An OpenAI Reasoning Model Just Disproved an 80-Year-Old Erdős Conjecture. Mathematicians Verified the Proof. Here's What Changes for Knowledge Workers Now.

On May 20, 2026, OpenAI announced that one of its general-purpose reasoning models produced an original mathematical proof disproving the planar unit-distance conjecture, a problem Paul Erdős posed in 1946 and that has resisted resolution for eighty years. The model's argument connected the geometry problem to algebraic number theory, used techniques from that field to construct an infinite family of counterexamples, and produced a polynomial improvement on the previous best lower bound.

Timothy Gowers, a Fields Medalist at Cambridge, verified the result. So did Thomas Bloom, who seven months ago publicly criticized GPT-5 for fabricating math proofs. This time the external mathematicians wrote a companion paper explaining the argument and providing context for the significance.

If you sell your brain for a living (strategy work, legal analysis, financial modeling, research, code architecture, anything where the value proposition is "I think about hard problems for you"), the timeline you assumed you had just got shorter. Here's the honest read.

What actually happened, and what didn't

The Erdős unit-distance problem asks: if you place n points in the plane, how many pairs of those points can be exactly distance 1 apart? In his 1946 paper, Erdős showed that a √n × √n piece of the integer lattice gives a lower bound of roughly n^(1+c/log log n) unit pairs, and he conjectured this matched the true magnitude, i.e., that the upper bound is the same shape. The conjecture has stood as one of the central open problems in discrete geometry for eighty years.

The OpenAI result is a counterexample that pushes the lower bound up: a new infinite family of point configurations that exceeds Erdős's lattice construction by a polynomial factor and, as several mathematicians have noted, disproves the conjecture in the form Erdős most likely intended. The novelty is not the size of the improvement. It's that the configurations come from algebraic number theory, specifically from algebraic integers in number fields with specific factorization properties. Geometry problems being solved by number-theoretic constructions is a known move in mathematics; the OpenAI model finding the move autonomously is the unprecedented part.

This is materially different from the GPT-5 episode last October. Then, OpenAI claimed GPT-5 had solved several long-standing open problems, including some by Erdős. Bloom and others quickly showed the proofs were not novel. The model had retrieved existing literature solutions and presented them as original. The episode set the credibility bar for any future "AI does math" claim. The May 2026 result clears that bar because the construction did not previously exist in the literature, the argument is mathematically correct, and external mathematicians verified it independently.

What the result is not: a single proof of a famous conjecture by a model is a single data point. Research mathematics is not just isolated combinatorics problems. Most working math involves theory-building over years, not a single combinatorial flash. The wet-lab parts of science, the field-work parts, the parts that require physical intuition or experimental design: those are not affected by this result. Yet.

The capability that moved

For two years, the consensus position on AI in research-class work has been: AI can summarize, retrieve, code at the level of a strong junior engineer, and pattern-match against existing solutions. AI cannot produce novel research-class output autonomously. The Erdős result moves the second claim.

The specific capability that moved is something the OpenAI announcement calls "cross-domain transfer": the ability to recognize that a problem in one mathematical field has structural similarities to a problem in a different field, and to import techniques across the boundary. This is the specific skill that distinguishes mid-tier mathematicians from research-class mathematicians. It is also the specific skill that has historically been the human moat in knowledge work generally.

The argument the model produced uses techniques from algebraic number theory (specifically, properties of algebraic integers in extension fields with non-trivial unit groups) to construct points in the plane whose pairwise distances have algebraic constraints that classical lattice constructions can't replicate. The trick is genuinely clever. It is the kind of trick a research mathematician would be pleased with.

If a model can do this for a 1946 problem in discrete geometry, the question is which research-adjacent fields the same capability transfers to. Bloom's companion paper explicitly notes that the construction has implications for incidence geometry beyond the unit-distance problem. Whether cross-domain transfer generalizes to applied research (chemistry, materials science, drug discovery) is the question I'd most want answered in the next six months.

What this means for "I'm paid to think"

I work as a solo operator, and a lot of what I sell is variations on "I'll think about your hard problem and give you a structured answer." Strategy work for portfolio companies, product spec writing, technical architecture reviews. The Erdős result does not mean my job disappears tomorrow. It means the timeline I had in my head was wrong.

The honest read is in three parts.

Research-class knowledge work has a moat. The moat is some combination of (a) tacit knowledge that doesn't appear in training data, (b) cross-domain transfer that requires real-world context, and (c) the social capital to be trusted with the question in the first place. The Erdős result attacks (b) directly. It does not attack (a) or (c) yet. The bet I was making, that (b) was safe through at least 2028, is now wrong.

Productized research output gets cheaper before commoditized research output disappears. The first thing that happens after a capability lands in a frontier model is that it becomes an API call. The second thing that happens is that an indie product wraps the API call. The third thing is that the indie product is replaced by a vertical SaaS at a tenth the price. That sequence usually takes 12-24 months. The Erdős result is at step zero. It's still inside OpenAI's internal model. The API-call version is months away. The indie wrapper is a year away. The vertical SaaS is 18 months from now.

The window to reposition is concrete. If your billable hourly work today is "I do research-class thinking for clients who can't or won't do it themselves," the window between now and "your client can get a defensible version of the same output from a $200/month SaaS" is somewhere between 12 and 30 months. Not zero. Not forever. The middle scenario you should plan for is 18 months.

What I'd actually do

I'm doing two things in my own consulting work, both small, both immediate.

One is shifting my pitch from "I think about the problem for you" to "I get the thinking actually used inside your organization." The Erdős result moves the technical-thinking part down the value chain. It does not move the political-implementation part down, because that part requires being inside an organization and earning trust from people who don't trust AI to make decisions for them yet. The work that survives is the work where a Claude-generated PDF cannot replace a human in the room.

Two is being much more public about my reasoning process, not just my conclusions. The reason research-class human work has historically been hard to replace is that the work isn't actually the conclusion. It's the path to the conclusion, the dead ends, the abandoned hypotheses, the part where you talk to a domain expert and get the corner-case nobody had thought of. If your output is a polished deliverable, AI replaces you faster than if your output is "here's how I got here and here's what I learned along the way." The version of consulting that survives is the version where the client is buying the journey, not just the destination.

If you sell your brain for a living, here's the test I'd run on yourself. Look at your last five paid deliverables. For each one, ask: could a sufficiently advanced AI, fed the same inputs, produce a deliverable a sophisticated client could not distinguish from yours? If the answer is yes for three or more, you have a repositioning problem this year, not a 2028 problem. If the answer is no for all five, the parts of your work that are AI-resistant are the parts you should be making more visible to clients.

Where this argument could be wrong

I want to be clear about the parts of this that are not load-bearing.

A single proof is not the same as sustained research capability. The OpenAI model produced one verified result. It did not produce the second one this week. Until reasoning models produce a stream of novel results across multiple fields, the right read is that the capability exists but isn't yet reliable enough to be a research tool you can depend on. Bloom himself made this point in his commentary: the result is real, but generalizing from one data point to "AI mathematicians" is premature.

The economic transmission is also slower than the capability advance. Even when AI can do research-class work autonomously, the institutional friction in industries that buy research-class work (law firms, consultancies, financial services, government) means the adoption curve is years, not months. The lawyer who could have been replaced by AI in 2024 is still billing in 2026 because the law firm structure and the client expectations haven't caught up. That friction is real and it buys time.

It's also possible that the next round of AI-math claims regresses. OpenAI has a track record of overclaiming. The GPT-5 episode is recent enough that the institutional memory should make this announcement more cautious than the last one, but the company has commercial reasons to overstate capability. I would be more confident if a second model from a different lab produced an independent novel proof within 60 days. If that doesn't happen, this is a single capability data point, not a trend.

The honest take

The Erdős result is the moment I would mark as the start of the credibility break on "AI can't do real research." It is not the moment knowledge work as a category becomes obsolete. The gap between those two things is roughly where my consulting work needs to live for the next 18 months.

If you sell your brain, the action is not panic. The action is to read the OpenAI announcement, read Bloom's companion paper, and then look at your own work with fresh eyes. The work that survives the next two years is the work that's not legible to a model trained on text. The work that doesn't survive is the work you could already write a prompt for. Most of us have some of both. The repositioning is moving more of your hours into the first bucket while you still have the time to do it.

Author

Lukas

@lukcombinator

Sources

• An OpenAI model has disproved a central conjecture in discrete geometry — OpenAI
• OpenAI claims it solved an 80-year-old math problem — for real this time — TechCrunch
• AI just solved an 80-year-old 'Erdős problem,' and mathematicians are amazed — Scientific American
• 80-year-old geometry puzzle cracked by OpenAI using number theory — Interesting Engineering
• OpenAI solves 80-year Erdős geometry problem — explainx.ai
• Amazing: Erdős' Unit Distance Problem was Disproved! — Gil Kalai, Combinatorics and More