In a wide-ranging interview released this week, Google DeepMind CEO Demis Hassabis laid out a vision for the future of artificial intelligence that extends far beyond chatbots, predicting that the convergence of “world models” and scientific discovery will define the next phase of the technology.
Speaking with mathematician Professor Hannah Fry on the season finale of Google DeepMind: The Podcast, Hassabis offered a candid assessment of the current state of AI, describing today’s systems as possessing “jagged intelligence”—brilliant in specific domains yet surprisingly flawed in basic reasoning. He also addressed the industry’s overheating economics, warning that while the technology is “overhyped in the short term,” its eventual impact will dwarf the Industrial Revolution.
The “Jagged Intelligence” Problem
Despite a year that saw the release of Gemini 3 and significant advances in multimodal capabilities, Hassabis acknowledged that the path to Artificial General Intelligence (AGI) remains obstructed by fundamental reliability issues. He noted the paradox of models that can win medals at the International Mathematical Olympiad yet stumble over simple logic puzzles.
“It’s very uneven still, the performances of these systems,” Hassabis said. “They’re very, very impressive in certain dimensions… but they’re still pretty basic in others.”
He attributed these inconsistencies—often manifested as “hallucinations”—to a lack of grounding in physical reality and the absence of a rigorous internal monologue. Current models, he explained, are often “literally telling you the first thing that comes to their mind.” To achieve AGI, systems must develop the capacity for “reasoning and thinking,” allowing them to double-check outputs and plan ahead.
“My betting is you’re going to need both to get to AGI,” Hassabis said, estimating that the solution lies in a split approach: “50% of our effort is on scaling, 50% of it is on innovation.”
Solving “Root Node” Problems
DeepMind’s strategy continues to focus on what Hassabis calls “root node” problems—scientific challenges that, if solved, unlock massive downstream benefits for humanity. Following the success of AlphaFold in biology, the lab is now turning its attention to material science and clean energy.
Hassabis highlighted a deepening partnership with Commonwealth Fusion Systems to apply AI to nuclear fusion, specifically to help contain plasma within tokamak reactors.
“If energy really was renewable and clean and… super cheap or almost free, then many other things would become viable,” Hassabis told Fry, citing desalination and continuous production as examples. “That’s the holy grail, and of course, that’s one of the ways we could help with climate.”
The Rise of World Models
A central theme of the discussion was the necessity of “world models”—AI systems that understand the physics, causality, and spatial dynamics of the physical world, rather than just statistical patterns in language.
Hassabis pointed to internal projects like “Genie” and “SIMA” (Simulated Agents) as stepping stones. By training agents in simulated environments, DeepMind aims to create AI that can understand cause and effect. “If you can simulate it, then in some sense you’ve understood it,” Hassabis noted.
This capability is viewed as essential for the next generation of robotics and “universal assistants.” Hassabis envisions a future where AI agents are not merely passive chatbots but active participants capable of executing complex tasks.
The AI Bubble and the Industrial Revolution
Addressing the frenetic pace of investment in the AI sector, Hassabis admitted to concerns regarding a financial bubble, particularly surrounding early-stage startups raising billions with unproven products. However, he maintained that the long-term transformative potential of the technology is not being exaggerated.
“It’s probably going to be 10 times bigger than the Industrial Revolution, and it’ll probably happen 10 times faster,” Hassabis predicted.
He warned that existing societal institutions might be ill-equipped to handle such rapid disruption, advocating for international cooperation and potential economic restructuring to ensure the benefits of AGI are widely distributed.
The Ultimate Limit
Ultimately, Hassabis remains driven by the philosophical question of whether the human mind is fully computable. While acknowledging theories regarding quantum effects in the brain, he operates under the hypothesis that intelligence is a process of information processing that can be replicated.
“Nobody’s found anything in the universe that’s non-computable so far,” Hassabis said.
He suggested that building AGI might be the only way to truly understand the mysteries of the human mind. “I’ve always felt this,” he concluded. “If we build AGI… and then use that as a simulation of the mind, and then compare that to the real mind, we will then see what the differences are.”