The AI found something humans had not, but humans had to make it matter.
Oitenta anos após Paul Erdős formular uma questão aparentemente simples sobre distâncias em um plano bidimensional, uma inteligência artificial de uso geral encontrou o caminho que matemáticos humanos não haviam trilhado. O modelo da OpenAI não foi treinado para resolver esse problema específico — ele chegou à solução por meio de uma conexão inesperada entre campos distintos da matemática, surpreendendo até os especialistas. Validado por pesquisadores externos de Cambridge e Manchester, o resultado levanta uma questão mais profunda do que qualquer teorema: o que significa descobrir algo, e quem — ou o quê — merece o crédito?
- Um problema aberto desde 1946 foi resolvido não por um matemático, mas por um sistema de raciocínio geral que ninguém havia programado especificamente para essa tarefa.
- A solução empregou conceitos da teoria algébrica dos números em um contexto geométrico — uma conexão que décadas de pesquisa humana não haviam estabelecido.
- A credibilidade da OpenAI estava em jogo: um episódio anterior, em que a empresa afirmou falsamente ter resolvido outros problemas de Erdős, deixou a comunidade matemática em alerta.
- Desta vez, a empresa inverteu a ordem — primeiro a validação externa, depois o anúncio — e matemáticos de peso, incluindo Tim Gowers, de Cambridge, confirmaram que o resultado merece publicação nos melhores periódicos do mundo.
- Ainda assim, pesquisadores humanos foram essenciais para refinar, interpretar e expandir o que a IA havia esboçado, mantendo a colaboração no centro da descoberta.
Na semana passada, a OpenAI anunciou que um de seus modelos de inteligência artificial havia resolvido o problema da distância unitária no plano — uma questão formulada pelo matemático húngaro Paul Erdős em 1946 e que resistiu a oitenta anos de tentativas humanas. O problema pergunta: em um plano bidimensional, qual é o número máximo de pares de pontos que podem estar separados por exatamente uma unidade de distância? O melhor limite superior conhecido havia sido estabelecido em 1984. O modelo da OpenAI encontrou configurações que superaram esse limite.
O que surpreendeu os especialistas não foi apenas a resposta, mas o caminho até ela. O sistema não havia sido treinado para matemática nem para esse problema em particular. Trata-se de um modelo de raciocínio geral que, segundo o artigo da empresa, aplicou conceitos da teoria algébrica dos números a um problema geométrico — uma conexão que os pesquisadores humanos não haviam feito. Thomas Bloom, da Universidade de Manchester, validou a solução em um artigo complementar, mas ressaltou que matemáticos humanos foram fundamentais para refiná-la e interpretá-la. A contribuição humana não se limitou à verificação: ela foi parte ativa da descoberta.
Tim Gowers, de Cambridge, classificou o resultado como um marco para a matemática gerada por IA, afirmando que recomendaria a publicação do artigo nos Anais de Matemática sem hesitação. A OpenAI aproveitou o endosso para argumentar que a IA pode contribuir para pesquisas de fronteira — mas foi cuidadosa em apresentá-la como ferramenta de apoio, não substituta dos matemáticos.
O anúncio ganhou peso adicional pelo contraste com o passado recente. Em outubro do ano anterior, executivos da OpenAI haviam afirmado que o GPT-5 havia resolvido dez problemas de Erdős ainda em aberto — afirmação que Bloom e outros rapidamente desmentiu, mostrando que os problemas já tinham solução conhecida. A declaração foi retirada. Dessa vez, a empresa inverteu a lógica: apresentou a validação externa antes do anúncio, transformando uma reivindicação em evidência.
OpenAI announced last week that one of its general-purpose AI models had solved a mathematical problem that had resisted solution for eighty years. The achievement concerns what mathematicians call the unit distance problem in the plane—a deceptively simple question posed by the Hungarian mathematician Paul Erdős in 1946. Before the announcement could settle, the company invited external mathematicians to scrutinize the work. They did, and they confirmed it was sound.
The problem itself asks: in a two-dimensional plane, what is the maximum number of point pairs that can be separated by exactly one unit of distance? Erdős suspected the answer would grow slightly faster than the total number of points involved, but establishing precise bounds has eluded mathematicians for decades. The tightest upper bound anyone had managed came in 1984. OpenAI's model found configurations that exceeded that limit.
What made the result noteworthy to the company was not just the answer, but the path taken to reach it. The model had not been trained specifically for mathematics or for this problem. It was a general reasoning system that, according to OpenAI's research paper, employed an approach considered novel—substituting a theory not typically associated with the unit distance problem. The concepts themselves were already familiar to specialists in algebraic number theory, but their application to geometric questions surprised researchers. The model had made a connection that humans had not.
Thomas Bloom, a mathematician at the University of Manchester who maintains a website tracking Erdős problems, reviewed the solution and confirmed its validity in a companion paper. Yet he was careful to note that the AI's initial demonstration, while correct, had been substantially refined by human researchers at OpenAI and by other mathematicians involved in the review. The human contribution remained essential—not just for checking the work, but for discussing it, interpreting it, improving it, and exploring what it meant.
Tim Gowers, a mathematician at Cambridge, called the result a landmark moment for AI-generated mathematics. If a human researcher had submitted this paper to the Annals of Mathematics, Gowers said, he would recommend acceptance without hesitation. No AI system had previously reached this level of rigor and novelty. OpenAI seized on the endorsement, arguing that the result demonstrated AI's capacity to contribute to frontier research, though the company was careful to frame the technology as a tool supporting mathematicians rather than replacing them.
The announcement carried weight partly because of what had come before. In October of the previous year, OpenAI executives including Kevin Weil and Sebastien Bubeck had claimed that GPT-5 had solved ten unsolved Erdős problems and advanced progress on eleven others. Mathematicians, including Bloom, quickly pointed out that these problems had already been solved by humans. Bubeck withdrew the claim and deleted his original post. That episode had left a mark on how the mathematical community received OpenAI's announcements. This time, the company had done the work of external validation first, presenting not just a claim but evidence that others had examined and endorsed.
Citações Notáveis
If this paper had been submitted by a human researcher to the Annals of Mathematics, I would recommend acceptance without hesitation.— Tim Gowers, Cambridge mathematician
The AI's demonstration was valid, but it was substantially improved by human researchers and other mathematicians involved.— Thomas Bloom, University of Manchester
A Conversa do Hearth Outra perspectiva sobre a história
Why does it matter that this wasn't a model trained specifically for mathematics?
Because it suggests the AI isn't just pattern-matching against a database of math problems. It's reasoning across domains. A general system solving a specialized eighty-year-old problem is more interesting than a specialized system doing what it was built to do.
But humans still had to check the work. Doesn't that undercut the achievement?
Not really. The checking matters, but so does the finding. A mathematician still has to verify a colleague's proof. That doesn't make the colleague's insight less real. What's different here is the speed and the novelty of the approach.
What about the October incident with the ten Erdős problems?
It was a credibility wound. OpenAI made claims that turned out to be false—the problems had been solved already. So this time they brought in external validators before announcing anything. That's the right move, but it also means people are watching more carefully now.
If AI can solve this, what can't it solve?
That's the question everyone's asking. The honest answer is we don't know yet. This problem was eighty years old and very specific. It doesn't mean AI is suddenly solving every open problem in mathematics. But it does mean the ceiling is higher than people thought.
Is the tool replacing mathematicians?
Not according to Bloom and Gowers. They're saying the AI found something, but humans had to refine it, interpret it, understand why it works. The tool amplifies what mathematicians can do. Whether it eventually replaces some of what they do—that's a different question, and it's not answered yet.