GPT-6 Astra Cracks a Goldbach Conjecture Variant: AI Moves from Computing to Thinking

A Letter Written 284 Years Ago That Stumped the Entire Mathematics World

In 1742, Goldbach wrote a letter to the mathematician Euler proposing what seemed like a simple conjecture: every even integer greater than 2 can be expressed as the sum of two prime numbers. This is what mathematicians famously call the "1+1" problem — one prime plus one prime.

To make it concrete: the even number 28 can be written as 3 + 25, but 25 equals 5 × 5, making it a composite number rather than a prime. A valid decomposition under the original conjecture would be something like 28 = 5 + 23, where both terms are genuine primes. The bar sounds low, yet it has stumped humanity for nearly 300 years.

The difficulty lies in the fact that primes are distributed with no discernible pattern — scattered along the number line as if at random, with no universal formula to apply. Chinese mathematician Chen Jingrun devoted half his life to proving "1+2" (one prime plus the product of two primes), the closest anyone has gotten to the original conjecture in the past century. But the product of two primes is fundamentally a composite number, and humanity has never managed to clear that final hurdle.

The "Stand-in" Version: The Liouville–Goldbach Conjecture

Faced with the towering wall of pure prime sums, mathematicians stepped back and constructed a lower-bar version — the Liouville–Goldbach conjecture. It relaxes the requirement of "two primes" to "two numbers whose combined count of prime factors is itself a prime number." Since a prime has exactly one prime factor, primes naturally satisfy a related condition.

The logical relationship here is worth clarifying: if the original conjecture holds, the stand-in version must also hold; but the reverse is not true — proving the stand-in version does not prove the original. It's essentially a way of lowering the threshold to probe the underlying rules connecting addition and multiplication.

Even with the relaxed requirements, human mathematicians spent several years making only limited progress: they proved the conjecture holds for "sufficiently large even numbers," but only under two additional assumptions — small even numbers were excluded, and the Generalized Riemann Hypothesis had to be assumed true. That amounts to dancing while wearing two layers of unproven assumptions — not a complete proof by any measure.

The count of prime factors is the key to understanding this stand-in version. Every integer greater than 1 can be uniquely decomposed into a product of primes (the Fundamental Theorem of Arithmetic); the total count of those primes, with multiplicity, is the number's prime factor count. For example, 12 = 2 × 2 × 3 has three prime factors (3 is itself prime); 8 = 2 × 2 × 2 also has three. A prime number like 7 = 7 has just one prime factor, and since 1 is not a prime, primes themselves do not satisfy the condition "prime factor count is a prime." The Liouville–Goldbach conjecture replaces "sum of two primes" with "sum of two numbers each having a prime number of prime factors" — a broader number-theoretic statement that is both related to and fundamentally different from the original conjecture.

GPT-6 Astra's Two-Step Breakthrough: From Conditional to Unconditional

According to an analysis by Bilibili creator Duyu, OpenAI's GPT-6 Astra shattered both of those constraints.

On day one, it produced a two-page PDF proving the conjecture holds unconditionally for all multiples of 4 — with no dependence on any unproven hypothesis. On day two, it found an entirely new line of reasoning that extended the result to all even numbers greater than 2, with no exceptions and no prerequisites.

What drew the most attention wasn't the result but the method. Rather than brute-force enumeration, it deployed proof by contradiction combined with properties of finite fields, building a step-by-step chain of logical deductions. The entire proof uses elementary algebraic reasoning that is reportedly accessible even to high school students. This ability to "find the right direction and construct a proof" has historically been considered a skill exclusive to elite mathematicians. More critically, the complete proof has passed formal verification — the logic is airtight.

It's worth staying clear-eyed: what was solved here is the Liouville–Goldbach conjecture variant, not the original Goldbach conjecture. The gap between "numbers whose prime factor count is prime" and "pure prime sums" remains considerable.

Finite fields are one of the core tools in this proof and deserve a brief explanation. A finite field (also called a Galois field) is an algebraic structure containing only a finite number of elements, closed under both addition and multiplication — the most familiar example is the set of integers modulo a prime p. Their advantage is that many problems intractable over infinite integer sets become more structured in finite settings, allowing algebraic rather than analytic methods. "Projecting" a number theory problem onto a finite field and then using the field's algebraic properties to derive a contradiction or construct an existence proof is a well-established, elegant technique in modern mathematics. Formal verification refers to translating every logical step of a mathematical proof into machine-checkable symbolic derivations using software such as Lean or Coq, fundamentally eliminating the logical gaps that human reviewers might miss — and it is a key pillar of this result's credibility.

From Swarm Intelligence to Single-Model Insight: A Watershed Moment for AI Research

Placing this breakthrough in context with earlier milestones makes the significance clearer. In the recent formally verified proof of the Navier–Stokes equations that went viral, OpenAI deployed ten thousand agents working for 88 hours to complete the task. That was an engineering victory — breaking a massive proof into countless modules, parallelizing the verification, and winning on sheer system orchestration and computational scale.

This time is an entirely different story: no cluster, no swarm tactics — just a single model following a single logical thread from start to finish, relying on insight into mathematical structure and the ingenuity to construct a proof.

If the Navier–Stokes proof was AI gnawing through a hard bone with brute numbers, this time AI walked a new path using a mathematician's mind. The leap from "computing" to "thinking" is a far more significant breakthrough. For the first time, it demonstrates that AI can produce mathematically elegant proofs that satisfy human aesthetic standards — not merely a calculator that piles on computation.

The Navier–Stokes equations are a system of partial differential equations describing the motion of viscous fluids, with applications ranging from weather forecasting to aircraft design. The existence and smoothness of their solutions is one of the seven Millennium Prize Problems posed by the Clay Mathematics Institute, carrying a one-million-dollar prize that remains unclaimed. What OpenAI completed previously was the formal verification of an existing proof of the equations — re-checking in machine language what human mathematicians had already written — rather than independently discovering a new proof from scratch. That distinction is fundamentally different from GPT-6 Astra constructing an original proof path on its own. Understanding this difference helps more accurately gauge the level of AI capability each breakthrough represents.

Closing Thoughts: AI Is Shifting from Tool to Research Partner

The significance of this event extends far beyond one conjecture variant. It effectively opens a new door for AI in scientific research — AI can now not only help scientists process data, but also participate in generating ideas and building derivations.

Our generation may be witnessing a historical inflection point: AI is gradually moving from a pure execution tool toward a genuine research partner, and in some domains, it is beginning to step ahead of us. For ordinary people, rather than feeling anxious, the better response is to understand the logic behind it — and keep pace with it.