In a very public, very awkward lesson on the difference between “solving” a problem and simply “finding” the answer, OpenAI’s latest model, GPT-5, has stirred up controversy.
Top researchers at the company took a premature victory lap on social media, claiming the AI had cracked a set of notoriously difficult mathematical riddles. The celebration, however, was short-lived, earning the AI giant a stern rebuke and a fresh wave of industry ridicule.
It started October 12, when Sebastien Bubeck, a former Microsoft vice president and now a member of OpenAI’s technical staff, posted on X: “GPT-5-pro is superhuman at literature search. It just ‘solved’ Erdős problem No. 339 … by realizing it had been solved 20 years ago.”
The post sounded modest enough, but things soon snowballed. Mathematician Mark Sellke followed up, saying GPT-5 had identified 10 Erdős problems that were “listed as open,” and had made “significant partial progress” on eleven others.
Kevin Weil, OpenAI’s VP of Science, then amplified the excitement with a now-deleted post declaring, “GPT-5 just found solutions to 10 (!) previously unsolved Erdős problems, and made progress on 11 others. They have all been open for decades.”
It sounded like GPT-5 had made history. But the math community wasn’t buying it.
When ‘open’ doesn’t mean unsolved
Mathematician Thomas Bloom, who runs the Erdős Problems website, quickly poured cold water on the celebrations. Responding on X, Bloom wrote, “Hi, as the owner/maintainer of http://erdosproblems.com, this is a dramatic misrepresentation. GPT-5 found references, which solved these problems, that I personally was unaware of.”
He clarified that when a problem is marked as “open” on his website, it simply means he hasn’t seen a paper that solves it, not that the problem is still unsolved. In other words, GPT-5 didn’t actually produce new proofs; it merely discovered papers that already contained the solutions.
Even Google DeepMind’s CEO, Demis Hassabis, chimed in, calling the situation “embarrassing.” And Meta’s AI chief Yann LeCun took a sharper jab, quipping, “Hoisted by their own GPTards.”
Subtracting the hype, finding the real story
Despite the drama, the underlying story still holds value. GPT-5 did show impressive capability, not as a mathematician, but as a literature detective. It managed to surface obscure papers that even a domain expert had missed.
As the renowned UCLA mathematician Terence Tao pointed out, AI’s real power in math lies not in solving open problems, but in accelerating the grunt work: literature review, pattern search, and hypothesis generation.
That’s exactly what GPT-5 did: it “industrialized” the research process. Just not in the way its users first claimed.
A teachable moment
The incident is another reminder that AI’s achievements should be checked as carefully as a math proof. For OpenAI, it’s a humbling equation: Overconfidence + poor phrasing = one big miscalculation.
GPT-5 didn’t unlock new math, but it did open a discussion about how hype, even when unintentional, can distort scientific credibility.
We live in divisive times. OpenAI has halted the use of Martin Luther King Jr.’s likeness in Sora after users created fake clips depicting the civil rights leader in “disrespectful” ways.
The post Math-etic! GPT-5’s Math ‘Breakthrough’ Gets F-Minus for Hype appeared first on eWEEK.
No Responses