AI Just Cracked a 30-Year-Old Math Problem. Or Did It?

“AI solves a 30-year-old math problem.” Read that sentence and your pulse quickens a little. The story making the rounds this week is that GPT-5.6 completed a long-open proof in convex optimization. But headlines like this deserve a hard look before you share them. What does “solved” actually mean here? Where does the fact end and the marketing begin? Those are the questions worth asking.

Let me be upfront about one thing. There isn’t a rich trail of expert community debate on this specific claim yet. So instead of leaning on upvotes or comment threads, this piece is about something more durable: how to read a claim like this in the first place.

Why Convex Optimization Is a Big Deal

Let’s set the stage. Convex optimization sounds intimidating, but the core idea is intuitive. Picture a bowl-shaped function that curves downward. The problem is finding the lowest point at the bottom. The magic is that the bottom is unique. In rugged terrain full of valleys, you can never be sure which dip is the deepest. In convex terrain, the local minimum is always the global minimum. Find any bottom, and you’ve found the bottom.

Why does that matter? Training machine learning models, building financial portfolios, routing logistics networks, processing communication signals — an enormous share of real-world problems boil down to “find the best value.” Convex optimization is the mathematical bedrock that guarantees you can actually find it. So if a long-standing open question in the field genuinely closed, that’s not trivial in theory.

The Trap Hiding in “30-Year-Old Problem”

Here’s the first note of skepticism. The phrase “30-year-old problem” is seductive — and slippery.

Math difficulty spans a huge spectrum. On one end sit monsters like the Riemann Hypothesis, where hundreds of mathematicians have thrown themselves at the wall and bounced off. On the other end sits a problem that stayed “open” for 30 years simply because nobody bothered to compute the optimal constant in some particular inequality. The second kind isn’t hard so much as untouched.

When an “AI solved it” story lands, the question to ask is precise. Was this a problem that many experts attacked and failed to crack? Or one that nobody had prioritized? Blur that line, and an ordinary result starts to look like the breakthrough of the century.

What Does It Even Mean for AI to “Prove” Something?

The second thing worth interrogating is the word “proved.” There are at least three very different scenarios hiding behind it.

One: the AI conceived an entirely new proof strategy on its own and carried it to the finish. That would be a genuine event. Two: the AI cleverly combined known techniques to fill in a gap. Impressive, but a different flavor of creativity. Three: a human researcher set the direction and the AI helped with computation and case-checking. A useful tool — but the subject of the sentence “solved it” is a person.

Most headlines don’t distinguish among these. They flatten all three into “AI solved it.” And there’s one more verification checkpoint that matters more than any of them: whether the proof was machine-checked by a formal verifier like Lean or Coq. A proof read only by human eyes can hide subtle errors. A proof that clears formal verification has its logical consistency guaranteed mathematically. Any “proof” that doesn’t mention this step deserves a question mark until proven otherwise.

The Trend Itself Is Real

If you’ve read this far, you might think I’m dismissing the whole thing. I’m not. Individual claims need scrutiny, but the broader trend is unmistakably real.

Over the past few years, real contributions from AI to mathematics have piled up. Solving International Math Olympiad-level problems. Finding new lower bounds in combinatorics. The rapid maturation of automated formal proving. Mathematicians have genuinely started pulling AI into live research as a proof assistant. So the idea that GPT-5.6 made a meaningful contribution to one optimization problem is entirely plausible.

The key is balance. “AI has replaced human mathematicians” is hype. “It’s all marketing” is cynicism. The truth sits somewhere in between: for certain classes of problems, AI is becoming an increasingly useful collaborator.

The Signal Worth Watching

More important than solving any single problem is the direction underneath it. For AI to earn trust in mathematics, two things ultimately decide the outcome: reproducibility and verifiability. Is the proof public? Has an independent researcher confirmed it? Did it pass a formal verifier? Only when all three line up does “solved” carry any weight.

Flip it around. If there’s no paper, no verification, no independent confirmation — just a press release and a demo circulating — then you aren’t consuming an achievement. You’re consuming anticipation. Telling the two apart has become an essential skill in this era.

The age of AI cracking hard math problems is genuinely arriving. But how far that door has actually opened will be told by verified proofs, not headlines. So here’s a question to sit with. If an AI-generated proof clears formal verification, does it matter whether the idea inside came from a human or a machine? Would you count it as real math either way?