Google DeepMind and OpenAI are matching step for step in AI breakthroughs such as Math Olympiad and ICPC gold medals, but Google seems to be in no mood to take kindly to OpenAI’s claims which it believes don’t have much basis.
Google DeepMind CEO Demis Hassabis has called OpenAI’s claims of having solved several Erdos problems as “embarrassing”. Erdos problems are a collection of unsolved and solved mathematical problems posed by the prolific Hungarian mathematician Paul Erdős, including famous ones like the Distinct Distances Problem and the Discrepancy Problem. These problems are known for their difficulty and are often the subject of intense study, research, and even cash rewards for solutions.
OpenAI employee Mark Sellke had earlier said that GPT-5 had been able to “find” solutions to as many as 10 Erdos problems. “Using thousands of GPT5 queries, we found solutions to 10 Erdős problems that were listed as open: 223, 339, 494, 515, 621, 822, 883 (part 2/2), 903, 1043, 1079,” he posted on X. “Additionally for 11 other problems, GPT5 found significant partial progress that we added to the official website: 32, 167, 188, 750, 788, 811, 827, 829, 1017, 1011, 1041. For 827, Erdős’s original paper actually contained an error, and the work of Martínez and Roldán-Pensado explains this and fixes the argument. The future of scientific research is going to be fun,” he had added.
OpenAI’s Sebastien Bubeck had commented on his findings. “Science acceleration via AI has officially begun: two researchers found the solution to 10 Erdos problems over the weekend with help from gpt-5 … PS: might be a good time to announce that Mark Sellke has joined OpenAI :-)” he had posted.
But Google DeepMind CEO Demis Hassabis wasn’t impressed with the claims. He posted “this is embarrassing” in a reply to Bubeck’s post.
When prompted by an X user for more information, Hassabis pointed them to a thread posted by Thomas Bloom, who owns the website ErdosProblems.com website, which the OpenAI employees had referenced. “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. The ‘open’ status only means I personally am unaware of a paper which solves it,” Bloom had posted on X.
As such, GPT-5 hadn’t actually solved any Erdos problems — it had found existing solutions on the internet to these problems that the owner of the site ErdosProblems.com hadn’t been able to find. The OpenAI employees had said that GPT-5 had “found” solutions to these problems, which was technically correct, but would’ve led people to believe that it had discovered these solutions independently.
And while there might be value in GPT-5 finding solutions to math problems on the internet better than a human maintainer of a website, the posts by OpenAI employees would’ve led most people to believe that GPT-5 had actually solved these difficult math problems. And with Google DeepMind calling out these claims in no uncertain terms, it shows that how competitive things currently are between Google and OpenAI as the two companies race to get AI to make newer and newer breakthroughs in science and math.