Evidence at a glance
Behind the Breakthroughs Is a Question of Persistence
The material titled “AI Is Getting Good at Math. Here’s Why.” discusses recently reported mathematical work from OpenAI and Anthropic, and asks a harder question than whether models can solve math problems: which open problems are they more likely to advance? Its examples include an internal OpenAI model finding a counterexample to the unit-distance conjecture, open since 1946, followed by other reported results. These should be read as claims described in the material, not as independent confirmation of every result.
The question matters to technical leaders because research mathematics differs from contest problems by more than difficulty. Researchers must decide whether an approach deserves weeks or months of effort, while models can delegate more attempts to search and computation. The change may therefore be less about machines acquiring human-like intuition and more about lowering the cost and patience threshold for pursuing an idea.
A Hard Problem Is a Sequence of Hints, Not Just an Answer
The material draws on mathematician Timothy Gowers’s commentary on the unit-distance problem to offer a useful, though not directly computable, model of difficulty: solving a problem requires a sequence of hints, with work between them. Gowers’s sequence begins by looking for a counterexample, then generalizing the best known construction, and then considering number fields of increasing degree while restricting the norm of the prime ideals used. He represents the complexity as 2^k+t, where k is the number of hints and t is the work needed to develop them into a result.
The formula is valuable not as a score for hard problems, but as a way to describe the structure of a breakthrough. A hint can change how a problem is framed and reveal choices that were previously invisible; it is useful only in the right order. The material also says that mathematician Jacob Tsimerman explored a related direction but stopped when he lacked evidence of progress. That is not timidity so much as resource allocation: when early work cannot show whether a path is promising, the bar for continuing rises.
Verifiers Turn Compute into Feedback, but Not into Direction
One reason model search can work in mathematics is that some candidate results can be checked. Formal proof systems can verify whether steps follow their rules, while programs can validate answers to certain computational problems. A system can therefore generate multiple attempts, discard completed but invalid results, and direct compute toward other paths. This changes the search loop—generate, check, revise or retry—rather than relying on a single generation to be correct.
A verifier, however, helps only once an argument has reached a checkable stage. It can determine whether a completed proof holds, but it cannot tell a researcher whether an unfinished idea deserves more effort or supply the next conceptual leap. The parallel exploration and records of failed paths described in the material may help later attempts avoid known dead ends; they still do not provide a reliable early signal that a path is promising. As the number of branches and required insights grows, brute-force search is increasingly vulnerable to diminishing returns.
Results from Different Systems Do Not Measure One Ability
Available evidence suggests that verifiability is an important condition, but not a complete explanation of mathematical capability. The MATH benchmark, released in 2021, contains 12,500 contest problems; its paper argued that scaling models alone was not enough to solve them. AlphaProof represents another approach: it searches for proofs in formal systems such as Lean and uses reinforcement learning to improve its search policy. Its paper reports roughly 80,000 TPU-days of training compute and says that, on specific held-out benchmarks, additional inference-time search compute can further improve performance.
These figures describe system costs and results under particular test conditions, not a general guarantee for all models. The FrontierMath: Open Problems pilot included 14 open research problems selected so that candidate answers could be verified programmatically. Its benchmark page reports that none of the six models in its initial evaluation solved more than 2%. This shows both that open research problems remain difficult and that a benchmark built around verifiable problems cannot represent the whole of mathematical research. Contest performance, formal proof search, and original research results measure different stages.
Start Deployment Decisions with Feedback and Cost
For a technical team, the promising use case is not the vague instruction to “let a model do research,” but a task that can form a feedback loop. Can candidate results be checked reliably? Can failures become useful information for later searches? Is the compute cost lower than repeated human trial and error? Where no clear verifier exists, experts still carry much of the review burden. Where progress depends on repeated conceptual leaps, adding parallel attempts may not produce a proportional gain.
A more grounded evaluation is to choose a direction that researchers would hesitate to pursue for months, but for which candidate results and checking criteria can be defined. Let the model explore and record both successful and failed paths, then let experts decide what merits deeper work. The goal is not to replace research judgment, but to free expert time from mechanical trial and error. The limits must be counted too: search shifts costs to compute, verifiers cannot assess unfinished ideas, and benchmark scores cannot replace expert review of correctness, significance, or originality.