Science Infrastructure in the Age of AI
Hypotheses generation is getting cheap, but verification isn't
Loosely based on my remarks at Johns Hopkins Carey Business School “Leading with AI” event
Transformative AI is arriving and disrupting every part of our society and economy — the same is and will be true for Science.
This new technological era raises many important implications for the science policymaker: how do we drive adoption of coding agent tools among scientists? how does AI change the role of autonomous labs in scientific experimentation and the geopolitical scientific landscape? What are the needed changes for scientific publishing and curation to remain relevant?
Today, I want to sketch out the case for another implication that AI has for how we fund science.
If we view LLMs as commoditizing the task of hypothesis generation, then the binding constraint of scientific progress shifts to hypothesis verification. This naturally implies an expanding role of autonomous labs, but it also implies that science infrastructure will become more important in the age of AI.
What do I mean by science infrastructure? It can range from high-end microscopes costing tens of millions of dollars, such as cryoEM and Scanning Transmission Electron Microscopes (STEM), to billion-dollar synchrotrons, supercomputing clusters, and ion beam user facilities.
Funding large-scale scientific infrastructure has historically fallen to public agencies and philanthropists, for a simple reason: the high upfront fixed cost, low marginal cost of operation, and low utilization rate by any one institution make it uneconomical for a single university or company to build, even though many would benefit from access.1
Why do these become more valuable? As the cost of generating hypotheses falls toward zero, the binding constraint becomes the rival, excludable capacity to test them. Especially as more private AI-powered companies focus on hypothesis generation, the relative value of public funding shifts from underwriting ideas to underwriting instruments and infrastructure.
A further benefit is that investments in infrastructure are hypothesis-agnostic: the taxpayer takes on a lower risk, since the bet is not on any single hypothesis but on the capacity to test a broad range of them.2
The value of such infrastructure is clearest in the astronomy and particle physics communities. These are fields which already have a surplus of hypotheses and are constrained by infrastructure capacity to test them. The Decadal Survey for astronomers and Snowmass & P5 for physicists convene the entire academic community to align on the next big shots on goal to take in terms of which new telescopes and accelerators will help test the most promising hypotheses.3
My argument is that the value of scientific infrastructure, already legible for some fields, will become even more valuable for other experimental sciences as AI becomes more powerful.
And it is not just a matter of capacity — new types of infrastructure and instrumentation often advance science in non-linear fashion, opening up previously inaccessible modalities of perception. In other words, the value of new scientific instruments also increases in an age of AI alongside large-scale scientific infrastructure.
Unfortunately, the specific value provided by instruments and infrastructure to scientific progress is relatively understudied outside of history of science and STS fields. The lack of broader awareness of the importance of scientific tooling is something even Freeman Dyson recognized:
Scientific revolutions are more often driven by new tools than by new concepts. Thomas Kuhn in his famous book, “The Structure of Scientific Revolutions”, talked almost exclusively about concepts and hardly at all about tools. His idea of a scientific revolution is based on a single example, the revolution in theoretical physics that occurred in the 1920s with the advent of quantum mechanics. This was a prime example of a concept-driven revolution. Kuhn’s book was so brilliantly written that it became an instant classic. It misled a whole generation of students and historians of science into believing that all scientific revolutions are concept-driven. The concept-driven revolutions are the ones that attract the most attention and have the greatest impact on public awareness of science, but in fact they are comparatively rare.
In particular, the overuse of papers and patents in the “metascience” community has led to a blindspot in properly understanding the hidden value of science infrastructure. While there is much focus on reforming PI-based grant review processes — such as golden tickets for reviewers — there has been far less discourse around the value and processes of programs like NSF’s Mid-Scale Research Infrastructure grant program and DOE’s user facilities. As AI progress accelerates, science policymakers should carefully consider the balance of public funding for PI-based grants and infrastructure capacity.45
By contrast, the logic of industrial science and infrastructure is one that China understands well. Some of the best technoeconomic analysis of U.S. national lab infrastructure is done by Chinese Academy of Sciences researchers. Today, China is building more than 90 megascience facilities with spending on scientific capital assets more than tripling between 2015 and 2024 even as traditional infrastructure investment has slowed. The U.S. may retain its lead in AI model capability and still fail to translate that advantage into accelerated scientific progress if we remain bottlenecked by hypothesis verification.
Finally, as someone who works in philanthropy, the value of science infrastructure is even more apparent. Rather than betting on a specific scientific idea or thesis, a key science infrastructure investment can bolster an entire field without requiring as opinionated or clear a view on a specific hypothesis or discovery. Warren Weaver demonstrated the value of promoting new science instruments in accelerating the pace of new biological discoveries and fields. In the modern era of science philanthropy, Eric Schmidt’s donation for a new collider at CERN and a new telescope and three observatories for astronomers, as well as Chan Zuckerberg Initiative’s support for breakthrough phase imaging in cryoEM, are more recent examples.6
Regardless of funding source, it remains the case that AI is steadily dissolving the assumption that better science comes from improving the supply of ideas through better grant mechanisms, more PhDs, more papers. As hypothesis generation becomes abundant, the binding constraint on scientific progress shifts toward our capacity to verify hypotheses. The scientific infrastructure capacity for large-scale hypothesis verification is precisely the club good that markets will not build, and that public agencies, national labs, and philanthropies exist to provide.
As we enter an age of AI and the “third wave of American philanthropy”, philanthropists and policymakers should more closely consider the value of scientific infrastructure.
i.e. a classic club good
For instance, Boudou and McKeon find that on the margin, additional public compute capacity allows researchers to “study less popular and newer topics, explore new topics that they have not studied in their prior work, and broaden the scope of their projects”.
If you are an astronomer or physicist who has experience in this kind of field planning and are interested in writing about your experience, please do reach out! These practices are often opaque and understudied by the broader science policy ecosystem and I’d love to support more writing on the interplay between scientists and funders in infrastructure planning.
We used to appreciate the value of scientific infrastructure before: from the Human Genome Project, a publicly funded effort to dramatically improve the accessibility of a key instrument that served to verify and validate a generation of biological hypotheses, to the Protein Data Bank, a data repository hosted by Brookhaven National Lab which also generated a significant amount of protein images from its beamline. Both of these efforts were foundational to enabling the advances in AI x Bio.
Indeed, the “Big Science” paradigm in the 1970s was so predominant that prominent scientists believed we had begun to overfocus on such infrastructure. But I would distinguish what I am advocating here as slightly distinct from the “Big Science” of the late 20th century, which was disproportionately focused on astronomy, particle physics, and space exploration i.e. basic science. My conception of science infrastructure is more neutral and indeed, far more applied in nature with beamlines and microscopes being key infrastructure for material and biological sample characterization.
See: Renaissance Philanthropy playbook on “mid-scale science” as a related theme.