I've been running servers since before Google existed. I've watched compute shift from bare metal to virtualization to containers to serverless to "just upload your model weights and call it done." But last weekend, something happened that made me stop and genuinely think: we may be at the beginning of something qualitatively different about what software is capable of doing.
On July 19th, while the World Cup final was playing in the background, a number theorist named Levent Alpöge — on staff at Anthropic, formerly a Harvard Junior Fellow — sat down with Claude Fable and, in the span of a few hours, helped disprove a mathematical conjecture that had been open for 87 years.
The Jacobian Conjecture. Posed in 1939. On Steve Smale's famous list of the most important open problems for the 21st century. Worked on by generations of algebraic geometers. Now, dead — dispatched by a degree-7 polynomial while someone had a soccer game on. Let me tell you why that matters even if you've never heard of the Jacobian Conjecture and never plan to.
What the Conjecture Actually Said
The Jacobian Conjecture is, at its core, a question about functions in high-dimensional space. If you have a polynomial map — a function from n-dimensional space to itself built out of polynomials — and its Jacobian determinant (a measure of how the function stretches space locally) is a nonzero constant everywhere, does that force the function to be globally invertible? Calculus tells you a constant nonzero Jacobian is sufficient for local invertibility. The conjecture said local ought to be enough for global.
Simple to state. Brutally hard. Mathematicians had been searching for either a proof or a counterexample since 1939. Prior computational searches had assumed any counterexample would be a monstrous polynomial of degree around 200 — that was the rough theoretical lower bound that shaped where people looked. Alpöge and Claude Fable found one at degree 7.
The counterexample is a three-dimensional polynomial function with a Jacobian determinant of constantly negative two. Three distinct input points — (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) — all map to the same output: (-1/4, 0, 0). The function looks invertible locally everywhere, but isn't globally. The conjecture is false.
As Jared Duker Lichtman at Stanford put it, this is "quite a remarkable result" — and that's coming from someone who works adjacent to this area. Fields Medalist Timothy Gowers called it "pretty amazing," noting it was the first time he'd personally witnessed an LLM help solve a major open problem outside his specialty that he'd definitely heard of. That's not a hype merchant talking. That's one of the five or ten best living mathematicians saying something surprised him.
The Degree-7 Shock
Here's the detail that hit me hardest: everyone searching for a counterexample assumed it would be gigantic. Degree 200 was the rough expectation based on existing theory. Degree 7 is almost insultingly small. If a human mathematician had found it in 1985, colleagues would have asked why nobody thought to look there earlier.
This is where I think the AI angle is genuinely illuminating rather than marketing. A human researcher working in this space carries a lot of assumptions — priors built up over years of failed attempts and theoretical bounds. Those priors pointed toward huge polynomials. They shaped where people searched, how they structured their brute-force attempts, what degrees they considered tractable. Alpöge working with Claude Fable apparently didn't get trapped by the same assumptions in the same way.
Bartosz Naskręcki at Adam Mickiewicz University, who works directly in this space, put it plainly: "Finding counterexamples like this to the Jacobian Conjecture is like searching for a needle in a haystack." The degree-7 result suggests the needle was in the first haystack all along, and everyone had been poking around in the fifth one because theory said the needle had to be big.
That pattern — where wrong priors constrain search in ways that are invisible until someone looks from outside — is one I recognize from infrastructure work. The cache bug that everyone assumed was in the application layer because it had always been in the application layer. The network latency that nobody looked at because the box had always run fine. Sometimes the constraint on finding the answer is not the search space, it's the assumption about what the answer looks like.
What the Verification Looked Like
One thing I find operationally interesting — and telling — is how the AI models reacted when mathematicians tried to verify the result. Researchers in the Hacker News thread, which drew 761 points and substantial participation from working mathematicians, noted that when they ran the numbers through various LLMs to double-check, the models exhibited what one described as "psychosis": they kept verifying the result multiple ways, apparently reluctant to accept it, before finally conceding it was correct. There's something almost relatable about that. I've had the same reaction to a monitoring alert that contradicted eight years of operational intuition — I check the dashboard twice, then grep the logs, then finally accept that something genuinely changed.
The result hasn't yet gone through formal peer review or appeared on arXiv as of this writing. The math community is appropriately cautious — false proofs and disproof claims for famous conjectures are not unheard of, and there are natural questions about reproducibility and the degree to which the AI contributed versus Alpöge's own insight. Alpöge works at Anthropic, which some observers have flagged as a potential conflict of interest in attribution. The full session logs haven't been publicly released.
That said, the counterexample itself is elementary to verify independently. You don't need to trust the AI or the researcher's narrative about how it was found — you just compute whether those three input points map to the same output and check that the Jacobian determinant is -2. It either holds or it doesn't. The math community seems broadly convinced that it does.
The 2026 Pattern Worth Tracking
This isn't an isolated event. 2026 has already seen a string of AI-assisted mathematical breakthroughs: AlphaProof Nexus resolved several open combinatorics problems earlier this year; Claude Mythos contributed to a significant number theory result; AxiProver has been working through formal proof obligations at a pace that would have been inconceivable three years ago.
The Jacobian Conjecture result is different in kind from those, though. This isn't a benchmark or a curated competition. It's a result that drew Fields Medalists out of their inboxes to comment publicly, that researchers who've spent careers on related problems are treating seriously. Daniel Litt at the University of Toronto was measured but optimistic: he described himself as "very bullish" on AI's near-term impact on mathematics, even while noting the Jacobian Conjecture itself "may not be particularly generative scientifically" on its own. Przemek Chojecki, who holds a PhD in mathematics, predicted the result would inspire "extensive new research across related mathematical domains." And Vidit Nanda at Oxford offered what I think is the right frame: "Let us celebrate the present and beware the future."
Not hysteria. Not dismissal. Calibrated attention to something that is actually happening.
What This Means if You Ship Software
Most of us are not algebraic geometers. But there's a direct read-across to the work we do.
The dominant framing of AI coding assistance has been autocomplete-on-steroids: it fills in the boilerplate, suggests the obvious next line, completes the function signature. Useful. Not very interesting. What the Jacobian result represents is something different: genuine exploration in a search space too large for human brute force, guided by a domain expert who knew how to frame the question and recognize a surprising answer.
In software terms, that's the equivalent of an AI helping a security researcher find a zero-day not by running a known exploit, but by exploring the state space with someone who understands what "unexpected" looks like in context. Or finding the performance regression in a distributed system by exploring parameter combinations no profiler would suggest. Or discovering that a cache invalidation policy you've run for years has a degenerate case that only manifests at production scale — and that the fix, once you see it, is embarrassingly simple.
The degree-7 shock is precisely the point. We keep expecting hard things to be hard in the ways we're used to. We build our search strategies, our tooling, our code review processes around where we've found problems before. Sometimes the answer is simpler than our accumulated priors let us look for — and a collaborator that doesn't carry those priors the same way we do can find it.
What to Watch Next
Once Alpöge's result is formally published and peer-reviewed, it will need to be absorbed by the algebraic geometry community. The mathematical downstream is somewhat open — as Litt noted, the conjecture's disproof may not immediately unlock major new theory, but the methodology is the real story. If this approach works for an 87-year-old open problem in pure mathematics, it works for hard engineering problems that are well-specified but too large to search exhaustively.
On the infrastructure side of my world, I'm watching what this means for formal verification tooling. Lean-based proof assistants can potentially create a flywheel: each AI-assisted result expands the library of verified mathematics that future systems can build on. That has implications for security auditing, protocol verification, and correctness proofs for distributed systems that most people in my industry haven't fully thought through yet. When a model can both generate a proof candidate and verify it in a formally checked system, the feedback loop gets very tight.
Eighty-seven years is a long time for a question to stay open. The fact that it closed on a weekend afternoon, with a World Cup final on in the background, while someone was running sessions with a language model — that's a dispatch from a future that's already here. The question now is what other search spaces we've been avoiding because theory told us the answers were supposed to be found elsewhere.
We should be paying attention.