Issue 25 May 22nd 2026
The theme of this newsletter is Ghosts.
When we analyze what went wrong within systems of science, we do a thorough job of cataloguing failure. We track retraction rates, measure citation decay, study the proportion of clinical trials that fail to replicate. We have become quite good at auditing what we attempted and fell short of. What we almost never do is audit the omissions: the cures that were never developed, the instruments never built, the cross-disciplinary connections that were never made, because the architecture of the system made them impossible to attempt in the first place.
These are the ghosts of innovation. And unlike failures, ghosts leave nothing behind.
A failed project leaves a paper trail. It leaves a funded grant, a research team, a set of negative results, perhaps a retraction, possibly a lesson. All of that is measurable, debatable, and can eventually enter the conversation about reform. But a question that was never asked because it couldn’t survive the grant review process leaves nothing. A collaboration that was never formed because the funding structure rewarded individual labs over interdisciplinary teams leaves no record. A researcher who filed away her most ambitious idea “for later”—and later never came back—contributes her ghost to a ledger that nobody is keeping.
The mechanism responsible is not conspiracy or malice. It is what the literature calls paradigm lock: once a framework becomes dominant, it shapes not only how results are interpreted, but which questions are considered worth asking and which methods are considered legitimate. Researchers, responding rationally to the incentives, pursue work that is likely to be publishable, fundable, and recognizable to their peers. What this produces isn’t the selection of ambitious exploration—it is the selection of what is testable within the existing frame. Progress within a constrained space is real progress. But it is not the same as progress across the full range of the possible.
Venkatesh Narayanamurti and Jeffrey Tsao (both subscribers to this newsletter!), in their work on the structure of technoscientific revolutions, offer a useful illustration of how this plays out in the adjacent possible. The multitouch display technology and the specialized glass that became Gorilla Glass® both existed as latent combinations before the iPhone. They could have remained forever in what they call the “shadowy and unrealized adjacent possible” (borrowed from the biologist Stuart Kauffman)—latent solutions that never found problems to solve. It took Steve Jobs, committing to pull specific latent combinations into contact with a specific problem, to make them actual. The ghosts are all the latent solutions in that adjacent possible that never found a Jobs. The ghost framing forces us to ask: how many of those latent solutions are there, undiscovered because it is impossible for anyone to go looking?
What distinguishes ghosts from every other “failure” of science funding is that they are truly unmeasurable—and unmeasurable costs tend to be politically weightless. They do not appear in appropriations hearings, reform proposals, or reports on research productivity. The ghost leaves nothing to complain about.
Here is what this means for anyone serious about science reform. Almost every conversation in the field is about what we funded and how well it worked. Very few conversations are about the topology of the attempted: which kinds of questions were structurally possible to pursue, and which were not. If you want to change what gets discovered, you have to first change what gets attempted. And that requires an audit not of outcomes but of architecture—asking not what failed, but what the current system makes impossible to try. The reform discourse that focuses only on waste and replication is working inside the paradigm. The more demanding question is whether the paradigm itself is selecting for a particular shape of future, and whether that shape is the one we would choose.
Growth is Getting Harder to Find, Not Ideas | NBER
When the paper “Are Ideas Getting Harder to Find?” landed, it caused quite a splash. In a follow-up, Teresa Fort and her colleagues ask whether perhaps the ideas are coming, but they are not successfully running the gauntlet to realization. Compelling evidence suggests that this is true (but yet, as far as we can tell, does not dispel the original claim).
How an ‘Impossible’ Idea Led to a Pancreatic Cancer Breakthrough - The New York Times
Daraxonrasib, the first drug to meaningfully extend pancreatic cancer survival, rests on 44 years of patient academic work—from Robert Weinberg’s 1982 discovery of RAS’s role in cancer to Kevan Shokat’s 2013 cracking of the “undruggable” KRAS protein—through decades of false starts and field-wide skepticism that the effort was “lunacy.”
How PhD Programs Limit Scientific Progress
We are often the hammer that sees every nail as a science-funding problem. This piece looks to an adjacent space—the minting of new scientific talent—and suggests that there is likely low-hanging fruit (architecture) to be rethought there, too.
A multi-agent system for automating scientific discovery | Nature
In nearly every newsletter, we’ve commented on AI in science; often sounding a bit bearish on the possibilities, especially when it comes to truly novel discoveries. This paper adds some fuel to the fire, showing that a multi-agent system can propose and test a hypothesis. Even so, as the paper makes clear, the exercise was largely a case of drawing new insights from what is already known. The debate will thus rage on.
Inside the Interstitium, the Human Body’s Hidden Pathways - The New York Times
The interstitium—a third circulatory system threading through the body’s organs—emerged not from a single breakthrough but from the slow combination of century-old anatomical observations, modern imaging, and even 4,000-year-old acupuncture traditions, finally synthesized by Neil Theise and Rebecca Wells in 2018. Its implications for cancer, diabetes, and pain medicine show that transformative discoveries are often combinatorial—recombinations no near-term funding cycle would reward.