Seeing Like a Gardener: A New Framework for the Public Funding of Science

This article was previously published as part of The Reconstruction Papers. The entire book is available in print and as a PDF.

The Trump administration in 2025 took a…different approach to science than we had seen before. It took the unprecedented step of cancelling (not just refusing to renew) thousands of ongoing grants, including clinical trials. It fired staffers at science agencies like NIH and NSF, and proposed massive budget cuts on the order of 40%-50% or more. It engaged in culture-war battles with many American universities, using the threat of withholding grants to everything from cancer research to cosmology research (which had never been done before). And via the Department of Government Efficiency, it paradoxically imposed new rules that made everything more inefficient than anyone could have imagined (such as one that required advance approval when spending more than $1 on a federal credit card).¹

A future administration will have some obvious decisions to make. But if all that happens is restoring the status quo, we will have wasted an opportunity to rethink the relationship between the US government and science, one that has persisted since the post-World War 2 era.

Most scientists will privately (and sometimes publicly) agree that the pre-Trump scientific ecosystem was not entirely healthy. As our friend Tom Kalil, a science-policy veteran, is fond of saying, it’s not as if scientists thought that the system was heaven on earth as of December 2024. Long before Trump, the system was too bureaucratic, too slow, too conservative in how it selected projects, too dependent on a narrow set of institutional forms, and too dysfunctional in terms of career pathways.

To fix these problems, we need to rethink the institutional design principles we use to allocate many billions in public dollars every year.

Since the post-WW2 explosion in science funding, we have designed scientific institutions and funding from the top down. That is, we created large bureaucracies at NIH, NSF, the Department of Energy’s National Laboratories, and more, all within a broad framework (credited to Vannevar Bush’s Endless Frontier) about the relationship among government, universities, and public knowledge.

Those were acts of institutional design that, consciously or not, reflected a narrow set of assumptions about governance, incentives, careers, and what kinds of organizations ought to exist.

One of the most important books for thinking about institutional design is James Scott’s Seeing Like a State. Scott’s basic argument is that modern states often try to make complex adaptive systems more “legible” to centralized planners: easier to measure, classify, standardize, and control. But in doing so, governments can unwittingly undermine the local knowledge and organic complexity that made those systems work in the first place.

Perhaps his most famous example is the ill-fated attempt to make forestry more “scientific.” Professional foresters disliked messy, diverse forests with multiple species, undergrowth, irregular spacing, and all the seeming inefficiencies of nature. Instead, they tried planting neat checkerboard patterns of the same species of tree.

On paper, this looked more scientific and rational. It was easier to count, map, and administer. But over time, according to Scott, these rational forests often performed worse. They were more vulnerable to disease, pests, soil depletion, and other failures that the older, more diverse forest ecology had buffered against. What looked irrational from the perspective of the planner turned out to be functional from the perspective of the living system.

American science policy, we would argue, has often made a parallel mistake.

Over time, we have built a research system that is increasingly organized around legibility to bureaucracies. Scientific projects must be described in advance, broken into specific aims and deliverables, and evaluated through standardized procedures that make them easier for large bureaucracies to process.

Institutions, likewise, are forced into a strict set of administratively familiar forms (“institutional isomorphism,”² as sociologists would say). The university PI lab funded by project grants is legible, as is the top-down national initiative with an inspiring name (such as a “Cancer Moonshot”³ or a “National Plan to Address Alzheimer’s Disease”⁴).

But legibility is not the same thing as vitality. It’s the paradox of progress:⁵ attempting to manage and control future progress via top-down mechanisms can actually inhibit such progress from occurring.

Outside of certain narrow aims, science is not engineering. In other words, science is not a machine whose outputs can be reliably maximized by central planners if only the metrics are precise enough. In an ideal state, science would be closer to an ecosystem: a mostly emergent process in which real progress often comes from diversity of methods and institutions, strange side paths, local knowledge, tacit skill, and forms of exploration that may look inefficient or irrelevant at the time.

That is why the right metaphor for science policy is not top-down industrial-style planning. It is gardening.⁶

Gardeners cannot predict the exact path of every root, or force every living thing into identical form. Instead, they create the conditions under which plants can flourish. That means tilling the soil, providing water and fertilizer, pulling weeds, protecting fragile growth from the weather and from invasive pests, and introducing variety as a way of maintaining soil quality. Good gardeners try to take a strong hand in producing their desired outcomes, but they do so with humility about the fact that, ultimately, they cannot control all of the complex organisms and ecosystems in play.

We should think of science policy in much the same way. The state should absolutely fund science and build institutions around it, and yes, there should be space for top-down initiatives at DARPA, ARPA-H, NASA, and elsewhere.

But in most cases, the state should act less like that of a central planner dictating how to do science and more like that of a gardener cultivating a rich and diverse ecosystem. That means:

Supporting many kinds of organizations rather than just one dominant form.
Funding shared infrastructure, tools, datasets, and public goods that allow unexpected work to emerge.
Making room for originality, heterogeneity, and local initiative rather than demanding that everything be justified by the same bureaucratic norms and practices.
Paying attention to whether the ecosystem is healthy overall, not just whether a handful of centrally chosen programs can be given an impressive title and a press release.

Indeed, we need to be wary of the recurring temptation toward top-down “moonshot” initiatives that politicians seem to love. Some such efforts are worthwhile, especially where a problem is well-specified and where more coordinated engineering really is the bottleneck. But in science policy, the moonshot model is overused. Cancer and neurodegeneration are not solvable engineering problems like the Apollo missions or the Manhattan Project. They require more varied experimentation and more tolerance for the fact that progress may emerge from directions nobody predicted at the outset.

For too long, we have leaned too heavily on seeing science like a state. It is time to learn how to see it like a gardener.

What was already broken

The postwar research system did many things extraordinarily well. The United States experienced decades of scientific and technological progress, from the internet to self-driving cars to miracle drugs for cancer (e.g., Gleevec⁷ for a particular type of leukemia).

That said, for the past few decades, many observers have been pointing to deep structural and organizational problems in the way we fund and perform science.

First, grant writing and report writing are too time-consuming. Highly trained scientists spend countless hours⁸ writing proposals, tailoring their goals to what they think reviewers would want, revising applications that are rejected in 90% of cases, and sitting on panels to review overwrought applications from everyone else.

We should not want our scientific talent wasting their efforts navigating complex bureaucracies rather than advancing our knowledge, any more than we would want Steph Curry or Taylor Swift to spend half their time on compliance work rather than playing basketball or making music.

Second, the funding system is structurally conservative and risk-averse. This is because panel-based peer review was never designed to identify the most transformative ideas. It is reasonably good at filtering out bad or unserious proposals. It is much worse at selecting projects that have tremendous potential, but seem irrelevant, risky, premature, or hard to explain. By definition, no true scientific breakthrough would have been the consensus view of the entire field five years before it was made!

Third, there is the problem of organizational monoculture. To be clear, the federal government does not only fund university labs. It also funds things like the Department of Energy’s national labs, NASA centers and affiliated institutions like the Jet Propulsion Laboratory, large clinical-trial networks at places like MD Anderson Cancer Center, major scientific collaborations like the Whole Earth Telescope, and other structures that do not fit the simple PI-grant template.

But the modal grant in the American research ecosystem—especially in biomedicine and much of basic science—is the university lab led by a principal investigator (PI), staffed largely by graduate students and postdocs, and funded through project-specific grants.

That form is suitable for some kinds of work. But it is not the only kind of science that needs doing. Long-term data collection, scientific infrastructure, replication, systems engineering, benchmark creation, instrument design, and interdisciplinary projects are often unsuitable for the standard PI lab model. Again, it is not that these items never get funded, it’s that they should be a much higher priority than they are currently.

A fourth problem is how we treat the scientific workforce. The system relies heavily on trainees who do much of the day-to-day work while chasing a vanishingly narrow set of stable jobs. To put it bluntly, we typically fund several times the number of postdocs in biomedicine⁹ as there are academic jobs in that field. We are trapping thousands of people in low-paying jobs throughout their thirties while dangling the prospect of an academic position that will never arise.

This arrangement is exploitative and inefficient. Institutional knowledge walks out the door constantly, as projects become dependent on people who are not meant to stay. Many capable researchers eventually leave not because they lack talent but rather because they realize that the bargain on offer is quite literally a pyramid scheme.

None of that started with Trump. And it is worth saying something impolite but true: The way that the status quo makes use of many thousands of graduate students and postdocs primarily serves the interests of established scientists, not the public. A system can be highly dysfunctional in the aggregate while still being comfortable for the handful of prestigious people and institutions at the top.

AI could change the bottlenecks of science, and therefore the institutions we need

Much of the current discussion treats AI as an immensely promising tool to be inserted into the old scientific machine. That’s true, but it also misses a broader point.

The more important question—as yet unresolved—is how AI could change the location of scientific bottlenecks through lowering transaction costs (as per Ronald Coase) and streamlining previously tedious tasks.

In some domains, literature review, coding, exploratory analysis, hypothesis generation, and parts of experimental design are already becoming cheaper. If that continues, then other inputs will arguably become more important: access to high-quality data, well-maintained ontologies, automated experimentation, wet-lab execution, instrumentation, validation, and the like.

If generating plausible ideas becomes easier, then adjudicating among them also becomes more important. Simply put, if AI can help produce more hypotheses, then the value of reliable experimental systems, gold-standard datasets, replication pipelines, and organizations with strong judgment rises accordingly. After all, the real world is still far too complicated for any conceivable AI system to predict in full.

This could well matter as to how organizations are structured. How so? The institutional form that made sense when the bottleneck was a lone researcher with an idea and some graduate students may not be what makes sense when the bottleneck is a tightly integrated loop among compute, automation, data, validation, and deployment.

So the right response to AI is not merely to sprinkle such tools into existing labs and hope for the best. We should ask which organizations are best suited to scientific work when some activities become much cheaper and others become much more valuable.

The answer to that question likely points toward a more pluralistic ecosystem.

Different kinds of science need different organizational forms

Another oddity about our current system is that it treats different kinds of scientific work as if they were basically the same thing.

They are not.

Some science is discovery-oriented, and its key qualities are conceptual insight, theoretical imagination, and the ability to notice what others missed.

Some science is platform-oriented. The goal here is building tools, datasets, assays, atlases, instruments, model systems, software, and shared technical capabilities that make many future discoveries possible.

Some science is validation-oriented, with the goal being replication, benchmarking, quality control, and adjudicating among conflicting claims. In a healthier ecosystem, replication centers or multicenter validation networks would be treated as core infrastructure rather than as side projects.

Some science is translational. The main challenge is not how to prove something under sterile lab conditions but how to connect research to real users, institutions, hospitals, schools, and the like.

And some science is mission-oriented. The problem is not whether any one investigator has a clever idea but whether a sustained and coordinated effort can make progress on an ambitious, identifiable goal. The Apollo program, disease-focused clinical networks, and DARPA (and its imitators) all fit this mold better than the usual investigator-initiated grant.

In all of these cases, there is no reason to think that the organizational or bureaucratic form should be the same.

A reform agenda should explore a larger design space

For decades, the science-policy discussion has been rooted in an oddly stunted imagination about institutions. Folks can picture a university lab, or a national lab, or DARPA, or a corporate R&D group (with an obligatory mention of Bell Labs and Xerox PARC). And in the past few years, they may mention Focused Research Organizations (FROs). That is often about as far as the institutional imagination goes.

But the design space for research organizations is enormous.

These organizations can vary along many dimensions, including:

whether they are oriented toward discovery, validation, or platform-building;
whether projects are expected to produce results in two years or ten;
whether funding goes to people, projects, milestones, or missions;
whether the main personnel are trainees or permanent staff;
how much internal hierarchy there is;
whether research agendas are chosen by committees, program managers, users in the field, or some hybrid;
what their fundraising/revenue strategy is;
how long the organization has existed;
whether the organization is aimed at publications, technical benchmarks, some public mission, or actual end users.

FROs vary two of those factors (focus and time), and quite successfully: There is a set of narrow and definable scientific/engineering problems that demand an intense focus over five years. And the genius of FROs is that by defining these problems, and giving a name to a new institutional form, they became more fundable and legible.

But there are many types of problems that go unaddressed in our current system. Likewise, there are many more dimensions or factors that we could attend to within various scientific organizations in order to solve them. The possibilities here are wide, limited only by our imagination. Institutional isomorphism would be an issue even if we lived in a time of stability and status, but it is an especially bad idea in the age of AI.

The most vital structural reform is straightforward in principle, even if it’s politically difficult: The federal government should stop treating universities as the default home of publicly funded research.

To be sure, universities are among the major achievements of human civilization, and they remain excellent at many things: training, disciplinary depth, curiosity-driven inquiry, and the long cultivation of scholars.

At the same time, it is unlikely that the German research university model adopted in the U.S. in the late 1800s to early 1900s was the single best way for most scientific research to be organized and performed. While universities have been integral to the post-WWII ecosystem, they should be treated as one potential grantee among many, not as the presumptive destination for most research dollars. Institutions should change as society, the economy, and technology continue to evolve.

That means expanding support for other kinds of institutions (which can include new types of centers affiliated with universities, of course—universities themselves can evolve too).

The FRO model makes sense for initiatives like building shared tools, developing field-wide datasets, constructing infrastructure, and solving coordination problems that are too large for an ordinary lab but not yet attractive to industry.

Another approach that makes sense is the permanently staffed research institute outside the university system. Places like Janelia and the Allen Institute have shown what can happen when an institution is designed around long-term staff, internal collaboration, and organizational coherence rather than the endless turnover of trainees and the incentives of individual grant-seeking PIs.

Yet another is a hybrid public-interest lab that is a cross between a contractor, a nonprofit institute, and an R&D shop—an organization that can combine government work, philanthropic support, and independent technical exploration. An older example is BBN, which was instrumental in creating the original internet. We can imagine new versions in AI-enabled biology, scientific software, or lab automation.

And then there is an entire class of institutions that we will increasingly need: scientific commons institutions. Their job is to maintain public goods for the broader ecosystem, such as data resources, benchmarks, automation platforms, standards, interoperability layers, and negative-results repositories.

In other ways, we must broaden how the government supports research. Our funding architecture still mostly assumes that “science funding” means giving money to a project and letting the grantee assemble whatever else they need. But many research teams/labs require bundled capabilities such as compute, access to specialized models, robotic lab time, and engineering support.

Federal agencies should therefore do much more to experiment with larger funding packages that include access to and support for centralized resources. Imagine a team working on protein design. Instead of receiving money alone, it might earn access to a national scientific cloud and common benchmark environments, credits for shared wet-lab automation, and support from research software engineers. A materials-discovery team might be granted simulation resources, robotic synthesis access, standardized testing facilities, and embedded reproducibility support.

This would lower barriers to entry for new organizations, make it easier for small, high-quality teams to compete with large incumbents, and create reusable public infrastructure rather than requiring every lab to reinvent the wheel.

We also need selection mechanisms beyond peer review. Peer review became so dominant over the past sixty years that many people started to believe it was synonymous with legitimacy. It is not; it is one funding mechanism among many. Most of the highly significant scientific breakthroughs in history happened before peer review became common. This is no surprise: consensus committees are definitionally bad at funding ideas outside the existing consensus.

That is why the ecosystem demands parallel selection systems. The DARPA model is one alternative: empowered program managers with expertise, substantial budgets, and the authority to assemble portfolios without waiting for committee consensus.

But that should not be the only alternative. Lottery-based systems¹⁰ for proposals above a quality threshold can reduce pseudo-precision and the false legitimacy of tiny scoring distinctions. Long-term “people, not projects” funding can allow capable researchers¹¹ to change direction when reality changes. Challenge programs, milestone contracts, inducement prizes, and advance market commitments can work when the government has a concrete public goal in mind.

Validation and replication will become even more critical. If AI works even moderately well for science, then one consequence will be a dramatic increase in the volume of scientific publications and claims. That is not necessarily good. It could just mean more noise to sift through, which makes replication and quality control more important, not less. We should create more institutions whose primary job is to test, replicate, benchmark, and resolve uncertainty.

The workforce model also must be rebuilt from the ground up. Right now we take highly capable people, route them into decade-long training pipelines, rely on them for the daily labor of science, pay them very poorly relative to their skills and age, and then make most of them compete for a tiny number of jobs that may have little to do with the kind of work they are actually best at. Then we act surprised when many of them leave to work in fields like consulting and biotech.

A more rational system would provide serious and respected career pathways for research scientists, research engineers, automation specialists, software developers, data stewards, instrument builders, protocol designers, and other people whose main job is not to become a PI but to make the scientific enterprise work at a high level. CMU Robotics showed in the late 1980s¹² that it was possible to build a world-class research environment with durable roles for technical staff and builders, not just for faculty stars. More institutions should copy that lesson.

Again, none of this suggests that universities are ineffective institutions, let alone that we oppose them. It merely means that they are not the one-and-only solution for every type of scientific work. Public policy should say this explicitly: Some forms of research belong in universities, some in national labs, some in nonprofit institutes, some in start-ups or hybrids, some in public mission labs, and some in shared infrastructure organizations. The question should not be “How do we squeeze (nearly) every research topic into the typical university grant?” but “What team and organization is best suited to this task?”

Openness should be far more serious than it has been

Publicly funded science should be open by default, with data, code, protocols, workflows, and materials made readily available in usable, machine-readable form.

This matters for all the familiar reasons, such as reproducibility and broad access. But it matters even more in an AI era, because scientific progress will increasingly depend on these kinds of shared materials. Moreover, we should not be satisfied if researchers dump half-documented files into obscure repositories. The goal is to create a usable scientific commons: curated datasets, maintained ontologies, public benchmark suites, protocols, interoperable infrastructure, and institutions that keep these resources alive over time.

Nothing is free. Such openness requires organization, staffing, and money, which is why it belongs in the discussion of institutional design, not just in the discussion of norms. One of the biggest successes in AI-driven science (AlphaFold) happened only because of open databases¹³ of well-curated data like the Protein DataBank. We need to replicate that model many times over.

Our scientific success depends on democratic state capacity

Science policy typically crops up in national debate (if at all) as a matter of budgets, grants, and national initiatives like a “cancer moonshot.” But at a deeper level, it is about whether a democratic government can still create effective institutions.

The postwar generation did not inherit NIH, NSF, DARPA, and the national labs from the heavens. It built them by making decisions about how researchers would be funded, where they would work, and how public dollars would be funneled toward scientific capability.

We have the same responsibility now. We shouldn’t merely try to tinker with institutions and funding streams that are optimized for legibility to centralized bureaucracies. That is the temptation James Scott described: seeing a living system only in the simplified terms that make it easiest for the state to monitor and administer.

Science is not a plantation of identical trees growing in neat rows for the convenience of foresters. It is a living ecosystem: diverse, unpredictable, and more fruitful than any Soviet-style planner could have specified in advance.

The question, then, is not whether to reverse the chaos of 2025. The question is whether we can do more than that: i.e., admit that the old system had already become too narrow and rigid, tolerate genuine experimentation in organizational form, and build an ecosystem that is more pluralistic and aligned with the actual bottlenecks of science in an AI era.

The task is not to see science like a state. It is to see it like a gardener.

Natanson, Hannah, Emily Davies, and Dan Lamothe. 2025. “DOGE’s $1 Spending Card Limit Touches Everything from Military Research to Trash Pickup.” The Washington Post. March 9, 2025.
DiMaggio, Paul, and Walter W. Powell. 1983. “The Iron Cage Revisited: Institutional Isomorphism and Collective Rationality in Organizational Fields (Translated by G. Yudin).” Journal of Economic Sociology 11 (1): 34–56.
“Cancer Moonshot.” 2025. Cancer.gov. April 29, 2025.
ASPE. 2015. “National Plan to Address Alzheimer’s Disease.” ASPE. November 23, 2015.
Buck, Stuart. 2024. “The Paradox of Progress: Trying to Predict Impact Often Prevents the Highest Impact.” Substack.com. The Good Science Project. December 13, 2024.
Komoroske, Alex. “The Magic of Acorns,” Medium, April 4, 2023.
Pray, Leslie. 2014. “Gleevec: The Breakthrough in Cancer Treatment | Learn Science at Scitable.” Nature.com.
“Faculty Workload Survey - the Federal Demonstration Partnership.” 2025. The Federal Demonstration Partnership. September 4, 2025.
Buck, Stuart. 2024. “Texas Gas Stations, NIH-Sponsored Post-Docs, and the Foolishness of Artificially Subsidizing Demand.” Substack.com. The Good Science Project.
Chawla, Dalmeet Singh. 2020. “They Wanted Research Funding, So They Entered the Lottery.” New York Times, February 14, 2020.
Azoulay, Pierre, Joshua S. Graff Zivin, and Gustavo Manso. 2011. “Incentives and Creativity: Evidence from the Academic Life Sciences.” The RAND Journal of Economics 42 (3): 527–54.
Buck, Stuart. 2024. “A Scrappy Complement to FROs: Building More BBNs.” Substack.com. The Good Science Project. October 16, 2024.
Callaway, Ewen. 2024. “The Huge Protein Database That Spawned AlphaFold and Biology’s AI Revolution.” Nature 634 (8036): 1028–29.