While 630 young math prodigies were sitting in a conference room on Australia’s sunshine coast, readying their pencils for the International Math Olympiad, a potential rival was still en route from the Brisbane airport: a team from Google, with an AI model tuned specifically to complex math.
For months, the Googlers had been training a variant of the company’s Gemini large language model, DeepThink, to solve Olympiad problems. For DeepMind, the contest was more than a chance to beat smart teens — in fact, the team winces at the suggestion that they’re in competition. Instead, it’s about taking a step on the path to a theoretical milestone known as “artificial general intelligence,” the point at which AI could be relied upon to think like a human, and be trusted with all sorts of tasks.
AGI “is going to be one of the biggest technologies, if not the biggest, that humanity will ever invent,” Demis Hassabis, the DeepMind CEO, told Bloomberg News in September.
The problems at contests like the Olympiad are designed to be novel, demanding creativity, long chains of reasoning and the ability to recover from false starts. They showcase the strengths of the brightest humans and the weaknesses of machines, which so far fall short when it comes to abstraction, multi-step planning and unsolved problems. That’s part of why Hassabis has become so focused on conquering them, entering Google in math and coding contests around the world. The stakes are high: At a moment when critics say progress has stalled, excelling in the contests could be proof their systems are still improving. Rival OpenAI does the same for its own system, ChatGPT.
2 Comments
*Janne Knodler for Bloomberg News*
While 630 young math prodigies were sitting in a conference room on Australia’s sunshine coast, readying their pencils for the International Math Olympiad, a potential rival was still en route from the Brisbane airport: a team from Google, with an AI model tuned specifically to complex math.
For months, the Googlers had been training a variant of the company’s Gemini large language model, DeepThink, to solve Olympiad problems. For DeepMind, the contest was more than a chance to beat smart teens — in fact, the team winces at the suggestion that they’re in competition. Instead, it’s about taking a step on the path to a theoretical milestone known as “artificial general intelligence,” the point at which AI could be relied upon to think like a human, and be trusted with all sorts of tasks.
AGI “is going to be one of the biggest technologies, if not the biggest, that humanity will ever invent,” Demis Hassabis, the DeepMind CEO, told Bloomberg News in September.
The problems at contests like the Olympiad are designed to be novel, demanding creativity, long chains of reasoning and the ability to recover from false starts. They showcase the strengths of the brightest humans and the weaknesses of machines, which so far fall short when it comes to abstraction, multi-step planning and unsolved problems. That’s part of why Hassabis has become so focused on conquering them, entering Google in math and coding contests around the world. The stakes are high: At a moment when critics say progress has stalled, excelling in the contests could be proof their systems are still improving. Rival OpenAI does the same for its own system, ChatGPT.
[Read the full dispatch here.](https://www.bloomberg.com/news/articles/2025-10-03/google-openai-use-math-contests-to-prove-ai-progress?accessToken=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJzb3VyY2UiOiJTdWJzY3JpYmVyR2lmdGVkQXJ0aWNsZSIsImlhdCI6MTc1OTUxMDg1NywiZXhwIjoxNzYwMTE1NjU3LCJhcnRpY2xlSWQiOiJUM0pZS0VHUEZIVUIwMCIsImJjb25uZWN0SWQiOiJEMzU0MUJFQjhBQUY0QkUwQkFBOUQzNkI3QjlCRjI4OCJ9.-TPKUqPEqOo8YOzWNFDhZ4OD8DCT4FVVRmOhuAg7n0o)
On one hand they say it will lead to AGI and on the other they have fine tuned the model for the competition.