Science Won't Wait for Its Institutions

AI in Science is moving fast. Just a few months ago, AI systems solved seven of ten novel research-level math problems at publication-level quality, for at most $1k of compute apiece.¹ The casual discussions have subtly moved on from how good the models have become and what they are capable of to an entirely different space. At one of the salon-style dinners we’ve been hosting with researchers and leaders to get a pulse on AI + Science, the core takeaway was that science is heading into a disruptive transition where the hardest problems we have to solve for are now more social and institutional than technical.

If you’re a Hacker News reader, you might have seen a steady stream of posts about new breakthroughs in mathematics. I’m sure every field is feeling this at a different speed, and mathematics feels furthest along. Terence Tao says AI assistance has become routine at the Erdős problems website,² and the teams building neural provers, whose machine-generated Lean proofs now run to thousands of lines, call simplification for human readers “a critical bottleneck”.³ Math may be closest to an “era of proof abundance,” with models generating formal proofs faster than humans can interpret them.

But a proof that no human can follow is an odd thing to celebrate. What exactly did mathematics gain as a field? Tao addressed this to his ICM 2026 audience, as the field enters “a crisis in the foundations of mathematical values and practices”.¹ So the job may shift: not producing proofs so much as finding conjectures worth pursuing, plus the unglamorous work of turning machine-generated proofs into explanations people can learn from. One recent position paper (Tao is an author) calls this the move from problem-solvers to research agents.⁴

Biology couldn’t be more different. You can’t flood a search space that was never well-defined in the first place, and biology is full of unknowns and under-observed mechanisms. Paola Lecca’s paper on machine learning for causal inference makes the point that a correlation, however strong, doesn’t hand you a mechanism or a target for intervention.⁵ Víctor de Lorenzo, writing in PLoS Biology, goes further: machine learning now enables “direct leaps to application without understanding the principles,” so will mechanistic studies decline?⁶ Our guess was that AI helps rather than hurts here. An emerging “scientific machine learning” literature pairs ML with mechanistic models for exactly this reason.⁷ None of that removes the need for experiments and new measurements, or for some better way to connect datasets that remain fragmented.

Another theme that kept coming up was the culture of data sharing in astronomy. That made sense; we had multiple astronomers in the room. For as long as I’ve known (since the early days of BIDS), the field has treated shared data as infrastructure.⁸⁹ Almost no other field I know has built anything like it, and part of that makes sense: telescopes are so expensive that not sharing the output was never really an option. Mostly, though, it comes down to incentives, or lack thereof. Journals mandate data archiving on paper, but it’s never really enforced.¹⁰ There are real IP concerns in some domains, and labs compete for limited resources.

That’s the real challenge, because the technical pieces for progress already exist in a lot of fields. What’s missing is usable data that’s openly available, not just serving a handful of labs.

We also briefly touched upon preprints, starting with arXiv. Most preprint repositories have always been designed for human speed, in both writing and reviewing. With the rapid advances in frontier models, nobody anticipated that we would soon be looking at a flood of AI-generated output. Journals were already struggling to find reviewers before any of this technology came along, and early research is starting to show that LLM-generated text is measurable at scale across arXiv, bioRxiv, and medRxiv.¹¹ One detection study classified roughly 20% of 2025 ICLR peer reviews as AI-generated.¹² As a quick response, arXiv has stopped accepting unpublished review and position papers in its computer science category, citing just how many of the submissions were LLM-generated.¹³¹⁴ While it is hard to detect low-quality and fraudulent material at the moment (this will change over time), the harder challenge is evaluating which of these contributions are valuable.

Despite this shift in tone, the room was optimistic about the upsides. AI accelerates simulations, gives more researchers access to sophisticated tools, and draws connections across fields that rarely collaborate. It could also allow researchers to revisit problems previously considered intractable. But the technology is moving much faster than universities, funders, journals, and regulators. The concern was not whether those institutions would adapt, but whether they would do so before the most important decisions had already been made.

References & Further Reading