Alphabet, mother -in -law of Google, and Openai announced that their artificial intelligence models won gold medals during a global mathematics competition, marking a major advance in AI mathematical capacities in the race for the creation of systems capable of competing with human intelligence.
This is the first time that AI systems have crossed the score threshold required for the gold medal at the International Mathematics Olympiad (IMO), a competition intended for high school students.
The models of the two companies have solved five of the six proposed problems, exploiting generalist “reasoning” models capable of treating mathematical concepts in natural language, unlike previous approaches privileged by AI societies.
While Google Deepmind has collaborated with IMO so that its models are assessed and certified by the official committee, Openai did not officially participate in the competition. The start-up revealed on Saturday that his models had obtained a score equivalent to a gold medal on this year’s events, on the basis of the evaluations of three former medalists from the external IMO.
According to Junehyuk Jung, professor of mathematics at Brown University and guest researcher at Google Deepmind, this success suggests that AI could, within a year, be used by mathematicians to solve even irresolus research on the border of the domain.
“I think the moment when we can solve complex reasoning problems in natural language will open the way to unprecedented collaboration between AI and mathematicians,” Jung told Reuters.
Openai’s breakthrough has been made possible thanks to a new experimental model focused on the massive increase in calculation power during test resolution. According to Noam Brown, a researcher at Openai, this was achieved by allowing the model to “think” longer and by mobilizing a parallel calculation power to simultaneously explore multiple lines of reasoning. Brown refused to specify the cost of computer resources for Openai, however evoking a “very expensive” operation.
For Openai researchers, this result demonstrates once again that AI models can develop extensive reasoning capacities, likely to extend to other areas beyond mathematics.
Google researchers share this optimism, believing that the capabilities of AI models could apply to other research issues, especially in physics, according to Jung, itself gold medal at IMO in 2003.
Among the 630 students who participated in the 66e Edition of IMO, organized on Sunshine Coast in Queensland, Australia, 67 candidates, or about 11 %, obtained a score of gold medal level.
Last year, Google’s Deepmind unit obtained a silver medal thanks to AI systems specialized in mathematics. This year, Google used a general model called Gemini Deep Think, a version presented during its annual developer conference in May.
Unlike the previous attempts of AI based on formal languages and long calculations, Google’s approach this year was fully based on natural language and resolved problems within 4.5 hours, the company said in a blog post.
OPENAI, which has its own reasoning models, has also developed an experimental version for competition, according to a message from researcher Alexander Wei on the social network X. He said that the company did not plan to publish a model with such mathematical capacities soon.
This edition marks the first official collaboration of the competition with certain AI developers, who have been using prestigious mathematical competitions for years like IMO to test the capacities of their models. IMO judges certified the results of these companies, including Google, and asked them to publish their results on July 28.
“We respected the initial request for the IMO Council, which wanted all AI laboratories to share their results only after they were verified by independent experts and the students received the acclamation they deserved,” said Demis Hassabis, CEO of Google Deepmind, on X on Monday.
Openai, who published her results on Saturday and claimed the first the gold medal, said during an interview that he obtained the authorization of an IMO council member to do so after the closing ceremony on Saturday.
On Monday, the competition authorized partner companies to publish their results, confirmed Gregor Dolinar, president of the IMO Council, in Reuters.
