AI Is Coming for Mathematics, and the Rest of Science Is Next

Something remarkable happened in the span of a single weekend.

Last week, OpenAI disproved a discrete geometry conjecture that had resisted 80 years of attempts by the world’s best mathematicians. Days later, Google DeepMind followed up by solving 9 of 353 open Erdős problems - a set of challenges that have stumped the mathematical community for decades. The math world is floored. And it should be.

But this isn’t just a story about mathematics. It’s a signal about where AI is headed next.

Why AI conquered software first

To understand why these math breakthroughs matter so much, it helps to understand why AI made its biggest early splash in software engineering.

The answer comes down to two things:

• Verifiability
• Money

Software is a uniquely feedback-rich environment. When you write code, you know almost immediately whether it works. Tests pass or fail. Code runs or crashes. There’s no ambiguity, no waiting for a peer review cycle, no years-long experiment to validate a hypothesis. The feedback loop is instant and unforgiving, and that’s exactly what makes it ideal for AI systems to learn from.

Combine that with the enormous commercial value of software, and you have a perfect storm. AI ate software engineering before it touched biology or physics, not because software is more intellectually interesting, but because it offered the clearest signal and the biggest reward.

The common thread: ground truth

Here’s what’s easy to miss: software isn’t uniquely special. It just happened to be the first domain where AI found a reliable source of ground truth.

And ground truth is the key ingredient.

Formal mathematical proofs have ground truth: a proof is either valid or it isn’t. Protein structures have ground truth: a folded protein either matches experimental data or it doesn’t. Materials science has ground truth: a material either has the predicted properties or it doesn’t. Physics simulations have ground truth: the simulation either matches observed reality or it doesn’t.

These fields have always had the intellectual rigor. What they lacked was the computational scale, and the AI systems capable of navigating their vast solution spaces. That’s changing fast.

Breakthroughs at the cost of hundreds of dollars

Perhaps the most striking detail in the recent math breakthroughs is the cost. These solutions, to problems that stumped human experts for generations, reportedly cost only hundreds of dollars each to compute.

That’s not a typo.

The economics of scientific discovery are being fundamentally rewritten. Historically, a major mathematical or scientific breakthrough might require years of a researcher’s time, expensive lab equipment, or massive computational clusters. Now, some of those same breakthroughs are within reach of a well-crafted prompt and a modest compute budget.

This doesn’t diminish the achievement. If anything, it amplifies it. When the marginal cost of a breakthrough approaches zero, the rate of breakthroughs can approach infinity.

What comes next

The implications extend far beyond mathematics. Every scientific domain with a clear notion of correctness, where you can verify whether an answer is right or wrong, is now in play.

Drug discovery. Climate modeling. Quantum chemistry. Genomics. Structural engineering. These are all fields where AI systems can, in principle, generate hypotheses, test them against ground truth, and iterate at machine speed.

We are likely at the beginning of a period where AI-driven scientific discovery becomes routine rather than remarkable. The math breakthroughs of the past week may look, in hindsight, like the first tremors before a much larger shift.

Exciting times, indeed.