Google DeepMind Warns: AGI is Just Years Away, Let's Prepare

General Artificial Intelligence (AGI)

General Artificial Intelligence (AGI), the scientific goal for decades, is no longer a distant milestone. For Demis Hassabis, CEO of Google DeepMind, its arrival is a matter of just a few years, no more than a decade. Speaking at an event at the Stanford Graduate School of Business, the AI pioneer made a clear prediction: AGI, the point where AI equals or surpasses human intellectual capabilities across a wide range of tasks, will arrive by 2030, "give or take a year." A staggering timeline that Hassabis frames as the start of a new human era.

Hassabis is blunt: "When we look back at this period, I think that in perhaps 10 years we will realize we were at the foot of the singularity." The year 2026, according to the DeepMind CEO, marks a crucial turning point: AI agents and tool-using capabilities have reached tangible utility in daily work, giving developers a clearer view of the remaining steps towards AGI. However, his greatest concern is preparation.

"Society needs to listen to this because we don’t have much time to prepare for what it means. It will be profoundly deep," he warns. The future, he says, is yet to be written, but the next few years will be "very critical" in determining its direction.

AGI is Close: DeepMind CEO Sounds the Alarm

Hassabis' observations are not new in the industry. Last year, Sam Altman, CEO of OpenAI, claimed that his company knew how to build AGI "as we have traditionally understood it," forecasting the entry of AI agents into the workforce. Dario Amodei from Anthropic and Elon Musk from SpaceX and xAI have also predicted AGI-level systems within a few years. Musk, in particular, goes further: "I think we will reach AGI by 2026," and confidently states, "by 2030, AI will surpass the intelligence of all humans combined." Some, like Shaw Walters from Eliza Labs, even believe that the milestone has already been reached and that current models fall within the definition of AGI, considering them "general intelligence."

Not everyone shares this enthusiasm, however. Skeptics highlight how current systems are still far from human general reasoning. In March, the ARC Prize Foundation released the benchmark ARC-AGI-3, designed to test AI systems' ability to learn and adapt in unknown environments. The leading models from Google, OpenAI, Anthropic, and xAI scored below 1%, while human participants achieved 100%. This gap reveals significant limitations.

Further complicating the picture is the lack of a consistent definition of AGI. Malo Bourgon, CEO of the Machine Intelligence Research Institute, emphasizes how conflicting definitions make it difficult to establish when the goal is actually achieved. "There are a lot of different definitions," said Bourgon, making it hard to determine what exactly qualifies as AGI.

Despite the uncertainties and definitional challenges, Hassabis remains firm in his belief regarding the acceleration of technological progress. "Everything will change in the next 10 years, probably more than people expect," he concludes, reiterating the urgency of social preparation that can no longer be left to technologists alone. The game for the future, and for harmonious coexistence with AGI, is being played now.