The Problem
In 1946, Paul Erdős posed a question about dots on paper. Place a bunch of dots on a surface. Arrange them so that the maximum number of dot pairs are exactly one unit apart. Simple to state. Impossible to solve , for eighty years.
Think of it like this: arrange a billion people at a party so as many people as possible are exactly one arm's length from someone else. Now prove mathematically that your arrangement is the best possible one. Every mathematician who tried since 1946 failed.
OpenAI's internal reasoning model just solved it.
Why This Is Different From the Last Time
Seven months ago, OpenAI claimed to have solved Erdős problems and got mocked. The reason: the solutions were already in the training data. The model was not reasoning. It was reciting.
This time, mathematicians independently verified the result. Not just checked the answer , verified the proof. The model discovered an infinite family of new solutions. This means it found a new class of answers that extends infinitely, not just one specific arrangement.
The method is what matters. The model did not brute-force through combinations. It connected ideas from algebraic number theory to a basic geometry problem about dot placement. Two fields that have nothing obvious to do with each other. The comparison the analysts used: like fixing your car engine using techniques borrowed from origami.
That is not what a sophisticated autocomplete does. That is what a thinking entity does.
What This Actually Means
OpenAI has been careful about the framing. The capabilities being demonstrated in public products have not matched the ambition of some internal claims. The math community has been particularly skeptical , previous AI math announcements had been debunked quickly.
This one has not. The proof was submitted to mathematicians who had no stake in the outcome and was not dismissed.
The ceiling question in AI research has always been: can these systems generate genuinely novel knowledge, or can they only recombine things they have already seen? The Erdős result is the strongest evidence yet that the answer to that question is changing.
Greg Brockman said OpenAI believes they are "70 to 80 percent of the way to AGI." Whether that framing is meaningful is debatable. Whether an AI system just solved an eighty-year-old open mathematical problem by combining ideas across disciplines is not debatable.
The result exists. The verification exists. That is what it is.