Our Take
When AI solves problems faster than humans can publish them, the research economy breaks; this is not hype, it is a structural problem for fields where novelty is currency.
Why it matters
Academic priority and publication timelines depend on a lag between problem statement and solution. When AI collapses that lag, researchers face scooped work, wasted effort, and incentive collapse. This matters now because math, physics, and theoretical CS face this immediately.
Do this week
Research leaders: audit your lab's 2-3 year projects against recent arXiv preprints and model capabilities benchmarks this week so you can deprioritize work already solved and redirect effort to open problems AI has not yet addressed.
Years of Work Rendered Obsolete by AI
Researchers working on long-standing open problems in mathematics discovered that large language models and AI systems had already solved or substantially advanced their work. The New York Times reports that mathematicians spent years—sometimes multiple years—on specific problems, only to find that AI models released during their research window had solved or superseded those same problems.
The incident is not isolated. It reflects a structural shift in how research gets done when AI can operate on problem sets at speeds that outpace human publication cycles. A mathematician who spent considerable effort on a problem now faces not just competition but retroactive obsolescence: the work was valuable when begun, but AI moved faster.
The Research Economy Assumes a Human Pace
Academic priority and career advancement rest on novelty. You publish first, you own the result. But that system assumes human-scale timelines: identify a problem, think for months or years, write it up, submit, get reviewed, publish. Peer review takes months. The entire pipeline assumes a year or more between problem statement and public solution.
AI compresses that to weeks or days. A model released in January can solve problems that were thought open in December. The researcher who started in January, planning a 2-3 year push, discovers in month six that their entire premise is solved and published by someone else (or a model).
This is not a marginal edge case. As AI capability expands into higher mathematics, theoretical physics, and formal verification, any researcher working on problems public for more than a few months faces real risk of AI scooping. The incentive structure cracks: why spend two years on something a model might solve in two months?
Fields that operate on long timelines—pure mathematics, theoretical physics, fundamental computer science—now face a choice: pivot to problems too new or too niche for AI to touch, or accept that priority and novelty claims are now contingent on AI release schedules, not research quality.
Reframe What Counts as Research Output
Individual researchers cannot outrun model release cycles. But research groups can change what they optimize for. Stop treating "first to solve" as the only win. Reframe around: interpretation, generalization, application to unsolved subproblems, or methodology that AI alone does not provide.
A mathematician whose proof was scooped by AI still has value in understanding why the proof works, where it extends, and what harder problems it enables. That is not the same output, but it is defensible work. Alternatively, shift early to problems where public formulation is newer or where the edge of AI capability has not yet reached.
For labs: audit your multi-year roadmap against current model capabilities and recent arXiv releases. If a project could plausibly be solved by a model release in the next 18 months, either deprioritize it, reframe it around interpretation rather than discovery, or accelerate it to publication before that window closes.