AI Solves Erdős Problems Without Explaining Them

The Erdős problems are falling to artificial intelligence, and everyone agrees this matters. What nobody has adequately defended is what "falling" actually means.

Paul Erdős spent the second half of the twentieth century scattering numbered conjectures across mathematics like someone leaving breadcrumbs. These were questions he found genuinely hard, often modest in appearance but resistant to the obvious approaches. A conjecture about the chromatic number of the plane. A claim about prime gaps. Problems that sat unsolved for decades, accumulated reputations as markers of mathematical substance. Then AI systems began producing answers. DeepMind's breakthrough on the graph coloring problem. GPT-derived solutions to combinatorial questions. The field celebrated this as evidence that machine learning could penetrate the deepest mathematical structures, that AI was doing what humans could not.

But the field has embedded an assumption in that celebration, and the assumption is doing all the work. When we say an Erdős problem has fallen, we mean a machine produced an answer that can be verified as correct. What we've quietly allowed is for "correct answer" to mean something radically different from what Erdős demanded.

The Erdős problems worth his attention typically required not just existence proofs. Constructive methods, proofs that illuminated why the answer was true, the logical architecture underneath. The mathematician wanted insight. The system we've built celebrates solutions that verify numerically or probabilistically without yielding that structure. The machine found that a certain combinatorial object exists, or that a bound holds true. It did not explain the mechanism. It did not teach us how to think differently about the problem.

This is not a failure of the machine. It is a redefinition of victory. AI excels precisely at problems where a numerical answer suffices, where verification can happen without understanding. The problems it cannot crack are those where humans insisted on knowing the shape of the solution, not just that it exists. We have learned something about AI's capacities, but we have learned it by changing the question mid-measurement.

This happens constantly in organizations when teams celebrate "progress" on metrics nobody defined together before the project started. A product launches and users engage with it. Engagement becomes the measure of success, even though nobody agreed that engagement was the goal. The system optimizes toward what it can track, and we call the tracking system alignment. By the time you realize the metric was wrong, it's generating its own incentive structure and defending itself.