Google DeepMind claims 'historic' AI breakthrough in problem solving (www.theguardian.com)

Google DeepMind says a specially trained version of its Gemini 2.5 model achieved a milestone at an international programming contest in Azerbaijan, becoming the first AI to earn a gold-level result at that level. The model solved a previously intractable real-world task — routing liquid through a network of ducts and interconnected reservoirs to minimize distribution time — by reasoning over an effectively infinite space of possibilities in under 30 minutes. Although it failed 2 of 12 tasks, Gemini’s aggregate performance placed it among the top competitors (Google described it as “as well as a top 20 coder in the world”), prompting DeepMind to frame the result as a “historic” leap toward more general abstract problem-solving and potential applications in drug and chip design. The achievement matters because contest problems require creative, correct, executable code, not just plausible answers, so success indicates progress toward reliable AI coding and reasoning. Key technical points: the model is a general-purpose LLM (Gemini 2.5) that was specifically fine-tuned on hard coding, math and reasoning problems; compute used was not disclosed but Google confirmed it exceeded what typical $250/month users of its Gemini Ultra tier receive. Experts caution against overhyping the result—arguing past game wins had limited real-world impact and noting the importance of efficiency, reproducibility and resource costs—but agree the demonstration is a meaningful step toward higher-accuracy, real-world problem-solving AI.