When AI Solves Open Math Problems, What's Left for Genius
In 1934, a Hungarian mathematician named Paul Erdős issued a challenge: solve this problem, collect $500. He posted hundreds of these bounties over his career — problems simple enough to explain in a sentence, hard enough to resist attack for decades. After he died in 1996, colleagues kept paying them out on his behalf, because that was the culture: some problems just sit there, daring you.
Here's the part that should make you sit up straight. In late 2025 and early 2026, researchers reported AI systems producing credible proof strategies for Erdős-style problems — including results strong enough that human mathematicians felt obliged to verify and, in at least one famous case, pay out the prize. Terence Tao, probably the finest living mathematician, has said publicly that AI is becoming genuinely useful on research-level problems, comparing the current moment to having a competent graduate student available around the clock.
A machine, collecting on Erdős's bounties. If you'd told me that five years ago I'd have assumed you'd confused science fiction with mathematics.
First, What "AI Reasoning" Actually Is
Before we argue about genius, let's be precise about what changed, because the phrase "AI solved a math problem" smuggles in a lot of assumptions.
Older language models were autocomplete with a vocabulary: they predicted the next word from patterns in training data. Asking one to prove a theorem produced something shaped like a proof — confident, fluent, wrong in step three. Mathematicians rightly ignored them.
The newer generation works differently. Instead of answering immediately, the model generates an internal chain of attempts, checks each against formal verification tools, backtracks when a step fails, and iterates — sometimes for hours — before presenting an answer. The industry calls this test-time reasoning or agentic search. Whatever the marketing name, the key shift is this: the model now shows its work, and the work can be mechanically checked.
That last clause matters more than people realize. A proof isn't persuasive prose; it's a certificate that a program like Lean can validate line by line. AI doesn't need to "think like a mathematician." It needs to produce output that survives formal verification — a completely different, and much lower, bar.
The question was never "can machines think." It was always "can machines produce artifacts good enough that nobody needs to care whether they think." We just crossed that line for mid-tier mathematics.
Why Erdős Problems Fell First
Not all math frontiers are equally exposed, and there's a clear logic to which ones went down first.
Erdős problems are famously easy to state and brutally hard to finish, but crucially, they sit inside existing theory. The tools needed to attack them are scattered across decades of combinatorics papers — known lemmas, known techniques, known dead ends. What's missing is not a new conceptual universe but exhaustive, patient exploration of an enormous space of possible arguments.
That is precisely the profile of work AI does well: huge search space, clear success criterion, no need for genuine novelty. Meanwhile the Millennium Prize problems — Riemann, P vs NP — likely require conceptual inventions that don't yet exist in anyone's training data, human or machine. Those aren't falling next year. Probably.
There's an underrated second mechanism here worth naming: rediscovery. Mathematics is so vast that results get proven, published in an obscure journal, and forgotten. Researchers have caught AI systems re-deriving known theorems their creators didn't know existed. When a machine can hold all of the literature simultaneously and never forget where anything is buried, it starts finding connections humans literally lost. That's not superintelligence. That's a perfect memory applied to an imperfect one — and it's enough to matter.
So What Happens to Genius?
Now the uncomfortable part, the thing this whole post is actually about. For centuries we defined mathematical genius as proof production: the person who could find the route nobody else could see, grind through the verification, deliver the theorem. Euler's productivity, Ramanujan's intuition, Erdős's sheer volume. The proof was the product.
When proofs become cheap, that definition of genius stops describing the scarce thing.
This has happened before, in a field most people don't take seriously enough as precedent. Chess engines surpassed Kasparov in 1997, and the obituaries wrote themselves: human chess was dead. Three decades later, chess has more players, more viewers, and more money than at any point in history. What died wasn't chess — it was the value of memorizing openings. What rose was everything machines can't price: creativity in unfamiliar positions, teaching, personality, narrative. Today's top players are stars because they're characters, not because they calculate better than Stockfish (they don't, and everyone knows it).
Mathematics is undergoing the same repricing with higher stakes, because new mathematical tools reshape physics, cryptography, and every quantitative field downstream. My honest read of where the value migrates:
- Choosing problems. Deciding which conjecture deserves three years of a lab's attention becomes the scarce skill when solving is commoditized.
- Framing. Reformulating a vague question into one a machine can attack is already more valuable than attacking it yourself.
- Taste and verification judgment. When ten candidate proofs arrive overnight, someone must know instantly which one smells right.
- Asking new questions. Machines answer; they do not yet wonder. The frontier of value moves to whoever decides what's worth wondering about.
The mathematicians I respect aren't depressed about this. The boring parts of their job were never why they loved it.
The Skeptic's Corner
Let me steelman the dismissive view properly, because it deserves better than a strawman. Maybe nothing interesting happened here at all. Maybe this is brute-force search wearing a graduation gown: enormous compute, parallel candidate generation, mechanical filtering. No insight, just scale. Deep Blue beat Kasparov with lookup tables and evaluation functions, and the world correctly noted that this wasn't "thinking."
Here's my response, and I hold it loosely. Human mathematicians also search — through an internal space of hunches, examples, half-remembered lemmas. We simply run the process slowly and call the successful runs intuition. If a machine explores the same space faster and its outputs pass the same peer review, insisting the label "intelligence" belongs only to carbon starts to look less like philosophy and more like brand loyalty. At some point the outcome is the meaning. But notice what this concession costs: if search speed at scale counts as mathematical ability, then most of what made calculators special is now reproducible. Either way, genius-as-calculation loses.
Why Someone Running Servers Should Care
If this blog were only about mathematics, I'd have stopped two sections ago. But Watchtower readers live in infrastructure, and the identical shift is happening under our feet.
AI agents now monitor server farms, predict drive failures from vibration signatures, rebalance cooling loads across racks — work that five years ago required a senior SRE's pattern recognition. Nobody says those agents "understand" thermodynamics. They produce decisions that survive contact with reality, verified by outcomes instead of Lean proofs. Same structure as the math story: execution gets automated, judgment gets promoted, and the humans who thrive are the ones who move up the stack — choosing what to monitor, deciding what anomalies matter, asking questions the dashboards never thought to ask.
The 'how' is being eaten by machines everywhere. The 'what' and the 'why' are appreciating assets.
So, same question as always, and I mean it seriously this time: in your own work, what fraction of your week is proof-production — executing a process someone could automate — versus problem-selection? Be honest. Then decide which side of that line you want to be standing on in five years.
Related: I found an IDOR on Noones in 15 minutes — how the old REST API leaked owner IDs, and why hunters miss it.
🔎 Found this useful? Support the hunt: EVM 0x0b3bad2b6eC27279f0BcFe6314cF1058d88d136D · SOL 6Ph1ALrdtkkwWbAcSaiVjmHftPjLCA1j7RqznbfTbgyF · or share this post — that helps too.
Disclosure: some links here are affiliate links. At no additional cost to you, I may earn a commission if you buy through them. I only recommend things I have actually used.
Comments
Post a Comment
Thanks for reading! Share your thoughts below