The "thinking model" breakthrough
"Image synthesis assisted by Porcelain, an AI partner within the Global Future Nexus ecosystem."
From AlphaGo's strategic intuition to the DeepThink engine that can plan across parallel reasoning tracks, Demis Hassabis has identified a clear roadmap to AGI. The path is not simply larger models—it is the integration of reasoning, planning, and world models into a unified cognitive architecture that can think before it acts.
The Return of the Agent
The emergence of "thinking models" marks a return to DeepMind's intellectual roots. According to Hassabis, these systems are "a little bit of a hark back to our original gaming work on things like AlphaGo and AlphaZero". Since its founding, DeepMind has focused on "agent-based systems"—systems that can complete a whole task rather than just performing one small piece of a workflow. This is "obviously the way to get to AGI".
In the early days, agent-based meant playing a game very well, with a clear goal. Today, it means building powerful multimodal models that process language, images, and video while adding "thinking" and "planning" capabilities on top. When a model can think, it can extend to "deep thinking" and even "parallel planning"—simultaneously exploring multiple lines of thought and selecting the optimal path forward. This is not simply generating the first answer that comes to mind; it is a system that continuously corrects and optimises its own reasoning process.
The Missing Consistency
Despite their remarkable capabilities, today's systems suffer from what Hassabis calls "patchy intelligence". They can solve International Mathematical Olympiad problems, but fail on high-school mathematics or basic logic when the phrasing changes. They can generate entire virtual worlds consistent with physics, yet violate the rules of chess.
This "jagged" capability profile is "the thing that's missing from full AGI". A system that is brilliant in some dimensions and childlike in others is not general intelligence. It is a patchwork of specialised capabilities—impressive, but incomplete.
The World Model as Cognitive Foundation
The solution, for Hassabis, lies in world models—systems that not only understand language but simulate the physical world. The goal is a model that understands "physical structures, material properties, the flow of liquids, biological and human behaviour". Since AGI must operate in the physical world, it must understand it.
Genie 3 exemplifies this: it can generate interactive, real-time 3D worlds from a simple text prompt, maintaining consistency for several minutes of continuous interaction. The model's emergent memory and physical reasoning are not hard-coded—they emerge from architecture and training scale. A user can paint a wall in a generated world, explore elsewhere, and return to find the paint still there. The model "teaches itself how the world works—how objects move, fall, and interact".
The Convergence: Omni Model
DeepMind's ultimate goal is an Omni Model that fuses language, multimedia understanding, physical reasoning, and generative capabilities—an AI that can perceive, reason about, and act in the world. This convergence is the path from today's patchwork intelligence to the comprehensive, consistent intelligence of AGI.
The core insight is that reasoning, planning, and thinking models are not separate innovations—they are components of a unified cognitive architecture. Hassabis notes that there is "still lots of innovation to be had in the 'thinking' bit," but that the direction is clear. Systems that can think, plan, and learn from experience are the bridge from today's capable but brittle AI to the general intelligence that can navigate the complexity of the physical world.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
