Is AI Solving Math Problems Just Formalized Brute‑Force? An In‑Depth Analysis

On August 1, OpenAI released results from internal test versions of its next‑generation flagship model Astra, which solved ten long‑standing open problems in mathematics and theoretical computer science. The release included a 249‑page manuscript, Lean 4 formal certificates for each result, and complete reasoning narratives. The repository shows a sorry_count: 0 entry, meaning no placeholder "sorry" steps remain; the token cost of the ten solutions is roughly $2,000 at Sol API pricing.

Mathematics Does Contain Formalized Brute‑Force

The hypothesis has historical precedent. Tarski proved that the theory of real closed fields is decidable, and Presburger arithmetic (addition‑only integer arithmetic) is also decidable. The Zeilberger algorithm turns a large class of hypergeometric identities into finite mechanical checks, and Gröbner bases reduce polynomial ideal membership to computation.

Another line of work reduces infinite statements to finite enumeration. The 1976 computer‑assisted proof of the Four‑Color Theorem exhaustively checked 1,936 configurations. Hales’ proof of the Kepler conjecture used interval arithmetic and linear programming, later formalized by the Flyspeck project in 2014. In 2016, a SAT proof of the Boolean Pythagorean triples problem generated 200 TB of data—humanly unreadable but machine‑verifiable.

Thus, "formalized brute‑force exists in mathematics" is a fact.

However, the difficulty usually lies not in the computation itself but in the reduction that makes the computation feasible. Hales spent most of his effort devising a formulation that limited the verification to a tractable size; the computer only performed the already‑tamed part.

Why the Brute‑Force Hypothesis Often Fails

1. Many propositions have no enumerable domain. The "existence of non‑sofic groups" question, posed by Gromov in 1999, asks whether every countable group can be approximated by finite permutations. The objects form an infinite, non‑enumerable class, so enumeration is impossible. Similar issues arise for asymptotic sphere‑packing density bounds, where the statements involve limits as n → ∞.

2. Even when enumerable, the scale is astronomical. First‑order provability is semi‑decidable, so enumerating proofs by increasing length is theoretically an algorithm, sometimes called the "British Museum algorithm." Yet the number of candidate proofs grows exponentially with length. A 20‑page Lean proof contains 10⁴–10⁵ symbols, implying an enumeration space of roughly 10¹⁰⁰⁰, a number without physical meaning.

3. Empirical counter‑example. Last month, mathematician Levent Alpöge used Claude Fable 5 to find a counter‑example to the Jacobian conjecture, one of Smale’s 21st‑century problems that had been open for 87 years. The counter‑example is a three‑variable degree‑7 polynomial with 360 coefficients, for which no a priori bound on degree or variable count exists, making blind brute‑force search infeasible. Subsequent directed searches with sympy found similar examples, illustrating that knowing where to look (guided search) is the real challenge, not raw enumeration.

A More Accurate Statement

Rather than saying "formalized brute‑force," we can split the claim into two observations.

First, mathematics is currently the only discipline with a huge "generate‑hard, verify‑easy" gap. As Terence Tao noted in Nature , AI's biggest weakness in most fields is the inability to audit its mistakes, whereas a mathematical proof can be automatically checked.

This explains why the first uncontested AI successes appear in mathematics: not because math is simple, but because tools like Lean provide a reliable verification mechanism.

Second, models perform highly prior‑driven heuristic search rather than blind enumeration.

OpenAI’s reasoning narratives illustrate this. In the sphere‑packing problem, the model initially pursued a global norm estimate using Cauchy–Schwarz, then abandoned it because it lost local mass information, switching to a local quality‑exclusion inequality. For the non‑sofic‑group problem, it first tried a randomized grid argument, found it dead‑ended, and switched to a deterministic median‑based argument. These choices reflect a compression of decades of mathematical “taste” into a strong prior that reduces the branching factor to single digits.

Who Built the Pieces Astra Uses

If mathematics were merely brute‑force, the components Astra employed for the non‑sofic‑group proof would be interchangeable. Instead, it relied on the binary Leavitt algebra’s unit group, the Kun–Thom extension graph theory, and Thompson’s group V—abstract structures that required decades of human development. The model is not wandering a barren search space; it drives on a road paved by human theory.

Terence Tao, speaking at the 2026 International Congress of Mathematicians, emphasized that AI’s success in mathematics hinges on the centuries‑old foundational theories built by human mathematicians.

Consequently, the claim that "mathematics is brute‑force" is misleading: the cheap verification makes search possible, but the search itself is powered by deep, human‑crafted priors.

Finally, the bottleneck in mathematical research is shifting. We are moving from a scarcity of proofs to an overabundance of AI‑generated proofs, many of which are correct but lack expert validation. As Tao argues, a proof that cannot be presented in a clear, expert‑level report should not be published.

In summary, AI can crack difficult math problems not because the underlying methods are fundamentally brute‑force, but because mathematics uniquely offers inexpensive, automated verification, allowing a model equipped with strong mathematical priors to conduct focused heuristic search.