AI is solving problems that stymied human researchers for years, and the pattern is too consistent to dismiss
From quantum mechanics simulations to century-old mathematical conjectures, AI systems are cracking problems that expert researchers could not. The results span enough domains that the accumulation demands a harder look than any single benchmark warrants.
The clearest signal that something structural has shifted in scientific research is not any single benchmark score. It is the accumulation of specific, domain-spanning cases where AI systems have solved problems that human researchers could not, often in a fraction of the time those researchers spent trying.
Alex Lupsasca supplies some of the most striking examples. A simulation of the SYK model, a technically demanding problem in quantum mechanics and gravity that multiple research groups had failed to complete over more than a year, was written by Codex in ten minutes. In a separate case involving gluon amplitudes in particle physics, researchers had spent a year unable to simplify the formula for single-minus gluon tree amplitudes. Lupsasca reports that “the final formula was first conjectured by GPT 5.2 pro and then proved by an internal OpenAI model.” The model also extended this work independently: anchored on the gluon paper, it completed the graviton amplitude calculation, a mathematically distinct problem, without further human input. Lupsasca’s conclusion from these episodes is direct: current models can produce papers indistinguishable in quality from those written by humans.
The mathematics competitions tell a complementary story. Carina Hong reports that Axiom Math scored 120 on the 2025 Putnam exam, above DeepSeek’s score of 103 and above the best human score of 110, on what Hong describes as a 120-point exam. That result is not just a benchmark achievement. It reflects a system architecture in which inference compute scales without hitting a wall, achieved by recursively decomposing proof goals into sub-goals and learning to backtrack. The same team has previously disproved a 30-year-old conjecture by finding a counterexample and solved a 130-year-old problem involving the global Leono function.
The final formula was first conjectured by GPT 5.2 pro and then proved by an internal OpenAI model. Alex Lupsasca
Formal verification is surfacing errors that decades of human review missed. Carina Hong also describes an episode in which auto-formalization using Lean caught an implicit assumption in Robert Aumann’s 1976 “agree to disagree” theorem that had never been made explicit in fifty years of teaching and citation. The prover not only identified the gap but patched it. That is a different category of contribution from solving a new problem. It is AI functioning as a more rigorous reader of existing human work.
The cost compression that accompanies these results reshapes who can attempt serious research. David Sinclair describes his lab completing work that would have required 160 years and, by his estimate, billions of dollars, now on a ten-thousand-dollar budget. Eric Jang puts a similar frame on software research: what required a full team of research scientists at DeepMind and millions of dollars of compute can now be replicated for a few thousand dollars of rented compute. Dylan Patel adds that a single person using Claude Code finished a research project he estimates would have taken a team of 200 economists a year. Terence Tao, describing work on Erdős problem 1026, notes that researchers used the latest AI tools to gather numerical evidence in the course of solving it, a quieter data point but meaningful given the source.
The internal organization of research is shifting in response. Mark Chen observes that at OpenAI and at other labs, work is becoming mostly orchestration-focused, with researchers generating ideas and models handling implementation and execution. That is not a description of AI as a productivity tool. It is a description of AI as a research partner that handles the parts of the work that used to require teams. Noam Brown offers a personal measure of the same shift: he expects that within six months to a year, a model will be able to produce a complete poker solver zero-shot, the equivalent of his entire PhD thesis in one pass.
Brian Greene, discussing string theory specifically, sees a real possibility that the nature of research changes in the next five to ten years, with AI solving problems that humans have been unable to solve. His assessment tracks with what the evidence from other domains already shows. The problems that AI is now cracking are not simple ones. They are problems where expert human researchers, working with full institutional resources and years of time, made no progress. That is the relevant baseline. The compression of time and cost is striking, but the more consequential fact is that the ceiling on what can be solved at all appears to be moving. Whether research institutions built around human-paced discovery reorganize quickly enough to direct that capacity productively is the question the current evidence raises but cannot yet answer.