Google DeepMind says it has achieved a milestone in artificial intelligence with its Gemini 2.5 model, claiming a “historic” breakthrough in abstract problem solving. The system became the first AI to win a gold medal at an international programming competition earlier this month in Azerbaijan. DeepMind likened the moment to IBM’s Deep Blue beating Garry Kasparov at chess in 1997 and its own AlphaGo’s victory against a human Go champion in 2016, placing it among the most notable milestones in the field.
The Gemini 2.5 model solved a highly complex real-world challenge that none of the human teams, including leading competitors from Russia, China, and Japan, were able to crack. It completed the task in under half an hour by calculating how to channel a liquid through a network of ducts and reservoirs as quickly as possible while evaluating an infinite number of possibilities. Although the model failed two of the 12 tasks in the contest, its overall performance ranked it second out of 139 of the world’s top college-level programmers, an outcome DeepMind described as a “profound leap” in AI’s ability to reason abstractly. Quoc Le, vice president at Google DeepMind, compared the achievement to Deep Blue and AlphaGo but emphasised that Gemini 2.5 was working on less constrained, more real-world problems with potential applications in areas such as drug and chip design.
DeepMind said Gemini 2.5 is a general purpose system specially trained for complex coding, mathematics, and reasoning challenges, performing “as well as a top 20 coder in the world”. In a statement, the company noted that such contests demand creativity, deep reasoning, and the ability to synthesise original solutions to problems never seen before. However, some experts urged caution. Stuart Russell, professor of computer science at the University of California, Berkeley, said claims of epochal significance might be overstated and pointed out that previous AI milestones such as Deep Blue had little direct impact on applied AI. He acknowledged, though, that solving an International Collegiate Programming Contest question correctly demonstrates progress towards building AI coding systems that are sufficiently accurate for high-quality software development.
Michael Wooldridge, Ashall professor of the foundations of AI at the University of Oxford, said the achievement sounded impressive and that solving problems at this level was exciting, but raised questions about how much computing power was required. Google did not disclose exact figures, only noting it was beyond what an average subscriber to its $250-a-month Google AI Ultra service could access using the lightweight version of Gemini 2.5 Deep Think in the Gemini App. Dr Bill Poucher, executive director of the ICPC, called the result a key moment in defining AI tools and academic standards for the next generation. DeepMind’s accomplishment follows a line of historic machine intelligence milestones including Rosenblatt’s Perceptron in 1957, IBM’s Deep Blue in 1997, AlphaGo in 2016, and AlphaFold’s advances in protein folding by 2020, underscoring a growing trend of AI systems competing directly with humans in tasks requiring skill and ingenuity.
Follow the SPIN IDG WhatsApp Channel for updates across the Smart Pakistan Insights Network covering all of Pakistan’s technology ecosystem.
