AGI when? Google DeepMind CEO says AI still makes simple mistakes despite big wins in elite math

India Today

a day ago

Big Tech giants like Meta, OpenAI, and Google are racing to build artificial general intelligence (AGI). It's the AI systems capable of thinking, planning, and adapting like humans. These companies are pouring billions into research and aggressively recruiting, even poaching, top talent to assemble the best teams. But Google DeepMind CEO Demis Hassabis believes true AGI is still years away, as the AI industry remains far from perfecting current models, let alone achieving human-level intelligence.advertisementSpeaking on the latest episode of the Google for Developers podcast, Hassabis said even the most advanced AI models today display 'jagged intelligence'. Meaning that, while they excel in some areas, they still stumble on basic tasks. He cited examples from Google's Gemini models, highlighting that Google's latest and most powerful Gemini AI model, which incorporates the company's DeepThink reasoning technique, can achieve gold-medal-level performance at the International Mathematical Olympiad, one of the toughest competitions in the world. Yet, he noted, those same models can still make avoidable errors in high school-level mathematics or fail at simple games.'It shouldn't be that easy for the average person to just find a trivial flaw in the system,' Hassabis said.
Hassabis argued that bridging the gap to AGI will require more than simply scaling up models with additional data and computing power. In his view, companies will need to focus on fundamental capabilities, particularly reasoning, planning, and memory, which remain underdeveloped in current AI systems.Another missing piece, he added, is the lack of robust testing. While many standard benchmarks are already saturated, giving the impression of near-perfect performance, he suggested they often fail to expose weaknesses. For example, Hassabis noted that Gemini models recently scored 99.2 per cent on the AIME mathematics benchmark, leaving little room for measurable improvement, even though the model still has flaws.To address this, Hassabis said companies need 'new, harder benchmarks' — not only in academic problem-solving, but also in areas such as intuitive physics, real-world reasoning, and 'physical intelligence'. He also emphasised the importance of safety benchmarks to detect traits such as deception. 'We're in need of new, harder benchmarks, but also broader ones, in my opinion — understanding world physics and intuitive physics and other things that we take for granted as humans,' he said.Notably, Hassabis has previously predicted that AGI could arrive within five to ten years, but cautioned that current systems, from Gemini to OpenAI's latest GPT-5, still lack critical capabilities. He stressed that the focus of AI companies should first be on perfecting today's AI models before pursuing full AGI.- Ends