Did GPT-6 Astra Really Solve Goldbach's Conjecture? A Closer Look

A Math Story Blown Out of Proportion

The claim that "GPT-6 Astra has cracked Goldbach's Conjecture" has been spreading across social media with a distinctly clickbait flavor. The Caihua Bedtime AI Podcast took a calm look at what actually happened: Astra did produce a machine-verifiable mathematical proof — but it didn't tackle the classical Goldbach's Conjecture. Instead, it addressed a substantially weakened variant known as the Liouville version.

The classical Goldbach's Conjecture requires expressing any even number as the sum of two primes — a very strict condition. This version relaxes that requirement: both summands only need to have an odd number of prime factors (i.e., a Liouville function value of −1), meaning composite numbers qualify. That relaxation dramatically expands the range of eligible numbers and significantly reduces the difficulty of the problem.

In other words, calling this result "Goldbach's Conjecture" is a textbook case of bait-and-switch.

What Is the Liouville Version Actually Asking?

The idea of counting prime factors sounds abstract, but a concrete example makes it clear. Take 12: 12 = 2 × 2 × 3, which has three prime factors (counted with multiplicity) — an odd number, so its Liouville value is −1.

Now consider 20, which can be written as 8 + 12. Both numbers satisfy the Liouville value = −1 condition, but neither is prime. This example illustrates just how far the weakened version is from the original conjecture — it allows composites, while the classical version accepts only primes.

This weakened problem has been open since it was posed in 2018. The core difficulty is that additive structure and the parity of prime factor counts arise from two entirely different mathematical frameworks. According to the podcast, Terence Tao once noted in a Math Overflow discussion that problems of this kind still run into the inherent "parity problem" of sieve methods. Since then, researchers have proven the result for sufficiently large even numbers — but only under the assumption that the Generalized Riemann Hypothesis holds.

The parity problem is a fundamental limitation of sieve methods in analytic number theory, systematically identified by Atle Selberg in the 1940s. In short, sieves work by progressively eliminating multiples to count primes or near-primes — but they are inherently blind to the distinction between numbers with an odd versus an even number of prime factors. The two behave nearly symmetrically from the sieve's perspective, making estimates equally (in)effective for both. This is why Goldbach's Conjecture remains resistant to pure sieve approaches: even if you can show an even number can be written as a sum of two "almost primes," the sieve cannot tighten that constraint down to actual primes. The Liouville weakening remains hard precisely because the condition "odd number of prime factors" sits at the very heart of this wall.

Does "Code Compiles" Mean the Proof Is Uncontroversial?

The headline claim of this episode is that Astra produced a complete formal proof — written in Lean, and accepted by the machine compiler.

A successful compilation means the formal system accepted the theorem and its complete chain of derivation. The repository fixes the versions of Lean and Mathlib used, the build pipeline is clear, the axiom audit passes, and there are no unfilled "proof holes" (no sorry placeholders). From a pure correctness standpoint, this is a very strong signal.

But it doesn't mean everything is settled. A formal proof guarantees that the formalized statement holds under the given axioms — it does not automatically guarantee that the formalized statement is exactly the one the paper claims to prove. The correspondence between the two still requires human review. That step is unavoidable when evaluating any formal result.

Lean is an interactive theorem prover built on dependent type theory, developed by Leonardo de Moura's team at Microsoft Research, with Lean 4 being the current active version. Its core idea: mathematical propositions are encoded as types, proofs are encoded as terms of those types, and passing the type checker means the proof is logically complete. Mathlib is Lean's large-scale mathematical library, covering proven theorems from elementary number theory to algebraic topology — essentially the "standard parts library" of formalized mathematics. A sorry placeholder is a Lean keyword that lets you skip a proof step so the code compiles anyway. A proof containing sorry is formally incomplete; checking for leftover sorry instances is a basic step in auditing whether a proof is truly finished.

What Does the Kimi Incident Tell Us?

Around this paper, someone on Zhihu asked Kimi to find errors in it. What happened next was telling: Kimi initially concluded the paper contained an error, then reversed its verdict.

On review, Kimi had overlooked a key equation in the paper on the first pass, and only acknowledged its mistake after a second look. This small episode carries an important message: having another large language model read a paper is not a substitute for genuine verification.

A model's critical feedback can be useful for flagging suspicious spots — it has value as an assistive tool. But meaningful verification requires replayable Lean code and the math community's step-by-step review of both the statement and the proof. Models critiquing each other's work are prone to tripping on details.

Can This Count as Astra's Independent Mathematical Breakthrough?

The most critical question: is this an independent mathematical breakthrough by Astra? We're not yet in a position to say so definitively.

The paper's source code and verification records are available on GitHub, but the paper includes no author statement page, no account of the prompting process, and no disclosure of how much was revised by hand. OpenAI's Astra release page likewise doesn't call out this result.

The result is an awkward gap: whether the theorem holds can be checked via code, but the layer of "who discovered it and who revised it" lacks any verifiable record. Without a clear accounting of the human-AI division of labor, attributing the achievement entirely to the AI is premature.

Prompt transparency in AI-assisted research is an emerging issue that some journals have already begun to address. Unlike traditional software engineering, the output of large language models is highly sensitive to the wording, decomposition strategy, and number of iterations in the input. The same problem with different prompts can yield radically different results. If a researcher provided highly directive prompts at key reasoning steps, the line between "AI produced the proof" and "the human provided the proof strategy and had AI fill in the details" becomes blurry. Proposed responses from the research community include: publishing complete conversation logs, distinguishing between "AI-generated" and "AI-assisted" authorship attributions, and explicitly stating the human-AI division of labor in the methods section. The absence of such records doesn't imply dishonesty — but it does make independent evaluation impossible.

Mathematical Value and AI Capability: Don't Conflate the Two

What's really worth paying attention to here — the mathematical result, or AI's capabilities? The podcast's view: both matter, but they shouldn't be conflated.

From a mathematical standpoint, this provides a short proof with complete formal delivery for a weakened problem that has been open for years — that's meaningful in itself. From an AI standpoint, it demonstrates that models may be able to participate in "finding approaches, writing papers, and making formal engineering judgments" — that's an expansion of capability boundaries.

For the next wave of "AI conquers mathematical milestone" headlines, the podcast offers three practical criteria: How strong is the theorem? Can the verification be replayed? What did the human and the model each contribute? Ask those three questions, and it's much harder to get swept away by the headlines.