in operation; allowing those trips to take place represents new, beneficial economic activity that we’re currently blocking.

Perhaps the best examples of these Jones-prevented trips are the ones happening right now, under the waiver. The Cato Institute has been tracking waiver-enabled voyages, and there have been 195 to date as of this writing, on 161 different vessels, shipping some 50 million barrels of various fuels and fertilizers (most of the volume is gasoline and crude oil). Those 161 vessels represent a more-than-doubling of the US cabotage fleet, relative to the Jones Act’s 93:

As you can see from their maps, these journeys typically originate in the Gulf Coast, the center of US oil refining and export, and then make their way either to the northeast US, to Puerto Rico, or to the West Coast via the Panama Canal.

If you’re not especially interested in boosting oil shipments for climate reasons, don’t worry: the Jones Act is harming your priorities as well. A major effect of the Jones Act is to transfer goods from freight ships to freight trucks, which produce vastly more carbon emissions per ton-mile traveled than ships.

The Jones Act is also a major factor behind the stagnation of the US offshore wind industry. Installing a wind turbine in the ocean requires specialized ships called wind turbine installation vessels (WTIVs). Until March of this year, there were literally zero WTIVs built in the US. That meant that installation required sailing WTIVs from ports in Canada to as far as Virginia with turbine components, because foreign-built ships couldn’t be based in the US and do US port-to-port hops. As of March, there’s exactly one Jones Act-compliant WTIV, first installing turbines years behind schedule and far over budget, and we need at least four WTIVs to meet the Biden administration’s old offshore wind targets.

More formal modeling gives us some sense of the magnitude of the economic harms from the Jones Act. A 2019 paper from the OECD’s Karin Gourdon and Joaquim J. M. Guilhoto conducted a simulation estimating the economic costs of the Jones Act and an analogous domestic-content rule in Brazil. They suggest that total US economic output will grow by between 0.1 and 0.4 percent annually. That’s between $31.8 billion and $127.4 billion a year, per the most recent GDP numbers.

Here at Coefficient Giving, we often evaluate grant opportunities by estimating their impact in “Coefficient Giving dollars,” which we define as “the amount of welfare generated by a $1 gift to someone with an annual income of $50k,” roughly the US GDP per-capita when we started this framework. More precisely:

CG $v = $50k * w * ln(1+z%) * y.

Where v is the value of an intervention, w is the number of people affected, z is the percentage change in the income of the people affected, and y is the number of years they’re affected. The math is less important than the overall idea: we place a great deal of value on increasing people’s incomes, with longer-lasting changes affecting more people being worth more.

Jones Act repeal looks very, very good under this framework. Suppose we take the low-end figure from the OECD researchers (0.1 percent added to annual output); that’s z. Imagine for the sake of argument that we have a grant opportunity that has a 1 percent chance of leading to the US repealing the Jones Act ten years earlier (y = 10) than would otherwise happen. The US has 340 million people or so (w = 340 million). Do the math and you get about CG$170 billion in impact. Right now, our Global Health and Wellbeing teams have a 1000x bar for grants: they have to produce CG$1000 for every US$1 we spend. That implies we’d spend $1.7 million to generate a 1 percent chance that the Jones Act gets repealed ten years early, before even considering other factors like environmental benefits.

Obviously, no one can guarantee an advocacy campaign will even be that effective, and the nature of advocacy is that most people you fund have a low, sometimes under a percentage point odds of success. But when the value is this high, it’s still worth funding longshots. And with a Jones Act waiver currently in effect, serious reform looks like less of a longshot.

We need bipartisanship and creative thinking

The battle lines around the Jones Act are very familiar at this point. On one side are the US ship-building companies and allied unions, organized as the American Maritime Partnership, which push very hard against any weakening of the law. On the other side are anti-protectionist groups like Americans for Prosperity and the Cato Institute, the latter of which has done yeoman’s work documenting the law’s harms.

My sympathies are more with the latter, but the current ecosystem could use shaking up. For one thing, there are strong reasons for Democrats, both moderate and left-wing, to oppose the Jones Act. Behind its huge environmental costs, it particularly raises costs for consumers in Hawaii and Puerto Rico, which are entirely dependent on shipping for most goods but are forced to use the small Jones Act fleet for shipments from the US mainland.

As a result, there’s a reasonably large group of Democratic Jones Act critics in Congress, largely consisting of members with Puerto Rican ancestry. Alexandria Ocasio-Cortez (D-NY) is a vocal critic, having tweeted even before she entered Congress, “Congress must repeal the law that cements PR’s status as 2nd class citizens.” Nellie Pou (D-NJ) repeatedly urged repeal during her time in New Jersey’s state legislature. Ritchie Torres (D-NY) has proposed legislation exempting Hawaii, Alaska, Puerto Rico, and Guam from Jones Act requirements, just as the US Virgin Islands is currently exempt. While most of Hawaii’s Congressional delegation has sided with the shipbuilders over consumers and backed the Jones Act, Ed Case (D-HI) has been a welcome exception in supporting the waiver this year and urging permanent reform.

Conversely, there are Republicans like current House Speaker Mike Johnson who are vocal fans of the act. The Jones Act isn’t an obviously partisan issue, which is good: it means there could be bipartisan support for reforms, which is helpful even under unified government — but will be essential if, as expected, Democrats take the House this fall.

Realistically, any progress will involve real compromise. There aren’t the votes for an outright repeal of the Jones Act, and there are not likely to be the votes any time soon. There might be votes, though, for smaller measures:

• An exemption for Puerto Rico, Hawaii, Alaska, Guam, and Northeast states reliant on heating oil, or perhaps just one of that set

• Repealing the build requirement of the act but not the crewing requirements, so that labor unions representing US seafarers continue to benefit, and perhaps even gain as the new foreign-built ships hire their members

• Pairing adjustments to the act with buyout offers to current US-based shipbuilders, so those affected have resources to choose a new career path and the transition is truly win-win

• Pairing changes to the act with subsidies to US-based shipbuilders akin to those in the SHIPS for America Act, so US firms face more competition but also have more resources with which to face it

None of these are my first choice of policy regime, and I don’t suspect any of them are the first choice for any player in this debate. But the current Jones Act waiver makes compromises like these uniquely possible. Will pro-Jones Act members of Congress choose any of these over an in-effect Jones Act? Of course not. Might they be willing to budge if it means an indefinite Jones Act waiver comes to an end and they get at least some of the law back? Potentially.

Right now there simply aren’t many groups trying to work through what a good compromise could look like. If you are a think tank, advocacy group, or independent policy researcher interested in this area and crafting a compromise that reduces the Jones Act’s harms while providing some kind of win to its promoters, please get in touch. The waiver means the time for action is now.

1. Ignoring the price signal - Willow Latham-Proenca

2. The Hostile Turn in Administrative Procedure, and the Ad Hoc Exceptions — Dylan Matthews

3. Actually Existing State Capacity Reform — Matt Clancy

4. The Two McNamaras - Saloni Dattani

5. The Effects of U.S. Public R&D on Global Growth — Dylan Matthews

And Alex, Jordan, and Nisha are all traveling this week.

Ignoring the price signal — Willow Latham-Proenca

How much do prices affect how much electricity we use? It’s an increasingly relevant question, since changing when people and firms use electricity – usually to try to spread out short-term peaks – is often one of the first proposed responses to the current energy crunch. It would also seem to have an obvious answer – if electricity prices go up, from an Econ 101 perspective, use should go down. Some go so far as to say that exposing customers (and their devices) to real-time prices should even be enough to organize meaningful demand response, without virtual power plants or other middlemen.

However, it turns out people really like electricity. A new paper this month finds that prices matter less than you’d think – Kudela and coauthors find that price increases do little to change demand (a 10% rise in the electricity price cuts consumption by under 2% on average, over the short term). Even more surprisingly, the most careful studies reviewed find prices have next to no effect at all on electricity use. They also find that elasticity is higher over the long run than the short term – consistent with users adapting to higher prices through capital improvements, rather than behavioral change (think a more efficient dryer, not waiting until midnight to dry your clothes).

The findings here are not just surprising but also policy-relevant, implying both that price transparency won’t do much to smooth demand without paired technology improvements, and that firm capacity (generation that can be called on at times of peak demand) will likely continue to be needed, especially over the short term. To be clear, this isn’t an argument against peak-shifting as part of a larger energy abundance strategy (particularly when intra-day price-sensitivity can be programmed into automated devices), but it is a reality check on how most people think about and use energy.

Speaking of paired technology improvements, Massachusetts launched a residential vehicle-to-grid pilot program this month – exactly the sort of demand-smoothing technology the Kudela paper implies we’d need to actually get electricity demand to follow supply.

The Hostile Turn in Administrative Procedure, and the Ad Hoc Exceptions — Dylan Matthews

“Abundance” as a movement has a set of diagnoses of what’s wrong with US public policy, and how to fix it, but it also has a historical narrative behind it, which goes something like this: the US used to be able to build stuff, fast. Starting in the 1970s, we lost our ability to build stuff. Zoning got tighter, permitting energy projects got harder, red tape got thicker all around.

Alex Mechanick, a professor at Johns Hopkins and veteran of the Biden regulatory shop, calls this shift “the hostile turn,” and has a useful new post outlining what it actually consisted of. Drawing on Paul Sabin’s book Public Citizens (for my money, the single most important text on American abundance politics) he argues that the slew of new substantive rules in the ‘70s did not, by itself, lead to this sclerosis. The Clean Air Act and the Vehicle Safety Act did not, on their own, make it impossible to build. The hostile turn, instead, consisted in a set of procedural obstacles, sometimes set up in substantive laws like the National Historical Preservation Act, and sometimes in purely procedural measures like the National Environmental Policy Act (NEPA), “statutes that tied the hands of agencies through more and more onerous procedures, rather than a focus on regulating on the basis of scientific and economic expertise.”

That much of the story is probably familiar; NEPA in particular is a common target of criticism from abundance folks. But for Mechanick, the lesson is that procedural problems demand a procedural solution, and that begins by rethinking and revamping the Administrative Procedures Act (APA), the so-called “constitution” of administrative law in the US. This is a move that most commentators don’t make, and Mechanick takes it to some interesting places.

Actually Existing State Capacity Reform — Matt Clancy

Compared to some of the other areas we work on, improving “state capacity” can seem a bit nebulous. The end goal of improving the productivity of government is laudable, but what, specifically, does that mean? With that question in mind, I enjoyed Jen Pahlka’s rundown of North Carolina’s recently passed Public Modernization Workforce Act.

One concrete way of making government more productive is via civil service reform. Prior to the new bill, finding new job opportunities in the North Carolina government was byzantine, giving calibrated feedback to workers wasn’t happening, retaining high performers with pay increases was challenging, and letting go of poor performers was extremely protracted and difficult. The Public Modernization Workforce Act streamlines much of this process, and introduces a lot of new innovations to help recruit, retain, and effectively manage civil service talent: a track to more easily make temporary workers permanent, a pilot testing upward feedback (where reports rate the performance of their managers), and setting in motion a review of the underlying classification and compensation system. As Pahlka notes (and we agree), hopefully the Federal Government and other state governments can learn from North Carolina’s experiment.

The Two McNamaras - Saloni Dattani

I enjoyed reading Oliver Kim’s article on the two lives of Robert McNamara: one as US secretary of defense and chief escalator of the Vietnam War, the other as World Bank president and crusader against global poverty. He suggests that though it’s tempting to contrast these as different sides of the same man, they shared common threads: a drive for quantification, a bureaucratic command approach, and the ambition to reorganize large institutions. In Vietnam, he used the Hamlet Evaluation System, which graded 12,000 hamlets each month and fed the results into an IBM mainframe that invariably showed the US winning. At the Bank, he more than tripled the staff and turned an institution run by engineers into one run by economists; yearly lending grew from about $1 billion in 1968 to $13 billion in 1981.

In both eras, he seemed to focus more on what was easily measurable rather than strategically relevant. The North Vietnamese were willing to absorb losses that American planners had not anticipated, and land reform was not considered relevant until much later. The Bank did some good, virtually eradicating river blindness in West Africa, for example, but also drew criticism, with a culture of lending with little regard for how the money would be used.

Archives suggest that McNamara privately believed by November 1965 that the Vietnam War could not be won, but continued selling it publicly for years, boasting enemy body counts. Asked later why he stayed silent, he offered the excuse that his loyalty lay with the president. Oliver concludes, “if there is a thread we can trace through his career, from the Pentagon to the Bank, it is that no amount of technocratic skill can substitute for ethical judgment. McNamara was one of the 20th century’s great systems builders—a man who could tame vast bureaucracies, enlarge them, rationalize them. That doing what was right might require sometimes stepping outside the system simply did not compute.”

The Effects of U.S. Public R&D on Global Growth — Dylan Matthews

Last week, I mentioned a big Congressional Budget Office report estimating the economic impact of federal spending on non-defense research and development efforts (think basic and applied science at the National Science Foundation or National Institutes of Health, more engineering-like work at ARPA-H or ARPA-E, etc.). That report leaned heavily on the work of economists Andrew Fieldhouse and Karel Mertens, whose recent American Economic Review paper constitutes some of the best evidence we have on the economic returns to R&D spending in the US.

Now, Fieldhouse and Mertens have teamed up with three other researchers (Gustavo De Souza, Ishan Nath, and Valerie Ramey) to extend their work and estimate the global impact of US R&D spending. It stands to reason that science and engineering in the US should have effects abroad. When Bell Labs researchers invented the transistor, it didn’t take long for that to be powering radios in Japan. The team finds that the US captures only about half the returns to US-based R&D, with the other half benefiting the rest of the world. The US-only returns are already enormous, and well worth it even if you only care about US well-being. But if, like us at Coefficient, you try to look at well-being worldwide, this finding suggests US R&D is an even better deal than we thought.

One more thing

Finally, in case you missed it, Jordan published a new website last week, Ordinary Abundance, which commenter Zac Hill described as “almost the inverse of a memento mori - a reminder of the sheer infinitude of life, its almost banal fecundity of possibility.” Read Jordan’s blog post about it here.

In our work, we focus on the former, trying to improve the lives of people around the world by making housing more available, energy more abundant, and health more secure. To date, we have not funded the Institute for Lower Expectations (though they keep asking). That focus will continue. But to fully capture the value of improving reality, it is important to step off the hedonic treadmill every now and again, take a breath, and reflect on how far we’ve come.

There’s a bit of this thinking in the air these days. My colleague Matt Clancy wrote a piece last year exploring how new technology improves his own life; Charles Mann published a six-part series in The New Atlantis showcasing the complex, awe-inspiring societal infrastructure that many of us take for granted; he and Virginia Postrel also recently started a podcast, called Everyday Abundance, with a similar goal; and Jerusalem Demsas and Kelsey Piper wrote about all of the seemingly small joys of modern life that past generations couldn’t dream of.

In that piece, Kelsey writes, “I try — once a year or so, perhaps around Thanksgiving — to surround myself with the thoughts of all the people who have ever lived. Almost every single human who has ever lived was very poor. With them accompanying me, I walk through my house and try to see it through their eyes.” While she was drafting those words, I was working on a project to help people do exactly that.

The limit of human felicity

In their Everyday Abundance episode on music, Charles and Virginia highlight this quote from an 1888 novel by Edward Bellamy:

[I]f we could have devised an arrangement for providing everybody with music in their homes, perfect in quality, unlimited in quantity, suited to every mood, and beginning and ceasing at will, we should have considered the limit of human felicity already attained.

I find this sentiment breathtaking. Less than 150 years ago, “accessible music in the home” was a dream that some felt would represent the absolute pinnacle of human happiness; for most of my adult life, I have been able to listen to much of humanity’s discography whenever I like, and, without necessarily realizing it, I’ve used that privilege to make cherished memories, deepen relationships, and better understand myself.

Day to day, I find it difficult to conjure that sense of awe and appreciation. But Bellamy’s quote got to me, and I wondered whether similar historical expressions of hope, relief, and joy for the things we now find commonplace might help more people access those feelings themselves.

My attempt to capture those sentiments is ordinaryabundance.com.

The core of the site is a tour through past perspectives on our modern amenities. I loved hunting down these quotes, so if you know of any others that deserve to be highlighted, please do reach out (or comment here). I hope you take a look, and if you crack a smile next time you do a load of laundry I’ll call it a win.

For now, I’ll leave you with the closing paragraphs of the site, and what I personally took away from this exercise:

It is to humanity’s credit that we remain restless in the midst of all of this progress. We continue to look forward, pushing the frontier further with new treatments, new tools, and new institutions that will help future generations in ways we can’t even picture yet.

But our lives today are a gallery of past generations’ heroic efforts to do the same. It serves us, and honors them, to recapture whenever possible the old sense of awe at these wonders that have long since become commonplace.

Prediction markets are already ethically controversial because of concerns about insider trading and market manipulation. So Kalshi says that, for now, it will limit the markets to Phase 3 trials after they’ve been fully recruited, and will include some guardrails requiring employment verification to bar insiders like lead investigators, biostatisticians, safety-monitoring committee members, the drugmaker’s own staff, and trial participants from taking part.

But I wonder whether, or how, these markets would be useful. Anne Wojcicki, who runs the 23andMe Research Institute, says the idea is ‘very patient-empowering’ and that they could help people follow trials relevant to their disease, but I think the case is pretty weak: Kalshi says their markets would open only once a trial has finished enrolling, so they couldn’t influence a patient’s decision about whether to enrol, at least in that study; rather, the main change a patient could make based on the market would be dropping out. If that happened, it would make it even harder for trials to retain enough statistical power to reach their endpoints.

That kind of self-fulfilling outcome seems like it would have even more downside in an industry that’s already very conservative and finds recruitment challenging enough. (Maybe some doctors would ask for access to promising experimental drugs for their patients through ‘compassionate use’ based on prediction market signals, but in practice that depends on the drugmaker’s willingness to supply the drug before it’s approved, which is usually pretty limited.)

Still, prediction markets might be more useful to external researchers, investors, and policymakers, in shaping follow-on research or funding around them. They might also help surface suspected fraud earlier, the way some short sellers have done. But, because they would make the stakes more directly linked to each trial, I’d guess they’ll also raise the risk of it happening in the first place. (See, for example, a recent case of statistical manipulation in a high-stakes phase 3 trial that was only uncovered after the drug was approved!)

It’s probably going to become more important to maintain the integrity of clinical trials, with stronger guardrails in blinding researchers and patients in the trial, strengthening regulatory independence, and maybe also requiring researchers to pre-register the full statistical analysis before the trial begins.

1. Summer housing roundup and the end of parking minimums in NC — Alex Armlovich

2. Can you pay off the NIMBYs? — Nisha Austin

3. What’s a dollar of R&D worth? — Dylan Matthews

4. What's new in the economics of innovation — Jordan Dworkin

5. Advanced market clinical trial design commitments — Saloni Dattani

6. Sobering takes on Canadian wildfires — Willow Latham-Proenca

Stay tuned for a What we’re reading spotlight from Saloni tomorrow. And Matt is on holiday this week.

Summer housing roundup and the end of parking minimums in NC — Alex Armlovich

Halina Bennet has a summer housing roundup at Slow Boring: Beyond the already-covered ROAD Act, California's Governor Newsom signed AB 179, replacing the state's notoriously fragmented affordable-housing financing with a single application and penalizing localities that charge development impact fees on state-funded projects (saving $60,000–$70,000 per home). San Francisco slashed its unfunded inclusionary housing mandate from 15% to 5% after a court-ordered feasibility study found that even 5% could render projects infeasible under current costs and rates. Massachusetts folded a "YIGBY" (Yes in God’s Backyard) provision into its economic development bill, letting religious institutions build multifamily by right on land they've owned for three years. Finally, Zillow's latest census analysis, estimating a “national housing deficit” by comparing the number of doubled-up families to the available housing supply, held essentially flat in 2024 for the first time since the post-2008 underbuilding era began, driven by a 50-year high in multifamily completions. Even these crude accounting-based proxies for housing underproduction (which do not directly consider relative prices or inter-metropolitan mobility) nonetheless find the 4.7-million-unit deficit remains concentrated in the usual suspects: New York, LA, Boston, SF, DC.

Separately, North Carolina just became the first state to eliminate nearly all mandated parking minimums statewide. Governor Stein signed HB 162 (the Parking Lot Reform and Modernization Act) taking effect January 1, 2027. The coalition story is remarkable for most legislation but typical of the greatest YIMBY wins: the bill united environmentalists, farmers, developers, and business leaders. Reason notes the liberalization covers commercial uses too, not just housing, which again is unprecedented in statewide parking reform. The bill started as a stormwater management bill and the bipartisan framing (focused on stormwater regulatory relief first, and urbanism second) is what got it through a Republican-controlled legislature. Whatever the ideological context and framing, it remains fascinating to see parking relief laws begin to leap forward in car-heavy red states while trundling along incrementally in transit-rich blue states.

Can you pay off the NIMBYs? — Nisha Austin

A forthcoming paper by Michael Hankinson, Edward Glaeser, Joseph Gyourko, and Morris Davis surveyed roughly 1,700 residents of expensive areas and found that most people do have a price. Over 80 percent accepted at least one hypothetical housing proposal when offered cash compensation. The amount depended heavily on density: residents in dense urban areas needed only a few hundred dollars, moderate-density neighborhoods required low thousands, and low-density suburbs needed well above $5,000. Proximity and project type mattered too: proposals next door doubled the price, higher-density projects quadrupled it, and low-income-designated housing multiplied it fivefold. I was surprised to learn that when compensation was offered as a parks-and-streets fund rather than direct cash, residents demanded ten times more. As Rogé Karma writes in the Atlantic, this has prompted several think tanks to propose direct payments to residents in communities that hit housing production targets. What we'd love to see is similar research on infrastructure (in particular, transmission lines, wind farms, and industrial facilities). The evidence base there is almost nonexistent, even though siting opposition is just as central to those fights.

What’s a dollar of R&D worth? — Dylan Matthews

Here at AGF, we tend to think investment in research and development, whether by governments through agencies like the National Institutes of Health or in private businesses and universities, is worth an awful lot. When you get down to it, economic growth is about getting more benefit out of the same amount of physical stuff; you can only extract more benefit from the same stuff if you have better technology, and you get better technology through scientific research.

But just how beneficial is R&D? A big new report out of the Congressional Budget Office, coauthored by economists Sheila Campbell, Jaeger Nelson, Eli Schrag, Heidi Williams, and Caleb Wroblewski, has some of the most careful analysis of this question to date, using two radically different methodologies that come to quite different numbers. One approach results in an estimate that each federal dollar spent on R&D, financed by an increase in the budget deficit, grows GDP by $5.15. An alternative framework leads to an estimate of $1.98; the numbers in each case are higher if you pay for the funding through spending cuts, rather than borrowing it. Even the smaller number implies that R&D spending is an incredible deal, with each dollar invested doubling its value.

What’s new in the economics of innovation — Jordan Dworkin

This week I’m at the NBER Summer Institute, taking in lots of new work on the economics of science and innovation. Two of the papers being presented have already made appearances on this blog (one on idea diffusion, and one on the impact of AlphaFold). The rest of the agenda is also worth a look. Chen, Liu, & Ma take a new approach to estimating the economic value of science, building a text-based innovation network from 463 million papers and 81 million patents to trace how knowledge flows into technology; Songyuan Teng decomposes pharmaceutical innovation into novelty (new chemical building blocks) and recombination (new assemblies of existing blocks) and models their relative importance; and Mohnen & Gans have a new study on AlphaFold showing that it sped up experimental research but narrowed its scope (echoing other findings on contraction in AI-driven science).

Advanced market clinical trial design commitments — Saloni Dattani

I enjoyed reading a post by Witold Więcek on the Clinical Trials Abundance blog, titled ‘Market shaping and trial design belong together’. He explains why people interested in solving market shaping problems – such as economists designing advance market commitments and philanthropists funding multiple R&D projects for global health problems – should be interested in innovative clinical trials.

Most drugs today are tested in separate clinical trials: one drug vs one placebo, which is much more inefficient than running them together in a platform trial (with, for example, five drugs vs one placebo). That structure also makes it hard to make head-to-head comparisons between different drugs, since separate trials might enrol different kinds of patients, run at different times, and measure different outcomes.

That means the standard way to run clinical trials is less useful for decision-making. A market shaper, for example, might want to specify some criteria that different drugs have to achieve in order to receive a prize or payout from an advance market commitment. But differences in trial design and setting would leave open a lot of uncertainty about which ones qualify. So, he argues, why not have them play a role in designing the trial as well?

Sobering takes on Canadian wildfires — Willow Latham-Proenca

If you were in the Northeast US last week, you probably got at least a noseful of Canadian wildfire smoke. We’ve written before about the extraordinarily high health costs of wildfires, but Zeke Hausfather makes a convincing case that Canada’s fires – unlike many in the U.S. West – aren’t mainly driven by institutional mismanagement. Unlike in California, these Canadian fires are mostly occurring in boreal forests that have never been managed, so there’s no unnaturally accumulated detritus to burn. Devastating crown fires driven by weather (heat waves + lightning strikes) are the norm in these environments - so higher average temperatures directly drive more burning without additional human intervention.

Robinson Meyer makes a related case in Heatmap, arguing that the failure of climate research to integrate the potentially massive health effects of wildfire smoke is a potent illustration of how big the unknown-unknowns of climate change could be. While there was some pushback about the deep history of wildfire research, the fact that the resulting health effects weren’t integrated into flagship models of economic damages stands – as a result, policymakers had very little basis to understand and incorporate smoke risks before these types of fires became a regular occurrence. Part of this might be a map of science problem – wildfire risk, epidemiology (famously tricky for events like this), and mainstream climate modeling just didn’t overlap enough to stimulate research on their intersection until facts on the ground made the gap obvious. It’s worth wondering whether this is a problem that AI will help with, or one that the jagged frontier of scientific research will exacerbate.

Here are a couple highlights from our grantees:

• The Inclusive Abundance Institute published “The Biggest Housing Bill in 30 Years Is Now Law,” a look at what the ROAD Act means for the abundance movement.

• YIMBY Melbourne launched the Australian Zoning Atlas, mapping planning controls across all major Australian cities.

1. 124 implementation actions and counting — Alex Armlovich

2. 200,000 comments and the state of US science — Jordan Dworkin

3. Does upzoning work? — Matt Clancy

4. Designing a cancer vaccine — Saloni Dattani

5. Western transmission and DOE needs — Willow Latham-Proenca

6. Space mirrors — Dylan Matthews

124 implementation actions and counting — Alex Armlovich

Aaron Shroyer and Sarah Brundage at NAAHL have the essential federal implementation guide for the ROAD Act. None of the law's ~47 housing sections are wholly self-executing: NAAHL counts 124 discrete implementation actions across eleven agencies, with 58 due within the first year. HUD owns 88 of them, including nine rulemakings (environmental review streamlining, the chassis definition, offsite construction financing, HOME reform). It might take awhile: HUD finalized roughly four significant rules per year last term, and its current agenda already has 54 rules queued. The staffing math is worse: the three HUD offices holding three-quarters of the workload are down 27–30% from FY25, and the General Counsel's office that must clear everything is down 35%. Meanwhile a dozen new programs (pattern-book grants, the Innovation Fund, PRICE, single-stair Point-Access Block grants) exist only on paper until Congress appropriates funding, and several sunset within three to seven years…so every year of delay compresses their operating window. Think tanks and university research specialists in housing should be ready to step up and help HUD’s limited headcount move these research, guidance, and rulemaking procedures forward.

200,000 comments and the state of US science — Jordan Dworkin

This Monday was the deadline for public comments on OMB’s proposed rewrite of the rules governing federal grantmaking. The response was immense, with more than 200k comments submitted over the past few weeks. As expected, responses poured in from scientists, universities, societies, and publishing outfits. But because of the scope of the OMB proposal (it affects over $1 trillion in annual federal assistance), concern came from well beyond science, with the National Association of Counties filing its own comment, and the National League of Cities and Moody’s raising the potential negative impacts on project financing and credit for state and local borrowers. In Marginal Revolution, Alex Tabarrok highlighted broader, increasing alarm from organizations and thinkers not inclined towards hyperbole.

For Tech Policy Press, Chris Marcum and Abigail Haddad analyzed the ~50,000 comments posted by last week, and found roughly 94% opposed and 6% in support, with opposition largely grassroots (only ~1/8 of opposing comments came from a form letter) and comments in support almost entirely organized (nearly 9/10 supportive comments were from a single form-letter campaign). The dominant objections were the politicization of grant decisions and review, and the expanded termination authority. OMB has said it wants to finalize the rule by October 1 and has declined requests for an extension of the comment period. But litigation seems likely.

Even in the absence of the new rules, US science is continuing to strain under the pressure of this administration. NSF is on track for its worst grantmaking year since 1985 (~2,200 awards so far in FY26 compared to an FY21-24 average of almost 6,000); PhD admissions across top research universities fell 15% this year after a drop of 11% last year, with Caltech cutting its incoming class by 40%; and last week NSF moved to ban collaborations with nearly all Chinese research institutions, dropping an existing risk mitigation approach (TRUST) in favor of a flat prohibition.

Does upzoning work? — Matt Clancy

Does upzoning lead to new homes? Does it reduce the cost of homes? Michael Wiebe has an accessible review of the academic literature around upzoning in his living literature review, examining empirical studies on major upzoning cases in Auckland, Sao Paolo, Wellington, New York, and Zurich. Upzoning consistently increases the housing supply, and when studies are set up to measure the impact on prices, it brings costs down (at least relative to what would have happened without upzoning - sometimes that just means prices increase less fast). But the link between upzoning and these outcomes is a complicated one (as Alex has written about before); upzone in a place where building new housing requires a lot of demolition, and you’ll need to upzone by a larger amount to see a comparable effect.

Designing a cancer vaccine — Saloni Dattani

The great blogger known as owlposting has an awesome new podcast episode with Alex Rubinsteyn and Ben Vincent, two scientists who work on cancer immunotherapy, on ‘How to design a cancer vaccine’. The general idea behind a cancer vaccine is to prevent cancer cells from evading the immune system by training patients’ own immune cells to recognize them and kill them off.

Unfortunately, many past attempts to develop personalized cancer vaccines have failed. This is in part because they were designed based on the genetic profiles of cancer cells, rather than the actual proteins those cancer cells expressed on their surface. Measuring which proteins were actually presented would typically require samples to be collected differently than they usually are (fixed with formalin, rather than frozen) and the measurement was expensive and, in older versions, not very accurate. There were other problems too: mRNA vaccines were a new platform and faced their own regulatory scrutiny and immune cells were hard to steer well. They tend to over-respond to just a few target mutations, which might distract immune cells from the real ones, or they might be set loose on the wrong ones, which could be dangerously toxic.

But now, things might be different. Prediction and measurement have gotten better, the mRNA platform has been tested in billions of people, and there’s a clearer regulatory path for personalized drugs. I really enjoyed the episode and feel cautiously hopeful. It’s also worth listening to hear about how weird cancer cells are: they can, for example, reawaken ancient viral DNA buried in our genomes or even switch on genes that are normally only active in the sperm or placenta. You can listen here.

Western transmission and DOE needs — Willow Latham-Proenca

It’s been a busy few weeks in transmission – at the close of June, 11 western states agreed to improve coordination and permitting for transmission across their shared footprint (much of the western grid outside of California is operated by a patchwork of utilities, unlike the northeast and midwest, where regional transmission organizations plan and operate the grid). The pact follows a study by the Western Transmission Expansion Coalition earlier this year, which identified more than 12,600 miles of new or upgraded transmission needed in the region over the next decade, a quarter of which are yet to be even planned for.

The DOE also released its most recent draft report on national transmission needs. It’s no surprise that high-voltage interregional and cross-interconnection links continue to be massively underbuilt, with consequences for both energy prices and grid reliability - for example, the study found transmission congestion costs reached $12 billion in 2024. The draft also highlights another geography where transmission is critical – DOE estimates that the southeast (another non-RTO zone) is missing out on about $10/MWh in congestion relief value – essentially, cheaper power can’t get to where it needs to go – as a result of underbuilt connections to neighboring regions.

Space mirrors — Dylan Matthews

Solar panels only generate power when the sun is shining, and everyone knows you can’t make the sun shine on demand. What the startup Reflect Orbital proposes is: maybe you can? Cofounded by a SpaceX veteran, the company proposes to launch satellites with targetable mirrors that can redirect sunlight to the dark side of the earth. The main idea is to produce more frequently-available solar generation, but the company also proposes using the technique to assist in search and rescue and other disaster response.

The proposal is taking a step forward, with the FCC approving a demonstration satellite. This piece from Hiroko Tabuchi does a good job of laying out the appeal of the idea (plentiful solar power!) and the real possible downsides, like disruptions to astronomical observations and animals’ circadian rhythms.

Here are a few other highlights and announcements from our team and grantees:

• The ROAD to Housing Act was signed into law — the first major federal housing legislation in roughly 30 years. Alex’s blurb above covers what comes next on implementation.

• Caleb Watney published A Long Sequence of Small, Correct Decisions, announcing his new role at Coefficient Giving leading a team focused on strengthening the institutions that will navigate the transition to powerful AI, and overseeing the Abundance and Growth Fund and CG’s government relations.

• Matt published the Q2 2026 progress report, a look back at the fund’s work over the last quarter.

• California YIMBY published the latest Homework newsletter, highlighting YIMBY-endorsed candidates advanced in 17 of 19 competitive open seat races in recent elections, a strong showing for Open New York and California YIMBY.

• California YIMBY also published a summary of Brian Potter’s working paper on the divergence of housing prices from construction costs across US cities.

This report is also a team effort - we can’t list everyone as an author on substack, but this is the work of the entire Abundance and Growth Fund team.

Housing Policy

After quite a rollercoaster, the Senate passed the 21st Century ROAD to Housing Act 85–5 on June 22, 2026, and the House passed it 358–32 shortly after. Speaker Johnson sent it to the White House on June 29 and while President Trump declined to sign the bill, it will become law by default if he neither signs nor vetoes it within 10 days (Sundays excepted) while Congress remains in session — likely today. The bill’s 45-plus housing provisions cover a wide array of reforms: cutting red tape, overhauling funding and financing, and providing technical assistance plus carrots and sticks for state and local reform. Beyond incentives, the supply-oriented regulatory fixes and financing measures trim federal permitting friction, expand the definition of manufactured housing by relieving the HUD code chassis mandate, raise and index FHA multifamily loan limits, lift the bank public-welfare-investment cap from 15% to 20%, and stand up grant programs to convert vacant buildings and rehab aging homes.

While the ROAD to Housing Act provides direct cleanup of key paperwork and bad rules that exist at the federal level, most of the binding constraints on housing still live at the state and local level.

• Idaho — In April, Governor Little signed SB 1352, SB 1354, and HB 706, requiring cities over 10,000 to allow small-lot starter-home subdivisions and ADUs and permitting cities to authorize single-stair apartment buildings.

• Virginia — In April, Governor Spanberger signed a housing package including SB 531 (by-right ADUs in single-family zones, signed April 13) and SB 388 (by-right affordable housing on faith and nonprofit land).

• New York — This May, Governor Hochul signed FY2027 budget legislation exempting qualifying infill housing projects — up to 500 units in New York City — from SEQRA environmental review.

• Maryland — On May 26, Governor Moore signed HB 894, preempting local zoning near high-frequency rail stations, and HB 548/SB 325, vesting development rights at application and deferring impact fees to occupancy.

• Florida — Governor DeSantis signed HB 803 (permitting reform, including a permit exemption for residential work under $7,500) in May and HB 1389 (extending Live Local Act by-right approval to government- and faith-owned land) in June.

• North Carolina — H162, eliminating local minimum parking requirements statewide, passed the Senate 44–1 in June and the House by concurrence that week and became law this month.

• Washington, DC — The DC Council unanimously passed the One Front Door Act (B26-0227) in May, directing code amendments to permit single-stair apartment buildings up to six stories; the Mayor signed it in June, and it becomes law later this summer, after the 30-day congressional review.

Finally, Center for Building’s proposal to raise America’s “International” Building Code single-stair apartment height limit from three stories to four passed its final membership vote and will officially make the 2027 International Building Code.

Innovation Policy

In May, the National Science Foundation announced its new X-Labs program, a $1.5 billion, decade-long initiative to fund independent, milestone-driven research teams working outside of traditional university and corporate structures. The program aims to address a long-held concern that existing grant mechanisms and institutional models aren’t well suited for team-oriented, engineering heavy, long-term research efforts. The initiative draws on ideas from across the science policy community, including the Institute for Progress, whose X-Labs proposal last August laid out a framework for this kind of funding model.

While we’re excited about this program, it is currently proposed to be under 2% of NSF’s annual budget; though we hope that will grow after the model proves itself. It’s also worth noting it could end up a higher share in the short run, if the NSF fails to disburse conventional grants on time, which it is at risk of doing (see below).

Related to the X-Labs announcement, the Abundance and Growth Fund supported US-based cohorts of the Big if True Science (BiTS) accelerator and Brains research accelerator, both of which graduated and pitched their ideas for ambitious science projects this quarter. These are the kinds of ideas that (we hope!) will one day become X-Labs, or Focused Research Organizations, or ARPA programs. Speaking of FROs and ARPA programs, in April Convergent Research and ARIA continued the expansion of the FRO model across the pond, partnering to launch two new UK FROs: Meridial and Echo Labs.

Turning to the majority of the NSF budget that funds traditional research grants and training programs, we were encouraged to see the NSF’s 2027 budget request propose an NSF metascience unit to “champion agency-wide efforts to explore non-traditional R&D funding mechanisms and drive evidence-based optimizations to the R&D portfolio.” We’ll be closely watching to see if this is followed through on.

One challenge facing traditional grant systems is excess conservatism, or risk aversion. A first step in fixing this could be measuring it. The Metascience Novelty Indicators Challenge organized by UKRI (for which we funded the prize) is one recent attempt to build such a measure. The challenge was a global competition to develop scalable methods for assessing the novelty of research papers (see coverage in Science here). The winning team from Forschungszentrum Jülich was announced in June; they built an AI-based approach that compares the knowledge contribution of a paper to the state of knowledge, as represented by the text of the papers it cites.

Sometimes a prize like that one is the best way to make progress on an innovation challenge, sometimes a grant program would be, but other times a different policy is a better fit. Maybe an advanced market commitment, maybe a focused research organization, maybe something else to help policymakers decide which is which. The Institute for Progress and the Market Shaping Accelerator joined forces to create the Atlas of Innovation: An interactive guide for policymakers to identify the right funding mechanism for innovation challenges.

The above projects aim to improve how research funding translates into scientific and technological progress. But that presupposes funding is happening in the first place. And while we’re excited to see new innovative initiatives at the NSF, this administration has also been hostile to federal support for science (which we think has very high returns - see this, and this project we are supporting). After Congress rejected the administration’s proposals for deep cuts to science last year, spending on new awards (see figure below) has largely not kept up with historical averages, though there was improvement at NIH, DOE, and NSF over the last quarter.

Caption: FY2026 New Award-Making Pace vs. Historical Average. Source: sciencespending.org

Last year, spending accelerated in the back half of the year for many agencies, and there is still time for the same thing to happen this year. Even so, the delays in funding are causing real long-run damage. A combination of rising uncertainty, the awarding of fewer (but larger) grants, and tighter immigration and visa policies are jointly contributing to significant declines in graduate student enrollment.

Clinical Trials

The FDA announced a flurry of new policies aimed at making the US clinical trials system work more effectively. One reason that clinical trials are expensive and inefficient is that the data they generate are locked in siloed systems and manually transcribed. In April, the FDA announced a pilot project to fix this, which would create ‘Real Time Clinical Trials’. The idea is to let regulators see trial data in real time, as it’s collected, potentially compressing the dead time between study phases.

Additionally, in June, the US Department of Health and Human Services (HHS) announced a new initiative called ‘Operation Trial Blazer’ that aims to bring back early-stage clinical research, which has increasingly been outsourced to countries like China and Australia, by reducing how long it takes to get approval to start phase 1 trials as well as how long it takes the FDA to review data from phase 1 trials.

Operation Trial Blazer includes several parts. One is an FDA pilot program to set up a network of qualified research institutions that help researchers prepare the protocol and supporting components before they submit it, though the FDA still decides whether a trial can proceed. It also provides more guidance on what manufacturing data is needed for phase 1 and clarifies this is lower than for later-stage trials, since companies have so far tended to submit more data than necessary to secure approval. And it includes guidance on choosing the first doses given to humans, recommending these are based on statistical modeling of the drug’s pharmacology rather than animal toxicology studies, which were often poor predictors of side effects in humans. Finally, there’s a platform for the FDA to review components of the package as they arrive, as a ‘rolling review’, rather than all at the end.

Most of this sounds positive, but there’s a caveat, at least in the short term. The FDA’s main bottleneck right now is staffing capacity, given the cuts over the last few years (roughly a fifth of the agency’s workforce, around 3,500 FDA staff, were cut last year). Whether these reforms ease that problem or worsen it further is an open question. Rolling reviews especially require more staffing time, because reviewing at the end typically means some attrition from researchers who don’t make it to that stage; similarly, the requirement that the FDA still sign off on whether phase 1 trials can proceed adds to that load too, especially while the initiative is new and leaning on staff input. Still, there’s good stuff here: the guidance on manufacturing data and dose selection doesn’t depend on FDA capacity the way the pilot does, so some of these changes could result in improvements soon.

Energy

On federal energy permitting, a bill has remained elusive, but we’re holding out hope for a bipartisan package before the August recess. In the meantime, there are two bipartisan moves worth keeping an eye on. In June Senators Cotton and Cortez Masto introduced a Senate version of the FREEDOM Act, which would help prevent agencies from revoking permits or halting construction on fully permitted energy projects, and is a direct response to the administration’s stop-work orders on offshore wind. The Senate version rolls in much of the House’s geothermal package from earlier that month and is notably cleaner than the House’s: it strips the most stringent agency penalties and moves fees into a freestanding “Permitting Performance Fund,” which should quiet concerns that the original structure would create a cost spiral for agencies. Meanwhile, in mid-June Senator King added an amendment to the NDAA targeting the DOD’s increasingly creative use of flight path certification to block onshore wind projects. Neither energy bill is a sure thing, but both signal that permitting certainty has bipartisan legs even in this environment.

On the electricity market side, FERC issued its long-awaited response to DOE’s October 2025 rulemaking request. This took the form of six “show cause” orders directing each of its jurisdictional regional operators – nonprofit regional bodies that run most of the grid outside of Texas - to justify or modify its current tariff structure, as well as submit detailed information on whether its portion of the grid will have enough power to serve new demand. The tight deadlines (60 and 90 days, respectively) for these actions will mean we may see important changes in how large loads connect to the grid and pay for energy over the coming months.

1. 10,000 pieces of type — Nisha Austin

2. Science’s institutional history at 250 — Jordan Dworkin

3. Anthropic develops drugs — Saloni Dattani

4. The Astera metascience essay contest — Matt Clancy

5. Trump v. Slaughter and the end of independent agencies — Willow Latham-Proenca

6. Modular housing and the HUD code — Alex Armlovich

And Dylan is on holiday this week!

10,000 pieces of type — Nisha Austin

One of my favorite 4th of July traditions is to go to the historic fort in town for a reading of the Declaration and firing of the cannons by a historical reenactment group. This year, with the sweltering temperatures, we decided to go to a reading of the Declaration at the local town hall instead. It unexpectedly included a hand-printed copy of the Declaration that was made by setting roughly 10,000 pieces of type by hand and running each sheet one at a time on cotton linen paper, the same slow way it would have been done in 1776. There’s a fitting Massachusetts thread to this: in 1776, we were the only state to ask every town clerk to copy the text into town records and every minister to read it aloud from the pulpit. Two and a half centuries later, a version of this is still happening in our town halls.

There’s a line in the Declaration which defines our unalienable rights as life, liberty, and the pursuit of happiness, rather than the more familiar “property”. Jerusalem Demsas and Kelsey Piper dwell on that choice: the historian Arthur Schlesinger read “pursuit” as practice, as in a right not merely to chase happiness but to have it. Piper goes on to describe all the ways in which Americans live lives of staggering wealth: hot water on demand, a refrigerator, a cupboard holding more kinds of tea than most people who ever lived drank in a lifetime. In 1900, Americans spent 43 percent of their income on food; the White House had no indoor plumbing until the 1840s. Medieval kings would envy us.

Jason Crawford reads the anniversary in a similar way. American history is not a story of decline but one of long, messy accumulation of things built: the reaper, the light bulb, the airplane, the interstate system, the Internet, eradicating polio and smallpox. His piece lands where I did in our foundational texts post: on Lincoln, and what we owe the institutions we inherit. What we have is precious and rare, and it should make us generous. The work is never finished, and every generation has to show up for it.

Science’s institutional history at 250 — Jordan Dworkin

For America’s 250th anniversary, Brad Wible invited a trio of pieces for Science’s Policy Forum tracing the institutional shifts that built the modern US scientific enterprise.

Ronald Daniels writes about the founding of Johns Hopkins in 1876, made possible by a $7m gift (the largest philanthropic gift in US history at the time), which imported the German research university model and laid the foundation for the postwar funding system. Ashish Arora and Sharon Belenzon cover the rise and fall of the corporate research lab, positioning the pursuit of long-horizon research in commercial contexts as a break from the more distributed systems that dominated before (typically individual inventors licensing patents to firms) and after (typically universities producing novel science, startups translating it, and firms commercializing it). And Daniel Gross and Bhaven Sampat trace the federal funding system back to OSRD, which invented the federal R&D contract (described at the time as one of the “great inventions” of the war) and grappled with many questions that remain top-of-mind today, like indirect costs, patent rights, and geographic distribution.

Anthropic develops drugs — Saloni Dattani

Anthropic made two announcements in the last week about its plans in AI for science. First was the launch of Claude Science, a platform that integrates dozens of scientific databases and bioinformatics models and currently runs on Opus 4.81. Second, that it plans to develop drugs of its own.

I found it interesting for multiple reasons: Anthropic committed to working on pre-clinical research – drug development in the lab and potentially in animals – but didn’t rule out taking its drugs through clinical trials as well. It also plans to focus on “neglected diseases” with weak commercial incentives, which it describes as potentially including both rare genetic disorders and tropical diseases, although it hasn’t disclosed any examples of diseases or drug modalities that it might work on.

This follows Anthropic’s acquisition of Coefficient Bio,2 a stealth startup, earlier this year, with roughly ten people who came mostly from Genentech’s computational drug discovery unit; and Anthropic is also building its own wet lab.

But why work on drug development at all? According to Anthropic, it’s to get hands-on experience in the field it works with companies on and to build a feedback loop to learn how to make better tools and models. As their head of life sciences Eric Kauderer-Abrams described it, there’s no substitute for being “in the trenches trying to develop drugs.”

I’d be more optimistic about Anthropic’s chances for developing successful drugs for rare genetic diseases, which tend to be ‘simpler’ in biological terms – as they’re often caused by a single gene or mutation that can be targeted with drugs and gene therapies – than neglected tropical diseases, for which the biological data to understand them, and how to target them with drugs, is often lacking in the first place. And since the best validation of drugs is testing them in human clinical trials, rather than in the lab or animals, I’m left wondering if they’ve deliberately kept that possibility open. Even aside from that, it seems to me that clinical testing and manufacturing, rather than preclinical drug discovery, are more important bottlenecks to progress on both fronts, although most AI tools so far have been built for preclinical work.

News reporting also suggests that their stated focus, on neglected diseases, helps them avoid stepping on pharma companies’ toes. In a sense, thanks to the Orphan Drug Act, faster approval processes, and rare disease priority vouchers, rare diseases aren’t as neglected as they used to be and pharma companies do have some incentives to develop drugs for them (and they now make up roughly half of annual new drug approvals) – but there are so many different rare diseases that the rationale probably still works.

But building labs and hiring top wet lab biologists seems like a first step rather than an end goal in itself; and since clinical trials and manufacturing are the bigger bottlenecks, delivering on Anthropic’s public mission would probably mean going much farther than what they’ve announced so far.

The Astera metascience essay contest — Matt Clancy

The most interesting thing I read recently was the winning entrants in the Astera metascience essay contest, for which I was one of the judges. While not every winning essay was related to AI (see these two, for example), the opportunities and challenges of AI dominated the contest:

• It’s a bit of a truism that AI and data are strongly complementary; but what data should you collect? What if we could use scaling laws as an input into this decision process? Collect data and train models on it, until we can see the trendlines emerge, which will let us estimate how much data we would need to collect to achieve a given performance (more from Peter Koo’s winning essay).

• A major problem in science is the “file drawer problem”, wherein null results don’t get published, in part because writing up any result takes time and energy, and the rewards to doing so are very low for academics (more here). But if AI agents are integrated into the research process, they have all the material they need to do this for us at almost no cost. Could we finally open all (well, more) file drawers? (more on this theme from Niveditha Iyer’s winning essay)

• One of the most important problems in science is knowing what question to study. With AI that can finally parse text as good as most people, we are in a position to map out the enormous network of cause and effect that the collective literature has studied. With a map in hand, could we identify promising gaps we would otherwise miss? Or could we at least make progress on automating the identification of important questions? (more from Prashant Garg’s winning essay)

• The use of benchmarks to assess the quality of different AI models has been a notable hallmark of AI progress, and one that seems likely to spread across science more generally. But how do you design good benchmarks? Shaamil Karim and Jaeeon Lee both proposed ideas related to this.

One more interesting AI-related challenge: when we ran the winning essays through Pangram, some of them were flagged as using AI (though the share that used AI was even higher among the ones that had been triaged out, prior to the AI check).

I think all future essay contests are going to need to think hard about an AI policy. In our case, we cared about the arguments and ideas, not the authenticity of the prose (or even the provenance of the idea?), and so we decided not to factor AI use into our decisions. But questions remain: what’s the right size for a prize, when the cost of writing an essay has declined? To ensure the prize is incentivizing ideas that outcompete what you would get by simply asking an LLM for ideas, maybe contest organizers of the future should use AI to generate essays in response to the contest announcement and covertly submit them; a prize is only awarded if the judges pick essays that were not generated this way.

Trump v. Slaughter and the end of independent agencies — Willow Latham-Proenca

To add even more uncertainty to the fraught landscape of energy planning and regulation, the end of the Supreme Court opinion season last month brought a highly controversial ruling in Trump v. Slaughter, where the court held that the president can fire commissioners at independent agencies without cause. While some argue that this is an appropriate tightening of elected authority over independent bureaucracy, a less-hopeful view (articulated by Robinson Meyer in Heatmap last week) is that the ruling represents a continuing erosion of Congressional authority – in this case, powers intentionally delegated to independent technical experts – relative to the executive branch. While we’re unlikely to see major short-term policy changes on the energy front – the current FERC and NRC are already closely aligned with the goals of the current administration – over the long term, this could reshape policy at FERC and NRC in ways that are hard to predict. Robinson Meyer takes the strong view here, arguing that Slaughter essentially turns the commissions into agencies – implementers of the president’s agenda, rather than independent arbiters (Ben Schifman of IFP has an interesting take in the same thread arguing that ratemaking bodies like FERC will remain exempt - time will tell if SCOTUS will re-litigate that argument).

Modular housing and the HUD code — Alex Armlovich

Kimberly Burnett has a new piece at Niskanen on HUD's new best-practices guidance for states and localities, focusing on one key recommendation: regulate modular housing by harmonized and objective performance standards rather than prescribed construction methods. Modular is still only ~3% of US single-family homes against 28% in Japan, and we’ve lived through many modular “hype cycles” in the US. But factory-built manufactured housing using the national HUD code is still between 5% to 10% of single family housing completions in the US, suggesting a key binding constraint to modular methods is regulatory fragmentation: duplicative factory-and-site inspections, codes written for on-site assembly, and thousands of local building departments with no capacity to evaluate factory-built homes. National model code standards for modular homes, like ICC/MBI 1200 and 1205, are spreading but have not yet succeeded on the scale of Congress’ choice to directly preempt local building codes nationwide with the HUD Code in the 1970s.

And finally, a couple additional highlights worth sharing:

• Institute for Progress published a detailed report on reforming Section 106 of the National Historic Preservation Act, which has expanded well beyond its original “stop, look, and listen” function into a significant source of delay for energy and transmission projects. The report proposes right-sizing the review process by limiting scope, capping visual-impact studies, and setting statutory deadlines, in line with recent NEPA reforms.

• California’s SB 79, the Abundant and Affordable Homes Near Transit Act, took effect July 1, establishing statewide zoning standards that allow increased residential density near major transit stops across California’s urban counties. Cities are taking markedly different approaches to implementation, with some complying, some using the law’s exemption provisions to delay.

Housing

1. The 21st Century ROAD to Housing Act passed the Senate 85-5 and awaits the president’s signature (or, absent a veto, will have automatic enactment by July 10th). 45+ provisions cut red tape, overhaul financing, and provide carrots and sticks for local reform. (June 24)

2. New York took two big permitting steps: Governor Hochul’s SEQRA modernization is reportedly final, and the mayor launched the SPEED Task Force report projecting two years off rezoning timelines. (May 15)

3. “Are we kind of being pricks?“ A Marblehead resident’s viral question captures what happens when towns comply with zoning law by putting housing capacity on a golf course where nothing will get built. (May 8)

4. Jerusalem Demsas also shared her reflections on Marblehead. The town’s story shows what happens when local democracy mediates land-use decisions, and when compliance isn’t the same as housing production. (May 28)

5. Nolan Gray goes deep on California data: ADU permits have produced nearly 150K new units, the density bonus law is facilitating thousands more, and major reforms are less than a year old. (May 8)

6. Michael Andersen of Sightline argues the picture outside California is just as encouraging: states across the country have only recently passed reforms and results are beginning to show. (May 8)

7. A new Urban Institute piece asks what happens after big upzonings in strong markets. Answer: a lot. The evidence that significant reform leads to construction continues to pile up. (April 17)

8. Our Zoned Capacity post got shoutouts from NYC’s Deputy Mayor and Matt Yglesias. In high-rent cities, “Missing Massive” upzoning is essential for unlocking construction that pencils out now. (April 10)

9. Evan Soltas translated his LA permitting paper into a NYC estimate: each year of permitting time saved is equivalent to an 8% drop in construction costs. (April 3)

10. The housing burden is falling disproportionately on new buyers and renters: new homeowners now pay 26% of income, renters who recently moved pay a record 33%. (April 17)

11. Arpit Gupta and Steve Teles diagnose why factory-built housing keeps failing: boom-bust cycles, patchwork regulations, transport costs, and parcel assembly challenges. (April 3)

12. Everyone’s talking about building codes. California YIMBY summarized a UCLA Lewis Center report on how the code-writing process works and how we got where we are. (April 10)

13. Vincent Rollet’s MIT paper on zoning and redevelopment is getting attention from city planners; it’s the kind of structural model that can discipline a rezoning’s projected supply impact. (April 24)

14. Pew’s new address chain video explains how new construction frees up housing across incomes in months, not decades. They also launched a report on pre-approved building plans. (May 28)

15. The NYT editorial board endorsed Massachusetts’ “Legalize Starter Homes” ballot measure, which would require towns to allow homes on small lots in areas served by public water and sewer. (June 3)

16. A VoxEU column argues the relative price of US structures has risen 80% since 1970, making construction productivity a drag on economy-wide growth, not just a sector-specific concern. (June 10)

17. Brian Potter asks where the economies of scale are in construction. The “Idiot Index” (cost of a good vs. raw materials) for housing is roughly 2, not far from the highly automated car industry’s 1.8. (June 3)

18. Florida’s insurance crisis may be easing: after tort reforms, insurers report sharp drops in litigation costs and premium growth. The state had 76% of US insurance lawsuits from just 9% of homeowners. (June 10)

19. Sightline’s new report finds apartments are the climate solution hiding in plain sight: more energy efficient, more likely heated by electricity, and located where people can walk, bike, and take transit. (May 15)

20. Larry Katz’s 1981 article on how growth controls drove San Francisco’s home prices from the national median to the highest in the country in just one decade. A hauntingly contemporary time capsule. (April 17)

21. Oliver Kim reviews a book complicating YIMBYism’s invocation of Singapore: 99-year land leases create a looming political problem, and the city is already so dense that upzoning can’t unlock value the way it does in US cities. (May 8)

22. IFP launched its Transit Abundance Playbook: fifteen memos on why the US pays the world’s highest prices to build transit and how to bring costs down, including on bus procurement. (June 24)

    Energy

23. The House passed a bipartisan geothermal package with NEPA categorical exclusions and faster leasing. But the binding constraints (resource characterization and transmission) remain unaddressed. (June 10)

24. Travis Kavulla explains why more places will look like PJM as data center demand rises: the grid has little excess capacity, and new demand requires costly new investment. (May 28)

25. Matt Yglesias argues “conventional environmentalism” is too focused on efficiency and missing the upside of truly abundant energy. We think cheap energy helps, but some sectors have biotech or materials bottlenecks energy alone won’t solve. (May 15)

26. The President’s FY27 budget proposes 11% cuts to DOE civilian energy programs, and the department has been slow-walking grants it already committed to and slowing new solicitations that keep appropriated funds moving. (April 17)

27. Data centers aren’t the only new large loads: heavy industry is increasingly worried about being outbid for scarce generation resources by data centers willing to pay above-residential rates. (June 3)

28. Zane Kasher’s job market paper finds a 12% drop in home prices within 3 miles of a wind farm. Aesthetic responses vary by country, and blinking lights are a big but cheaply fixable driver of annoyance. (April 24)

29. The energy-growth link is less clear-cut than you’d expect: better energy sector productivity may not affect GDP much, but eliminating power outages could matter enormously and growth effects may be asymmetric. (May 8)

30. LBNL updated its Retail Price Trends and Heatmap/MIT launched an Electricity Price Hub with utility-level data. Price patterns remain fragmented and geography-specific. (April 3)

31. Loudoun County gets 45% of its revenue from data centers, meaning lower property taxes, better schools, and general abundance. But there’s a flip side: residents are fighting the transmission lines needed to power them, and the county chair says people no longer care about the tax benefits. (May 8, June 10)

32. Semianalysis published a comprehensive analysis of space-based data centers. Base case: they don’t reach cost parity with earth until ~2040, but we shouldn’t completely discount the idea. (June 24)

    Innovation Policy

33. OMB published a 412-page proposed revision to federal science funding guidance. Concerns include political pre-award review, mandatory termination clauses, and sweeping foreign collaboration bans. Comments due July 13. (June 3)

34. NSF launched the $1.5B X-Labs initiative: decade-long, milestone-driven funding for independent research teams working outside traditional university structures. A small bet (~1% of NSF budget) but an exciting one. (May 28)

35. A new RFI floats capping PIs at two to four grants. The last attempt (in 2017) was dropped within a month. Supporting younger scientists is the strongest case; responses due August 3. (June 17)

36. The NSF FY27 budget proposes a metascience unit and $50M for Tech Labs (exciting), but also proposes eliminating social sciences and cutting funding 55% (not exciting). Congress largely ignored a similar proposal last year. (April 10)

37. The NIH indirect cost cap fight is over (for now): the administration let the Supreme Court deadline pass. Jeremy Berg’s explainer on how IDC rates actually work is worth your time. (April 10)

38. New medical phrases increasingly come from more specialized scientists and are introduced in less accessible language, both of which slow idea diffusion. Our living literature review program was motivated by exactly this concern. (June 25)

39. Three reads on what happens when AI gets good at producing science’s proxies (theorems, papers) before the processes they stand in for. We highlight Bessis on math, The Paper Factory, and Ted Chiang’s prescient 2000 short story. (May 15)

40. Many funders say they support “high-risk, high-reward” research, yet science keeps getting more conservative. Standard review processes aren’t structured to evaluate risk, and HHMI’s selection model shows promise. (May 8)

41. Ben Reinhardt argues the Vannevar Bush era of “free play of free intellects” is ending. Researchers should think about who their customers are, and independent research labs need to prove they can deliver superior quality. (April 17)

42. Eroom’s Law — drug R&D costs keep going up — may persist not because of science but because only a few large firms can commercialize drugs. Innovation is rate-limited by their pipeline capacity, not discovery. (April 10)

43. Two very different visions for reforming science funding: Gibson in City Journal wants lotteries, scouts, and indirect cost caps; Fineberg in PNAS wants sustained federal investment and talent pipelines. Both want more support for young scientists. (April 3)

44. Grant Witness is now tracking NIH and NSF funding curves in real time which is essential for monitoring whether agencies are spending what Congress approved. (April 3)

45. arXiv’s CS chair announced year-long bans for authors submitting papers with unchecked AI content like hallucinated references. Holding authors accountable for accuracy seems worthwhile, even if the post-ban peer review requirement goes too far. (May 28)

46. Refine, an AI review tool for economics, published a benchmark study, winning ~90% of head-to-head matchups against frontier LLMs. More tool-builders should run and publish studies like this. (June 24)

47. Emergent Ventures announced new grants for metascience policy entrepreneurs and science communicators. (April 17)

    Clinical Trials

48. Dario Amodei’s latest post focuses heavily on clinical trials. AI will flood the regulatory pipeline with drug candidates, so we need major reforms to make sure therapies don’t languish for years. (June 17)

49. ASCO 2026 highlights: Daraxonrasib doubled survival for metastatic pancreatic cancer and lorlatinib showed 55% of ALK-positive lung cancer patients still progression-free at 7 years. On the screening side, the NHS Galleri blood test quadrupled cancer detection but mostly caught stage III rather than truly early. (June 4)

50. FDA Commissioner Marty Makary resigned after friction with RFK Jr. He announced promising ideas but also oversaw departures of career scientists and eroding evidentiary standards. (May 15)

51. Joe Lonsdale’s FDA reform agenda pushes promising ideas: surrogate endpoints, Australian-style Phase 1 trials, and eliminating wasted time between trial phases.(June 3)

52. The FDA announced “Real Time Clinical Trials” but the rollout was muddled. Adam Kroetsch writes it’s genuinely promising but requires the FDA to define how they use live data and build interoperability standards. (May 15)

53. We might finally get a vaccine against Lyme disease but the phase 3 confidence intervals were wide and the trial missed its primary endpoint. We had one in the ‘90s and lost it to unfounded fears. (April 3)

54. Drug development fails 95% of the time. Abhishaike Mahajan explains the financial engineering tricks, from hub-and-spoke companies to synthetic royalties, that keep biotech money flowing despite the brutal odds. (May 8)

55. In Development has a great article on pooling regulatory reviews across African countries so manufacturers don’t have to file separately in each one. The African Medicines Agency launched in 2025 but faces big challenges. (May 28)

56. Niko McCarty ran a bounty to surface ideas for cheaper biology experiments: 430 submissions, 20 prizes awarded. Ideas included a protein printer from DNA origami and gel-filtration protein synthesis. (April 10)

57. Saloni discussed clinical trial reform on the Works in Progress podcast, from why ethics reviews are inefficient to how Australia made earlier-stage trials simpler and faster. (April 24)

58. Niko McCarty and Saloni wrote a biotech/medicine roundup: organ donations are vastly more efficient, semaglutide failed for Alzheimer’s, and a transformative pancreatic cancer drug succeeded in phase 3. (April 24)

    Europe

59. Paul Krugman says Europe is keeping up with the US; Luis and Pieter Garicano say it’s falling behind. The biggest question: does US tech dominance actually matter for European living standards? (May 15)

60. European tourists visiting the US for the World Cup are going viral with their own “driving around tests“, though free hotel rooms and celebrity meet-and-greets don’t come standard with road trips. (June 17)

61. Philippe Aghion identifies Europe’s gaps: no single market, insufficient long-run research funding, no DARPA-like agencies, a culture that doesn’t celebrate risk-taking, too much red tape, and more. (June 17)

62. Samuel Hughes asks whether European housing politics should be Americanized. YIMBYism is effectively nonexistent on the continent, despite housing shortages arguably worse than America’s. (June 17)

63. Kelsey Piper and Alexander Kustov explain why America is better at immigration than Europe: flexible labor markets let immigrants work, integrate, and build public support. (April 3)

64. Europe 2031 paints a scenario where AI access becomes nationalized and Europe’s assumption that it can free-ride on US tech breaks down. Eerie timing, published shortly before the US banned foreign access to Anthropic’s latest model. (June 17)

65. The ERC announced stricter application rules for 2027, likely in response to AI-generated proposals flooding review panels. NIH also capped at six applications per year. The optimal solution is an open metascience question. (April 24)

66. The Centre for British Progress released a paper on street votes: residents collectively hire an architect, agree on upzoning, and vote to allow new housing with value capture to homeowners. (April 24)

    Team Announcements

67. Alex Armlovich published Zoned Capacity Is Like an Artificial Oil Deposit, on why cities can claim they have room for thousands of new homes on paper and still not build them, and

68. Alex Armlovich published Tokyo Land Is Still >$85 Million an Acre, looking at what happens to land values when a megacity achieves housing abundance.

69. Matt Clancy published How do you get Abundance and Growth?, an outline of the Abundance and Growth fund’s plans for what kinds of work to fund in 2026.

70. Matt Clancy published A Little Progress Is Worth a Trillion Dollars, with a web tool and interactive explainer for valuing different growth scenarios.

71. Matt Clancy published An Atlas of Innovation, on how to decide when to use a prize, a research grant, or something else.

72. Matt Clancy published Social Science at the NSF, on how history rhymes with today’s debates about federal science funding.

73. Jordan Dworkin published US Science Agencies Have Money; Can They Spend It?, on why and how to keep track of government spending on R&D.

74. Jordan Dworkin published Our Grandchildren’s AI-Science Bottleneck, on why slow fields may determine the future pace of discovery.

75. Dylan Matthews published How Do States Share Ideas?, a brief introduction to the study of policy diffusion.

76. Saloni Dattani published Why Were Covid Vaccine Trials So Fast?, on how the timeline to develop coronavirus vaccines blew many predictions out of the water.

77. The Abundance and Growth team published What got us into abundance and growth, about the works that were personally foundational to our interest in abundance.

    Writing from our Grantees

78. Witold Więcek co-authored a JAMA perspective on the FDA’s Bayesian guidance.

79. IFP published analysis of Section 901 of the ROAD Act, a prevailing wage report, and the Transit Abundance Playbook.

80. Ruxandra Teslo wrote on how the pancreatic cancer breakthrough came about.

81. Lauren Gilbert published “Immigration and House Prices“ in her Migration Living Literature Review.

82. Ben Holland joined David Roberts on the Volts podcast to discuss why climate funders don’t fund housing.

83. The Foundation for American Innovation hosted the Energy Imperatives Summit in June.

    More Abundance

84. Marc Dunkelman and Alex Mechanick grapple with how democracies make decisions when people disagree, and why our current processes put too little weight on growth. (May 8)

85. The USPS owns 8,500+ properties. Rebuilding at surrounding density levels could unlock 117,000 units and earn nearly $1B/year in leases. State capacity and housing supply in one policy? (May 28)

86. The Jones Act waiver data is in: 45 voyages from 35 ships in just two months, mostly moving oil from the Gulf Coast to the West Coast and Puerto Rico. That’s a lot of pent-up demand. (May 28)

87. Sam Peltzman documents a big negative break in American happiness starting around Covid that has persisted well past the pandemic. Both GSS and Gallup data confirm well-being remains near unusual lows. (April 10)

88. 31% of World Bank policy reports are never downloaded, but the bank still writes them. Todd Moss argues this is a tragic waste of talent that could be solving actual problems. (April 10)

89. The Census Bureau is debating noise infusion for anonymizing data. A Trump Commerce Department proposal to ban it entirely could lead the Bureau to just not release certain data at all. (June 24)

90. Charles Mann and Virginia Postrel launched Everyday Abundance, a new podcast. First episode: the history of brushing your teeth. If anyone can make that interesting, it’s these two. (June 10)

91. In Development launched with a piece on cash transfers. A dollar is worth roughly 250x more to someone in extreme poverty. Harvard’s research ethics committee nearly killed GiveDirectly’s first RCT on grounds that giving people money might harm them. (April 10)

92. A big FRI survey asked AI experts, economists, superforecasters, and the public to predict AI’s economic effects. Surprisingly, the groups weren’t far apart, and all fall well short of “double-digit growth” forecasts. (April 3)

93. Alex Imas on the rise of “relational work”: as AI makes information cheap, irreducibly human elements — trust, presence, judgment — become what people spend more on as they get richer. (April 17)

94. The Labor Automation Forecasting Hub collects predictions on everything from specific jobs to specific states. The forecasts will be wrong a lot, but the clarity is appreciated. (April 24)

Most people didn’t believe it was possible to develop vaccines in the timeframe needed for them to be useful. Some looked at past vaccine timelines, which had averaged roughly 8 to 12 years, and thought this one would be similar. Others thought that, even though it was an emergency situation, it would still take at least a year and a half, or two years, or maybe even four.

I had a different conclusion. In a piece I wrote in the summer of 2020, I explained why I believed that vaccines would most likely arrive within a year of the beginning of the pandemic (placing a 58% probability on enough doses for 25 million Americans being approved and available between October 2020 and March 2021, with my central estimate landing around February 2021).

We now know how the timeline panned out and my forecast was, if anything, slightly too pessimistic, since vaccines first became available three months earlier, in December 2020. In this post, I want to go into more detail and take a look back at what happened. Why were Covid vaccine trials so fast?

Subscribe now

The scientific foundation

One of the most surprising things about Covid vaccines is how many of them there are: mRNA vaccines (Pfizer and Moderna), viral vector vaccines (Oxford-AstraZeneca and Johnson & Johnson), protein subunit vaccines (Novavax), inactivated whole virus vaccines (Sinovac and Sinopharm), and others. Although their protection dropped as the virus evolved into new strains, I take this bounty of options as a result of the coronavirus being relatively easy to develop vaccines for.

Part of the reason is the disease itself. With some infections, like HIV, no one clears the virus naturally or develops lasting immunity to it, so it’s hard to know what a vaccine should mimic. Covid-19 was very different: it was evident early on that most people recovered and developed antibodies that could neutralize the virus. This suggested it was possible to prime the immune system with a vaccine.

Another reason was prior research into coronaviruses. Work on SARS and MERS from earlier coronavirus outbreaks meant scientists already understood some of their features. They had identified the key antigen the immune system reacted against – the spike protein – and had animal models and laboratory assays ready to go. Some vaccines against animal coronaviruses had already been developed. Candidate vaccines for the earlier SARS virus had also been developed for humans (and shelved after the 2003 SARS epidemic was contained). When SARS-CoV-2 arrived, lots of this groundwork could be picked up.

There’s another feature that explains why some vaccines ended up highly effective: a stabilized spike protein. The spike protein gives the coronavirus its crown-like appearance, and is the main ingredient in most Covid vaccines. But interestingly, it changes shape during an infection. Before it fuses with a cell, it sits in a compact “pre-fusion” form (on the left, below); afterwards, it springs into an elongated “post-fusion” form (on the right). The former boosts immunity most because it’s what antibodies generally encounter, before the virus has entered cells. But when the spike protein is isolated for a vaccine, it tends to collapse into the later, post-fusion form.

Over the past decade, advances in structural biology, especially in cryo-electron microscopy, have helped scientists see the pre-fusion shape of coronaviruses directly. With that knowledge, they introduced a few mutations to stabilize the spike protein in its pre-fusion form, better for the immune system to recognize. This is a big reason why vaccines with the stabilized pre-fusion protein (including the mRNA vaccines and Novavax) generated much stronger neutralizing antibody responses and were likely much more effective than vaccines without the stabilized form (including the AstraZeneca and Sputnik V vaccines).

All of these features meant the chances of a vaccine succeeding were somewhat high. They’re also why they could be designed very fast – of course, with a lot of work on the part of scientists. It took two days to design the Moderna vaccine, which then spent 63 days in preclinical testing before entering clinical trials, versus a weekend and about 100 days for the Oxford/AstraZeneca vaccine. But the vaccines would still need to be tested in clinical trials and gain regulatory approval afterwards, and this process usually takes several years.

Parallel trials

The timeline for vaccine development typically goes like this. A candidate vaccine is first designed and tested in the lab and in animals, as ‘preclinical testing’, which takes around 1.5 years on average. Next, in ‘phase 1 trials’, it’s tested in a small number of participants for basic data on safety and the immune response, taking around 2.5 years. Phase 2 trials expand this to hundreds of participants to refine the dose and gather more data on immunogenicity and safety, and take around 3 years. Then phase 3 trials test efficacy and safety in hundreds or thousands of people, and take around 2.5 years. After all these stages, all the data is submitted to the regulator as a package, which reviews it, usually taking another year.

Historically, successful vaccines have taken around 10.7 years to make it through the whole pipeline. Each phase happens sequentially and functions as a ‘go / no-go’ system, where candidates are dropped if they flop at a stage.

Covid vaccine trials took place quite differently. Rather than running sequentially, the phases ran in parallel, as you can see in the diagram below. For example, phase 1 and 2 trials were combined, collecting the relevant data for each at once, or phase 2 and 3 were combined in the same way. Ultimately, the same amount of data was collected in a shorter timeframe, by overlapping the phases rather than doing them one by one.

This was a major reason it was possible to condense the whole timeline so much. But even then, the trials still ran much faster than usual. For example, Pfizer’s combined phase 2/3 trial then took about 4.5 months, but it usually takes an average around 2.5 years for phase 3 trials alone. What made that possible?

Subscribe now

Easier recruitment

One of the biggest delays in clinical trials is the process of recruiting enough participants to run them at all. Only half of trials meet their original target for recruitment, and a fifth revise their target downward. But the Covid vaccine trials recruited huge numbers, fast. The phase 3 trials for Moderna and Pfizer enrolled around 30,000 and 44,000 participants respectively, making them some of the largest clinical trials in history.

I think there were a couple of reasons this was easier than usual. One was that people were simply more willing to volunteer. It was a pandemic after all, and everyone’s attention was on it, plus, many people were stuck at home without much to do anyway. Interest in volunteering for trials was very high. Dr Jim Kublin, who helped run a clinical trial network for Covid, recalled: “My colleagues were enrolling a phase one study for the Moderna vaccine, and they had like 10,000 people sign up on their website. They needed 40 people to enroll.”

The infrastructure to recruit participants was also more streamlined than usual. The National Institute of Allergy and Infectious Diseases (NIAID) created a clinical trial network with a shared volunteer registry called COVID-19 Prevention Network (CoVPN), from four existing NIAID trial networks, to supplement recruitment into trials. Volunteers could sign up with a short screening questionnaire to the network as a whole and be routed to whichever trial fit. The registry helped support recruitment for the late-stage trials of Moderna, AstraZeneca, Janssen, and Novavax, and by autumn 2020, almost half a million Americans had signed up through it to take part in Covid vaccine trials. Meanwhile in the UK, a similar shared volunteer registry recruited participants through the NHS. Over half a million volunteers signed up in total, feeding into trials including those for Novavax and Oxford/AstraZeneca.

Covid was Everywhere All At Once

It also took less time to test new drugs and vaccines because the disease was spreading rapidly. It’s challenging to run trials fast when diseases are rare or slow to develop. Taking a simple example, imagine a trial where 100 people are given a placebo and 100 people given a vaccine. If only 1 or 2% of them (so 1 or 2 people) caught Covid during the trial as the base rate, it would be very hard to spot a reduction from that already-low number. Even if vaccines reduced infections by half, you’d hardly be able to tell in this example, not with much confidence anyway. Without a much larger sample size, a difference between the two groups could simply reflect noise.

This reflects the ‘statistical power’ of the trial, and that it’s easier to figure out whether a vaccine is protective if new cases are arising quickly (i.e. the disease has a high incidence rate) or if the sample size is very large. You can see this in the model I made below. It’s also why it’s much harder to develop new vaccines for diseases we’ve reduced massively: it would be a struggle to test new polio vaccines in rich countries that have eliminated the virus, for this reason.

Fortunately for trial statisticians1 – but unfortunately for everyone else – Covid was spreading quickly. During Pfizer’s trial, the monthly incidence rate (the % of the population infected for the first time, per month) was around 0.5%. By September 2020, using seroprevalence data, the CDC estimated that up to 23% of people had already been infected in the hardest-hit areas, and just below 10% across most of the US.2 In this situation, what matters most is how fast new cases are accumulating while a trial is running: during the autumn and winter of 2020, outbreaks were developing fast, and cases were racking up quickly in the placebo group, which made it easier to see a reduction in the vaccine group. Within a few months, Moderna’s trial recorded 185 Covid cases in the placebo group versus 11 in the vaccinated group, and Pfizer’s trial recorded 162 versus 8.

Subscribe now

Emergency Use Authorization

People often point to the ‘Emergency Use Authorization’ pathway as the main reason Covid vaccines were introduced so fast – that they weren’t approved the regular way, and faced lower standards of evidence than usual. Let’s dig into this a little.

In general, an emergency use authorization (EUA) is a conditional authorization that lets the FDA allow a product to be used during a declared public health emergency before it has gone through full approval, and which can be revised or revoked at any time. And the bar is intentionally low: the FDA only has to find that the product ‘may be effective’ and that its known and potential benefits outweigh its known risks, given that there is no alternative that’s approved.

The standard of a product that ‘may be effective’ is much weaker than the evidence of effectiveness required for full approval. One example was the emergency authorization of hydroxychloroquine as a treatment for Covid in March 2020, which relied on observational data rather than efficacy trials. A few months later, when the RECOVERY trial showed, in a randomized study, that it had no benefit, its authorization was pulled.

Does this mean that the standards were so relaxed for vaccines? Not really. The vaccines faced a stricter EUA, called ‘EUA plus’. In its June 2020 guidance, the FDA demanded large randomized controlled trials demonstrating an efficacy of at least 50% before a vaccine could be licensed. Then in October, it added more conditions an EUA would have to meet: a median of at least two months of safety follow-up after the final dose, and review by an independent advisory committee in a public meeting.

This meant that the standards for Covid vaccines were similar to a full approval, with two notable exceptions. 1) While full approval required a complete manufacturing dossier, with validation and facility inspections, the EUA accepted a less complete package, and deferred the validation and inspection until later. 2) While full approval required a longer safety and durability record, typically at least six months, the EUA accepted a two-month median follow up time after the final dose.

It’s worth noting that the vaccines did fulfill the requirements of a full approval later on, starting with Pfizer’s vaccine in August 2021.

Incidentally, the EUA’s shorter follow-up requirement was the subject of political controversy. When Trump complained that the vaccine was being delayed until after the election, it was because of the follow-up requirement, which wouldn’t be complete until later – Pfizer only reached the median two months of safety follow-up in the third week of November, over two weeks after the election, and filed its EUA request on November 20.

I don’t have strong views on the requirement, but it seems hard to strike a balance between collecting enough safety data to build public trust versus speed up the roll-out, and unlike all the other reasons for acceleration, this was the only way in which the safety threshold was laxer than a full approval, which was achieved later on. It seems like the crux is how important public trust would be in increasing vaccination rates and whether the requirement achieved that.3

Rolling regulatory review

Typically, when researchers finish their trials, they compile a complete dossier with lots of information about the trial, the product, the manufacturing process, and so on, and submit it to the regulator, which then reviews the complete package.

It usually takes a while to review the complete package. In the US, the FDA’s goal is to review the data and make a decision within 10 months of receiving the application, or 6 months if it qualifies for ‘priority review; in practice it takes an average of eleven months. Meanwhile in the EU, the European Medicines Agency’s goal is 210 days (roughly 7 months), but the clock only includes the days the agency is actively reviewing; it pauses whenever the regulator sends questions back to the company and waits for answers, so in reality, a review usually stretches well beyond a year.

The Covid vaccines faced a different process. In Europe, they faced a ‘rolling review’ process instead, where regulators assessed the data – on the manufacturing, preclinical research, trial results, etc. – as it arrived, instead of all together right at the end. So by the time the final data was in, most of the review had already been done, and it took much less time to reach a decision. Pfizer submitted its formal EU application on 30 November 2020 and received conditional authorization about three weeks later, on 21 December, because the rolling review had already evaluated the data as it came in.

In the US, the rolling review process applied to the full approval, rather than the initial authorization, since the vaccines went through the Emergency Use Authorization pathway instead, which began on 20th November for Pfizer and granted approval on 11th December 2020. The complete application, which did involve rolling review, was completed in May 2021 and received full approval on 23 August 2021, roughly three to four months, still much shorter than usual.

Subscribe now

Can we apply some of these processes to speed up other drugs and vaccines?

By this point, you’re probably wondering why the same processes couldn’t apply to many other drugs and vaccines. I’ve wondered the same thing, and I think it’s true that various parts of this process could apply elsewhere too – such as clinical trial networks to recruit participants, more fundamental research into pathogens and candidate vaccines, and better trial design. (The Covid experience is a reason I became interested in clinical trial reform more broadly.)

But some parts would be challenging to scale up more widely. Rolling regulatory review, for example, is more demanding on the part of regulators. In the usual pipeline, most drugs or vaccines fail before they reach the finish line and are submitted for review, which means regulators are saved the time of reviewing lots of applications that would have been withdrawn anyway. (You could, of course, expand rolling reviews by hiring and training more regulatory staff, and reduce the trade-off!)

Parallelized trials are also challenging. When trials are run sequentially – with phase one, phase two, and then phase three – the sequence functions as a ‘go / no-go’ system, where developers can drop candidates they think are unlikely to succeed at the next stage. It saves time to run trials in parallel, but it also raises financial risk: instead of spending money on the next phase only once the previous one has paid off, a vaccine developer would have to commit to spending it up front.4 A similar problem applies to building production capacity during the trials, since that capacity would be wasted if the vaccine doesn’t work.

This would have been especially financially risky during a pandemic, because of the degree of uncertainty at the outset. What if the virus had actually been contained by the time the trials were complete, as was the case during the 2003 SARS outbreak? Or what if the fundamentals of the virus were misunderstood and the vaccines were designed incorrectly? By the time such large trials were complete, a lot of money might have been sunk into trials that would ultimately fail. This is where Operation Warp Speed comes in: in my view, it made it possible for a lot of the processes above to function at a lower financial risk, and in a more coordinated way.

Operation Warp Speed and its analogs

Operation Warp Speed was launched in May 2020 by the Trump administration, as a partnership between the Department of Health and Human Services and the Department of Defense. Its goal was to produce and deliver 300 million doses of safe and effective Covid vaccines, with the first doses becoming available by January 2021.

To make it happen, Operation Warp Speed funded the development of a range of candidates and their clinical trials, including the phase 3 trials of Moderna, Janssen, AstraZeneca, and Novavax.5 The CoVPN trial network was developed as part of Operation Warp Speed. Operation Warp Speed also paid for companies to start manufacturing their vaccines at industrial scale while trials were still running, so that millions of doses would be ready to ship once vaccines were authorized.

Importantly, it also committed in advance to buying hundreds of millions of doses of a vaccine that succeeded, meaning that the companies wouldn’t face a massive loss if the outbreak ended up petering out before vaccines were ready. This idea, called an advance market commitment, was originally developed by economists Michael Kremer and Rachel Glennerster to successfully speed up new pneumococcal vaccines in the 2000s, before it was adapted for the pandemic. As then-secretary of Health and Human Services, Alex Azar, described it, the government would invest more than 10 billion dollars to “de-risk vaccine companies’ development efforts and manufacture product in advance.” By the end of the year, it had committed around 13 billion dollars.

As I described earlier, running trials in parallel and manufacturing before trials are complete makes things faster but is financially risky, because a company commits money up front that would be wasted if the candidate vaccine failed. Operation Warp Speed absorbed much of this risk. (One thing that Operation Warp Speed didn’t do was change the regulatory bar: the EUA and rolling review procedures were set by the FDA separately.)

Other countries also used analogs to Operation Warp Speed to accelerate vaccine development and secure doses for their populations. The UK, for example, ran a Vaccine Taskforce, set up in spring 2020 and led by Kate Bingham, which also backed a spread of candidates across different technologies, made advance purchase agreements, and funded manufacturing. The EU ran a Vaccines Strategy, launched in June 2020, which used advance purchase agreements as well: the EU Commission financed part of companies’ upfront costs in return for the right to buy doses. Since it negotiated for all 27 member states, it had a lot of bargaining power, but it started much later; its first major agreement was signed months after the UK and US. And lastly, COVAX created a pooled procurement mechanism for countries around the world, aimed to secure doses for lower-income countries that couldn’t make their own deals (though they fell below their target of later delivering vaccines, delivering half of their target of 2 billion doses by the end of 2021 as high income countries struck bilateral deals quicker).

Operation Warp Speed-type efforts probably work best for large problems that need coordination across many actors, where the goal is well-defined and commercial markets won’t deliver it at the speed required, because the private returns are too small relative to the social value. But it’s also possible to work faster across the board, by creating clinical trial networks and shared volunteer registries to speed up recruitment, innovative trial designs that collect data more efficiently, investment into early-stage research, more regulatory capacity to review evidence as it arrives, and many other ideas we’ve shared on this blog.

Subscribe now

Faster trials saved millions of lives

The speed of vaccine development during the pandemic was unprecedented, but it wasn’t unforeseeable. It came from a coordinated effort that involved dedicated research funding, clinical trial networks, parallel trial phases, rolling regulatory review, and parts of the review being deferred rather than skipped. What eased the process were decisions to de-risk the process financially, with upfront funding and advance commitments to buy the vaccines if they worked.

The result was extraordinary. Millions of deaths globally were averted by Covid vaccines in just the first year of vaccination alone.6 We take it for granted today, but it was not a given that trials would be completed within a year. Nor was it a given that the vaccines would have around 95% efficacy against the original strain of the coronavirus, which was much higher than most scientists anticipated: the median forecast was 50%. Speeding up the process saved millions of people.

When it comes to clinical trials, it’s easy to assume the process is so bureaucratic that we couldn’t move quickly even if we wanted to. The pandemic shows that is not the case. When the need was obvious, thousands of people around the world – including scientists, operators, policymakers, regulators, economists and government officials – worked differently to make sure vaccines became available faster.

It also reveals how slow the current system is. If clinical trials can be sped up so much without compromising on safety, we should find ways to reform them outside of pandemics as well. Every year of delays in testing is a year patients go without potentially lifesaving treatment, risking their lives. We shouldn’t be dragging our feet for the sake of following the status quo. Rather than treating pandemic trials as exceptional, we should ask whether that speed should have been the norm all along.

It’s been six years now since the publication of Ideas are getting harder to find, a famous paper in the economics of innovation which documented that vast increases in research effort are not resulting in materially faster rates of discovery (since then, a number of other papers point to similar conclusions). Why not? My preferred version is that as sciences mature, a variety of stresses emerge. A partial list of those stressors is that, as a discipline matures…

• Ideas that can be quickly or cheaply discovered are the first to be found

• Further progress requires marshalling more detailed knowledge, which requires more years of training to acquire, and older scientists have a harder time with some kinds of breakthrough innovation

• Marshalling more knowledge also increases reliance on research teams, and teams have a harder time developing breakthroughs than individuals

• The requirement to bring more knowledge to bear on outstanding problems, leads people to specialize more. But this impedes making connections across different fields, which is an important part of innovation (see Jordan’s latest blog post on this, in the context of AI that differentially accelerates different fields!)

The last point - that specialization impedes the diffusion of ideas - has seemed likely true to me, but I haven’t ever seen it measured, until this week. A new paper by Gaetani and Berkes tackles this measurement problem in medicine. To measure the spread of ideas, they compare the abstracts of papers that mention a medical phrase in the first year in which it was coined, and then in the subsequent ten years. They then use natural language processing to measure how different the abstracts are that first mention a concept and the ones that use it in later years, with the idea being that a greater difference between the abstracts implies the idea is being used in fields that are further afield from where an idea was developed.

Gaetani and Berkes’ paper suggest that specialization has made it harder for ideas to spread: they find new medical phrases increasingly come from scientists who have worked in fewer fields, and that phrases originating from more specialized scientists don’t diffuse as far. But they also document a dynamic that compounds this issue: over time, scientists have been introducing new ideas in less and less accessible language.

Our living literature review program was motivated in part by precisely these concerns - accessible literature syntheses could help ideas diffuse more quickly over time.

Here’s what caught our attention over the last week:

1. The ROAD Act and the Transit Abundance Playbook — Alex Armlovich

2. A new threat to economic data — Dylan Matthews

3. Refine sets the benchmark for AI review — Jordan Dworkin

4. Remember data centers in space? — Willow Latham-Proenca

5. Operation Trial Blazer — Saloni Dattani

6. The limit of human felicity — Matt Clancy

We’re also planning to post a separate What we’re reading spotlight by Matt tomorrow. Stay tuned!

The ROAD Act and the Transit Abundance Playbook — Alex Armlovich

The Senate passed the 21st Century ROAD to Housing Act 85–5 on Monday; the House takes up the amended package today, and President Trump is expected to sign it within the week. The bill’s 45-plus provisions cover a wide array of reforms: cutting red tape, overhauling funding and financing, and providing technical assistance plus carrots and sticks for state and local reform. The supply-oriented measures trim federal permitting friction, expand the definition of manufactured housing by relieving the HUD code chassis mandate, raise and index FHA multifamily loan limits, lift the bank public-welfare-investment cap from 15% to 20%, and stand up grant programs to convert vacant buildings and rehab aging homes. Most of the binding constraints on housing still live at the state and local level, but this provides direct cleanup of key paperwork burdens and bad rules that exist at the federal level.

IFP just launched its Transit Abundance Playbook: fifteen memos from transit practitioners and researchers on why the US pays the world’s highest prices to build and run transit (and how to bring those costs down). Rohan Aras and I wrote on bus procurement: US agencies pay anywhere from $500,000 to over $1 million per bus, while peers abroad buy comparable vehicles for under $400,000. Bad incentives in federal-local cost-sharing compound with excessive customization at the agency level…plus burdensome procurement regulations, bans on trade in buses even with US allies, and weak federal state capacity, to produce the world’s least competitive & most expensive domestic bus market.

A New Threat to Economic Data — Dylan Matthews

Economists, sociologists, demographers, and other researchers need accurate, ideally individual-level data to understand what’s going on in the world. At the same time, most people don’t want information like their address or income available for everyone to see. How do we reconcile these two goals? Some countries, like Norway and Sweden, err on the side of openness, making even individual income tax records public. The US typically values privacy highly, which means the Census has begun to use “noise infusion,” a technique for anonymizing individual and business-level data by adding random changes to individual data points. At a high level, the random noise cancels out, in theory letting you still learn accurate facts about, say, a county or industry, without violating specific people or businesses’ privacy.

When the Census announced they would do this to the 2020 Census data, leading to strange conclusions like finding that 48 people live on the same island as the Statue of Liberty (in reality zero people do, at least legally), there was a major backlash and the agency eventually backed down. Now, the Trump Commerce Department is proposing to ban any use of noise infusion. Nathan Goldschlag, a former staff economist at the Census Bureau, has a post at the Economic Innovation Group’s blog arguing this goes much too far and risks leading the Bureau to just not release certain data at all. I am far from an expert on this and honestly think the US tends to overweight privacy in these matters, but it’s the kind of debate that could have far-reaching ramifications for how we understand the economy.

Refine sets the benchmark for AI review — Jordan Dworkin

There are a growing number of AI tools for various components of the scientific process, but there are rarely careful, public evaluations of how they perform and compare. Last week Refine, an AI-powered review tool for economics research, published the kind of benchmark and evaluation study that I would love to see more of in this space. They ran ~1300 head-to-head matches on 150 economics preprints, comparing the quality of Refine’s reviews to both single-shot frontier LLM reviews and scaffolded review systems. Refine won roughly 90% of the match-ups (~95% against single-shot models and ~85% against scaffolds). Worth caveating that the benchmark’s unit of evaluation is a “paper-grounded atomic concern”, a specific verifiable issue tied to a location in the paper; this is a reasonable operationalization of review quality, but is also one that informed Refine’s design. Refine’s win rate is also strongest in matchups where it raises more unique concerns than the competitor (~90-95% when it has more concerns vs ~70% when it has fewer) - that could very well be signal, but it could also be an evaluator bias towards longer/more detailed reviews (a preference that has been observed in human evaluators), which would be both useful and interesting to check. But overall, it seems clear that one major takeaway is “Refine is very good at producing high-quality, technical feedback, and if you’re an economist you should consider giving it a try”, and another is “more tool-builders should run and publish studies like this.”

Remember data centers in space? — Willow Latham-Proenca

(...probably). Semianalysis is out with a comprehensive comparative analysis of the costs and uncertainties around space-based datacenters, arguing that while technical, operational, and cost barriers mean that space-based data centers start out at at least 4x earth-based costs, we shouldn’t completely discount the idea. They’re not exactly bullish on timing, though – in their base case, space doesn't reach cost parity with earth until ~2040, driven by falling launch and hardware costs. Even then, space would be a genuine necessity only if earth-based data centers have exhausted all potential capacity from grid supply, BTM generation, converted bitcoin mining, and industrial capacity/manpower to build additional infrastructure, which their base case doesn’t show happening at a meaningful scale through the middle of the century. However, what’s interesting is how quickly regulatory barriers (and capacity barriers dressed up as regulatory barriers) could potentially start to bite – by the early 2030s, space-based data centers could be running at only a ~30% premium, easily in the range where permitting and regulatory barriers become the deciding factor on whether to launch if earth-based capacity remains artificially constrained. Depending on how fast we solve some of the technical hurdles that currently make space-based centers a financial non-starter – as well as the universal constraint on chip deployment anywhere, semiconductor production – political will could end up being of central importance, making the most important argument for space…well, the space (and even that is limited by narrow orbital regimes!).

Operation Trial Blazer — Saloni Dattani

This week, the US Department of Health and Human Services (HHS) announced a new initiative called ‘Operation Trial1 Blazer’ that aims to bring back early-stage clinical research, which has increasingly been outsourced to countries like China and Australia, by reducing how long it takes to get approval to start phase 1 trials as well as how long it takes the FDA to review data from phase 1 trials.

The initiative includes several parts. One is an FDA pilot program to set up a network of qualified research institutions that help researchers prepare the protocol and supporting components before they submit it, though the FDA still decides whether a trial can proceed. It also provides more guidance on what manufacturing (CMC) data is needed for phase 1 and clarifies this is lower than for later-stage trials – companies have so far tended to submit more data than necessary to secure approval. And it includes guidance on choosing the first doses given to humans, recommending these are based on statistical modelling of the drug’s pharmacology rather than animal toxicology studies, which were often poor predictors of side effects in humans. Finally, there’s a rolling platform for the FDA to review components of the package as they arrive rather than all at the end.

Most of this sounds positive, though I’d add a caveat of the main bottleneck facing the FDA right now: staffing capacity, given the cuts over the last few years (roughly a fifth of the agency’s workforce, around 3,500 FDA staff, were cut last year), and whether these reforms ease that problem or worsen it further. Rolling reviews especially require more staffing time, because reviewing at the end typically means some attrition from researchers who don’t make it to that stage; similarly, if the FDA still has to make the final decision on whether phase 1 trials can proceed, especially given this initiative is new and will take their input. Still, there’s good stuff here: the guidance on manufacturing data and dose selection doesn’t depend on FDA capacity the way the pilot does, so some of these changes could result in improvements soon.

For more thoughts on the new policy, Ruxandra Teslo and Adam Kroetsch also weighed in.

The limit of human felicity — Matt Clancy

I’ve been listening to the wonderful new podcast Everyday Abundance (no affiliation!) by Virginia Postrel and Charles Mann, based on Dylan’s recommendation two weeks ago. Each episode looks at the history of some aspect of modern life, which we tend to take as just part of the background now. The last episode I listened to was about listening to music, and featured this quote from Edward Bellamy, writing in 1888:

[I]f we could have devised an arrangement for providing everybody with music in their homes, perfect in quality, unlimited in quantity, suited to every mood, and beginning and ceasing at will, we should have considered the limit of human felicity already attained.

I estimate I spent about 16% of my waking hours in 2024 listening to music.

We have just one announcement to share this week: our colleagues at Coefficient Giving launched a $10-30 million RFP on global health and wellbeing in an era of transformative AI, looking for proposals on everything from LMIC clinical trial infrastructure to how labor market disruption could reshape health systems. Applications are open through August 21.

1. The driving around test (Euro edition) — Matt Clancy

2. You should read Dario Amodei's latest essay — Dylan Matthews

3. It's 2017 again at the NIH — Jordan Dworkin

4. 40+ ways to plug in a data center — Willow Latham-Proenca

5. Should European housing politics be Americanized? — Alex Armlovich

6. Oops, I ordered the wrong antibody again — Saloni Dattani

7. Updates from Congress — Nisha Austin

The driving around test (Euro edition) — Matt Clancy

As part of the month-long debate about the extent of the European/American gap in living standards, Luis and Pieter Garicano suggested a “driving around test”: “Go to the periphery of any modern American city and see a level of new-built material wealth that is extremely uncommon in Europe.” With the World Cup underway in America, some European tourists are doing their own “driving around” tests and going viral, though these driving around tests are hardly scientific (the US may be wealthy, but free hotel rooms and meet-and-greets with our favorite singers do not come standard with road trips).

Nonetheless, for those worrying about the US-EU living standards gap, there’s plenty to read from the last few weeks. To start, Simon Johnson’s interview with economist Philippe Aghion covered a wide range of plausible factors explaining the gap: a lack of a single market, insufficient funding for long-run research, lack of DARPA-like agencies, a culture that doesn’t celebrate risk-taking, and excessive red tape. For more on this theme, see also Alex’s post below about European YIMBYs (or the lack thereof). While Aghion believes these factors are real, he pushes back when Johnson characterizes Europe as troubled:

I wouldn’t say “troubled.” Europe is not on the verge of a major crisis. Instead, the danger is a long period of stagnation or sluggish growth.

Unfortunately, the authors of Europe 2031 (the latest AI scenario microsite) worry that we are in an era when stagnation will impose heavy costs on Europe in the years ahead. One theme of the US-EU living standards debate has been on the importance of having a thriving domestic tech sector. So far, many of the benefits of the US tech sector have been shared by Europe, which, after all, has as much free access to Google and Facebook as it wants. The Europe 2031 authors paint a scenario where this assumption breaks down, and access to AI becomes nationalized as its security implications grow and compute for inference remains scarce. The microsite’s timing is eerie, coming shortly before the US government banned access to Anthropic’s latest model for foreign nationals (though, at least in this case, the ban meant that US citizens also lost access).

You should read Dario Amodei’s latest essay — Dylan Matthews

If you’re staying on top of AI news, you probably don’t need me to tell you to read Anthropic CEO Dario Amodei’s latest essay. That said … you should read Dario Amodei’s latest essay. It’s the most detailed look at the company’s public policy attitudes to date, and particularly interesting to the AGF team because of how much Amodei focuses on clinical trials. As he writes, AI seems likely to "greatly increase the rate at which new drug candidates enter the regulatory pipeline"; but that pipeline is so long and slow that it will be hard for customers to enjoy those drug candidates in a timely way. That implies that we need major reforms to the way the Food and Drug Administration and European Medicines Agency evaluate drug candidates, to make sure new, promising therapies aren't languishing for years before approval.

It’s 2017 again at the NIH — Jordan Dworkin

It’s 2017 again: the Gorillaz are topping charts, Star Wars is in theaters, and the NIH is discussing a potential cap on the number of grants a single PI can hold at once. Last week the agency published an RFI floating potential caps at two, three, or four simultaneous research project grants, with the stated goals of spreading funding more widely across labs, institutions, and researchers, and supporting more early- and mid-career scientists.

The agency’s last attempt at this – the 2017 Grant Support Index, which would’ve capped support at roughly three R01s’ worth of funding – was dropped within a month following pushback from elite universities. Last week’s RFI cites a handful of metascience papers in support (specifically showing diminishing marginal returns to additional funding, and reduced disruption as a function of team size), though in my view the metascientific jury is still out on the cost-benefit of a hard cap; getting more money to younger scientists seems like the strongest case for the cap, but cases against include diverting funding from meritorious proposals, increasing process burden, and making large-team science more difficult to sustain. Since 2017, the agency has tried a few other approaches for achieving these goals, including a dedicated fund for early- and mid-career scientists (which directly replaced the GSI when it was dropped) and a softer version of a cap at NIGMS, in which explicit justification is required for funding above a certain threshold. STAT has a good overview of the background and the various perspectives. Responses to the RFI are due August 3.

Should European housing politics be Americanized? — Alex Armlovich

"Should European housing politics be Americanized?" asks Samuel Hughes. American policy frames usually map badly onto European debates, but housing is an exception: Europe's housing shortages are arguably worse than America's, with European house prices now above US levels and roughly 80% of the postwar increase attributable to regulatory restrictions on building. Despite that, YIMBYism is effectively nonexistent on the continent, and housing arguments often default to rent control, zero-sum expropriation without new construction, and public housing. Hughes's sharpest move is historical: zoning was arguably invented in Europe, and he walks through Berlin's 1905 villa-colony scheme to show the same map performing the same exclusionary function 120 years later. I went in a little nervous about how a British writer would explain American YIMBYism abroad, but Sam’s argument was novel and deftly executed.

40+ ways to plug in a data center — Willow Latham-Proenca

It may not be the most scintillating reading for the general audience, but the Lawrence Berkeley National Lab is out with an extremely cogent, clear rundown of over 40 proposals to speed up how fast large loads – data centers, of course, but also industrial users or fleet EV charging – can connect to power. These range from the frequently-discussed (BYOG, interruptible service) to the arcane (ERIS-like service for loads, one of several proposals to adapt tools from the generation interconnection toolkit to the load side). As usual, the national lab that brought you quantitative evidence that new load doesn’t (necessarily) increase prices is calibrated on how much to panic – they compare load forecasting errors from the 1970s over-forecast to now, and find we might actually be doing better. Worth noting that proposals to speed up construction of the transmission needed to connect many of these loads to electrons – which can take up to three times as long as the planning and study process, in the report’s estimation – are out of scope here.

Oops, I ordered the wrong antibody again — Saloni Dattani

The science sleuth Sholto David recently uncovered a mix up that’s affected hundreds of papers in biology. Scientists who were intending to work with the p16-INK4a antibody, to flag a tumor suppressor linked to cancer and aging, accidentally ordered another antibody p16-ARC instead, which tags a component of the cell’s internal skeleton.

Some of the consequences were mostly harmless, where scientists ordered the correct antibody but described the wrong one in their papers. Others were unfortunate, as lots of experiments may not have worked because of the mistaken antibodies. But some are potentially fraudulent, where in some experiments, scientists reported results using the wrong antibody that could have only worked if they were using the right ones.

The whole naming confusion reminds me of another related episode in biology, where hundreds of papers were affected by an error in which the gene SEPT1 got automatically converted into the date ‘1st September’ by Microsoft Excel. Eventually, the Gene Nomenclature Committee solved the problem by simply renaming the gene SEPTIN1. Let this be a warning about the dangers of dating.

Updates from Congress — Nisha Austin

The first major federal housing bill in roughly 30 years just took a big step forward. After months of ping-ponging between the House and Senate, all four corners (the chairs and ranking members of the Senate Banking and House Financial Services committees) released unified bill text yesterday for the 21st Century ROAD to Housing Act. Senate leadership is expected to bring it to a floor vote this week. We’ll be watching.

On energy permitting, there are two bipartisan moves worth keeping an eye on. Senators Cotton and Cortez Masto introduced a Senate version of the FREEDOM Act, which would prevent agencies from revoking permits or halting construction on fully permitted energy projects, and is a direct response to the administration’s stop-work orders on offshore wind. The Senate version rolls in much of the House’s geothermal package from earlier this month and is notably cleaner than the House’s: it strips the most stringent agency penalties and moves fees into a freestanding “Permitting Performance Fund,” which should quiet concerns that the original structure would create a cost spiral for agencies. Meanwhile, Senator King added an amendment to the NDAA targeting the DOD’s increasingly creative use of flight path certification to block onshore wind projects. Neither energy bill is a sure thing, but both signal that permitting certainty has bipartisan legs even in this environment.

Here are a few other highlights and announcements from our team and grantees:

• Jordan published Our grandchildren’s AI-science bottleneck, on why the fields AI accelerates least may end up determining the pace of discovery for everything else.

• Matt published a blog post about the Atlas of Innovation, an interactive guide to choosing between prizes, grants, contracts, and other tools for funding innovation, developed with IFP and the Market Shaping Accelerator.

• Saloni wrote up What’s new in biology for Works in Progress, rounding up recent developments in the field, and covering everything from antibody mix-ups to new cancer research.

• Abundance New York launched a voter guide ahead of the city’s elections.

• California YIMBY endorsed Xavier Becerra for governor.