OpenAI Dumps 722 Math Papers Overnight — Riemann Hypothesis, Hodge Conjecture, and Mathematicians Are Losing It

📌 Key Takeaway: OpenAI released 722 mathematical research papers in one night, covering 372 distinct research results across frontier topics including the quasi-Riemann conjecture, Hodge conjecture, and BSD conjecture. Each paper consumed roughly the compute of 3 hours of ChatGPT Pro. All papers, source code, and partial Lean formalized proofs are publicly available on GitHub.

The math world blew up yesterday.

Not because someone proved Goldbach's conjecture — but because OpenAI, the company behind ChatGPT, quietly dropped 722 math papers onto GitHub. Overnight.

That's 722. Not conference filler. Real mathematical research manuscripts covering 372 independent results. The topics read like a greatest-hits list of unsolved problems: quasi-Riemann conjecture, Hodge conjecture, BSD conjecture — each one a problem that has kept mathematicians awake for decades.

Why Is an AI Lab Doing Math?

Here's the context.

These papers aren't casual side projects. They come from an unreleased internal model built specifically for mathematical research. Think of it as an "AI mathematician" that worked around the clock and produced 722 papers worth of output.

The system attempted roughly 4,000 mathematical problems — not textbook exercises, but genuine frontier-level questions. About 372 of those attempts produced publishable results. That's roughly a 1-in-10 success rate.

Honestly, that's better hit rate than most human PhD students.

Each Paper Costs 3 Hours of ChatGPT Pro

The compute cost is another jaw-dropping number.

According to OpenAI, the average computational cost per paper is roughly equivalent to running ChatGPT Pro continuously for 3 hours. Multiply that by 722 papers, and you're looking at an astronomical compute bill.

But OpenAI clearly doesn't care about the cost. Because what they're trying to prove isn't "how much money we can burn" — it's "can AI actually do real science?"

The answer: at least in certain areas of mathematics, yes.

Some results in this collection have been formalized and verified using Lean — a proof assistant that acts like a compiler for mathematical proofs. If a proof passes Lean verification, it's rock-solid. The fact that OpenAI included partial Lean formalizations shows they're not just generating text — they're generating verifiable mathematics.

Everything Is Open-Source. Seriously.

What's really shaking the academic world isn't the volume — it's the attitude.

OpenAI released all 722 papers, all source code, and partial Lean proofs to GitHub. Anyone can read, verify, and build on top of these results.

This move is smart for two reasons.

First, open source means transparency. "You claim 722 papers? Here, look at them yourself. Verify them." It shuts down every "this is just AI hype" argument before it starts.

Second, open source is an invitation. Mathematicians worldwide can now collectively review these results — poking holes, filling gaps, extending ideas. The 722 papers may just be the starting point for an entirely new paradigm of AI-assisted mathematical research.

What Does AI Doing Math Actually Mean?

The obvious question: will AI replace mathematicians?

Not anytime soon. Most of these 722 papers are "half-step forward" work — finding new proof paths within existing frameworks, generalizing results to new settings, discovering new structural relationships. The kind of work that requires creativity at the "zero to one" level? AI can't do that yet.

But medium to long term, things change fast.

When an AI system can simultaneously attempt 4,000 math problems and produce 722 papers overnight, the role of human mathematicians shifts from "producers" to "curators" and "direction-setters." You tell the AI which direction to dig, and the AI does the digging.

Once this model matures, mathematical research efficiency could improve by an order of magnitude.

What OpenAI did here is fire a shot across the bow for the entire scientific community: AI isn't just good at writing blog posts and making slides. It can tackle hard science.

The question now is how the math world responds.