Here’s a sentence I never expected to write: an AI chatbot just helped solve a math problem that’s been open since 1983. Not by magically producing the answer, but by doing something arguably more interesting — being a genuinely useful research collaborator.
Ernest Ryu, a mathematician at UCLA (currently on leave as a Member of Technical Staff at OpenAI), along with PhD student Uijeong Jang, has proved the pointwise convergence of Nesterov’s Accelerated Gradient method — an optimization algorithm that’s been a workhorse of machine learning for four decades, but whose fundamental stability properties nobody could pin down.
The tool? ChatGPT running on GPT-5 Pro. The time? About 12 hours of active work over three days. The catch? It’s not what the headline makes it sound like.
The Problem That Wouldn’t Die
Nesterov’s Accelerated Gradient (NAG) method, introduced by Yurii Nesterov in 1983, is one of those algorithms that just works — so well that it’s embedded in practically every modern ML training pipeline. It adds momentum and a “lookahead” step to gradient descent, making optimization dramatically faster.
What nobody could prove for 42 years was whether NAG actually converges pointwise — that is, whether it settles on a single solution or oscillates forever. It’s the kind of foundational question that sounds abstract but matters enormously: if you can’t prove convergence, you can’t fully trust the algorithm’s stability, especially in safety-critical applications like medical imaging.
As Ryu put it: the problem had been “stubbornly resistant” to every technique mathematicians threw at it. Decades of attempts. No proof.
How AI Actually Helped
Ryu tried using earlier GPT models on this problem and got nowhere. GPT-5 Pro changed the game — but not in the way you might expect.
The workflow wasn’t “ask ChatGPT, get answer.” It was more like having a brilliant but unreliable research partner who throws out 20 ideas, 16 of which are wrong — but the 4 good ones are directions you’d never have thought of alone.
Specifically, GPT-5 Pro helped by:
- Generating creative proof strategies — including some “completely out of the blue” approaches Ryu hadn’t considered
- Restructuring equations — finding equivalent forms that made the proof tractable
- Cross-referencing literature — pulling connections from disparate mathematical fields
- Rapidly testing paths — filtering dead ends in minutes instead of the hours manual exploration would take
About 80% of GPT-5’s suggestions were dead ends. But the 20% that worked formed the backbone of the proof. Ryu and Jang then did the hard mathematical work of rigorously verifying, cleaning up, and formalizing everything.
The resulting paper — “Point Convergence of Nesterov’s Accelerated Gradient Method: An AI-Assisted Proof” — is careful about attribution. GPT-5 is credited as a tool. Not a co-author. Not an inventor. A tool that happened to be unusually good at mathematical brainstorming.
Why This Matters More Than You Think
This story is a Rorschach test for how you feel about AI. Let me offer my reading.
The optimistic take: AI just accelerated mathematical research by 3-10x, according to Ryu’s own estimate. A problem that sat open for 42 years fell in days. If this scales — and it will — we’re looking at a step-change in how quickly scientific breakthroughs happen. Terence Tao, one of the world’s greatest living mathematicians, highlighted the work as a significant example of AI assistance in mathematics.
The realistic take: This worked because Ryu already understood the problem deeply enough to recognize which of GPT-5’s ideas were worth pursuing. A less experienced researcher might have chased the 80% dead ends forever. The AI didn’t replace expertise — it amplified it. There’s a big difference.
The uncomfortable take: OpenAI, which makes ChatGPT, is also Ryu’s employer. They celebrated the result on their blog as evidence of GPT-5’s mathematical capabilities. That’s not wrong — but it’s also marketing. The proof stands on its own mathematical merits regardless, and Ryu’s academic reputation backs it. Still, read the OpenAI blog post with appropriate skepticism.
What This Means for NZ
For New Zealand’s research institutions, this is a wake-up call that’s not about replacing researchers — it’s about who gets amplified. The universities and CRIs that figure out how to integrate AI collaboration into their research workflows will produce breakthroughs faster. The ones that don’t will fall behind.
This also has implications for how we teach mathematics and science. The skill of the future isn’t just knowing math — it’s knowing how to productively collaborate with AI on mathematical reasoning. That means teaching students to evaluate AI suggestions critically, recognize productive vs. unproductive paths, and maintain rigorous verification standards.
And for a country that punches above its weight in research but has limited resources, a 3-10x acceleration tool isn’t a nice-to-have. It’s a competitive necessity.
🔍 The Bottom Line
A 42-year-old math problem didn’t get “solved by AI.” It got solved by a mathematician who had a tool that made him dramatically more creative and efficient. The distinction matters because it tells us where AI is genuinely powerful right now — not as an autonomous genius, but as an amplifier of human expertise.
The real breakthrough here isn’t the proof itself (though mathematicians are rightly excited). It’s the demonstrated workflow: human expertise + AI creativity + human verification = faster science.
That formula is going to reshape research. The question isn’t whether it happens. It’s whether you’re ready for it.
