Think of the smartest autocomplete tool you’ve ever used. Now ask it to predict what happens when you push a glass off a table. It can describe the shattering in poetic detail. It has absolutely no idea why the glass falls.

That’s essentially the argument Demis Hassabis, CEO of Google DeepMind, laid out in a January 2026 interview on CNBC’s “The Tech Download.” Large language models, for all their impressive capabilities, fundamentally lack an understanding of physics, causality, spatial dynamics, and long-term planning. Language, Hassabis contends, describes the world but does not fully contain it.

The fix, according to Hassabis, is something called “world models”: AI systems designed to simulate and predict real-world dynamics rather than just process and generate text. It’s a distinction that sounds academic until you realize it’s reshaping DeepMind’s entire approach to building artificial general intelligence.

Language is a map, not the territory

Here’s the thing about LLMs like Google’s Gemini. They can process text, images, audio, and video. They can pass bar exams and write functional code. But ask them to reason about what happens when two objects collide, or to plan a sequence of physical actions over a long time horizon, and the cracks start showing.

Hassabis’s point is that language is a compression of reality, not reality itself. A sentence like “the ball rolls down the hill” encodes a tiny fraction of the physical information involved in that event. Gravity, friction, momentum, the shape of the terrain: none of that is actually “in” the words. LLMs learn statistical patterns across language. They don’t learn the laws of physics that language attempts to describe.

From Genie 3 to AGI: DeepMind’s world model bet

Hassabis’s remarks didn’t arrive in a vacuum. Around the same time, DeepMind showcased Genie 3, a project that generates interactive environments based on natural language prompts. Rather than producing a static image or a block of text, Genie 3 creates simulated spaces that a user, or another AI, can actually navigate and interact with.

Hassabis has linked world models directly to DeepMind’s long-term AGI strategy. He views them as essential infrastructure for two domains in particular: robotics and scientific discovery. A robot that can only process language instructions but can’t simulate the physical consequences of its actions is, to put it mildly, limited. A scientific AI that can’t model causal relationships isn’t doing science. It’s doing pattern matching.

The timeline Hassabis has floated for AGI sits at roughly 5 to 10 years out. That estimate has been consistent in his public statements, and the world model pivot suggests DeepMind views the path to AGI as running through physical simulation rather than ever-larger language models.

What this means for investors and the AI landscape

DeepMind’s own trajectory illustrates the shift. The lab built its early reputation on game-playing AI systems like AlphaGo and AlphaFold, both of which required the kind of structured, rule-based reasoning that LLMs struggle with. Moving toward world models is, in many ways, a return to DeepMind’s roots: AI that understands constraints, consequences, and causality.

Tech media outlets reported Hassabis’s remarks widely on January 19–20, 2026, with no mention of crypto or digital assets. The risk for investors is timing. World models are still early-stage relative to LLMs. Genie 3 is impressive as a research demonstration, but the gap between a research demo and a commercially deployed system that outperforms existing approaches is historically where a lot of investment capital goes to die. Hassabis’s 5 to 10-year AGI estimate implicitly acknowledges this: the technology he’s describing doesn’t exist yet in mature form.