Demis Hassabis on World Models, Scientific Discovery and the Path to AGI

In a December 2025 Google DeepMind podcast, Hannah Fry asks CEO Demis Hassabis about AI progress, scientific simulation, safety and the path toward general intelligence. His forecasts are presented as views from that time.

Hassabis highlights advances in model reasoning and world models during 2025. Fry asks how those developments relate to the scientific problems he has long wanted AI to help solve.

He calls protein structure a “root node” problem: solving an underlying scientific challenge can open many other lines of work. He names materials and fusion as other ambitious areas, not as problems already solved.

Hassabis discusses a deeper collaboration with Commonwealth Fusion and imagines how better tools could support energy research. He describes reliable clean energy as potentially consequential for climate, industry and other fields.

Fry asks about “jagged intelligence”: a system may perform impressively on difficult mathematics yet struggle with tasks that seem simple. Hassabis says capabilities remain uneven across domains.

He suggests possible explanations including how information is represented and the limits of current reasoning. These are research hypotheses, not a complete account of why every failure occurs.

Fry asks whether an AlphaGo-like approach can help mathematics. Hassabis says a model can guide search or planning, but real scientific problems are harder to verify and structure than a board game.

The conversation turns to the tension between commercial competition and research. Hassabis recalls an earlier vision of incremental, carefully studied progress, with scientific applications released along the way.

He also welcomes useful assistants and broader access to AI, but says rapid product cycles make rigorous scientific work harder. Commercial resources and competition have accelerated progress while changing the environment in which research happens.

Asked about scaling, Hassabis rejects a simple binary between a hard wall and unlimited smooth improvement. Data, computation and model design remain research questions, and progress can slow or change form.

He says his organization divides attention between scaling and new ideas, describing the balance approximately as half and half. This is his characterization of the team’s effort in 2025.

Fry raises hallucinations and whether models can say “I don’t know.” Hassabis argues that systems need both a representation of uncertainty and training that lets them express it appropriately.

He also suggests that deliberate checking and planning can help a system revisit an answer before presenting it. That remains a design direction rather than a guarantee that hallucinations will disappear.

Hassabis explains his interest in simulation: language contains substantial information about the world, but physical and spatial dynamics are not always fully described in text. Interaction with environments can provide another source of learning.

He sees video and world models as possible ways to capture aspects of those dynamics, with applications ranging from robotics to games. Generating plausible scenes is useful, but visual plausibility alone does not prove physical accuracy.

For science, a learned model could approximate a process from data and make some simulations cheaper than brute-force computation. Whether it is accurate enough must be tested for the specific scientific question.

He describes combining Genie-generated environments with a SIMA agent that acts within them. A system could generate varied tasks and environments, then test whether an agent can carry them out.

Fry asks how to know whether generated worlds are realistic. Hassabis acknowledges that they can hallucinate or approximate mechanics and says grounding and more rigorous tests remain necessary.

He distinguishes an environment that looks convincing to an observer from one that would pass a physics experiment. The latter is a harder standard for using simulations in science.

The pair speculate about using simulations to study evolution and complex systems. Hassabis says controllable virtual environments could let researchers vary conditions and observe results, provided the simulation captures the relevant mechanisms.

Simulations also provide records that can be monitored and analyzed. He suggests using other AI systems to identify interesting or concerning patterns in large volumes of simulation data.

On the possibility of an AI investment bubble, Hassabis says the answer need not be binary. Expectations, valuations and individual projects may overshoot even if the underlying technology proves valuable.

He argues that a large company with existing products can apply AI across its ecosystem through different economic conditions. This is also an account of Google’s own competitive position.

Fry asks how to avoid personalized chatbots becoming echo chambers. Hassabis agrees that excessive agreement can reinforce harmful beliefs and says assistants need ways to challenge users appropriately.

He discusses measuring aspects of an assistant’s persona, such as authenticity and humor, while recognizing that users prefer different styles. The difficult design problem is personalization without surrendering reliability.

Asked about AGI, Hassabis points to advances in imaging, world models and reasoning as components that might eventually be combined. He uses terms such as “proto-AGI” prospectively rather than claiming the question is settled.

The hosts compare possible disruption from advanced AI with the Industrial Revolution. Hassabis predicts changes could unfold faster and affect more areas; the magnitude and timeline are his estimates, not established facts.

He speculates about new institutions and forms of collective decision-making, as well as questions of work and purpose in a more automated economy. Fry presses on whether governance can keep pace with technical development.

Hassabis says he worries that institutions are fragmented and may lack sufficient influence. He calls for responsible development and eventual international cooperation on basic standards.

The discussion turns to computability and whether intelligence or consciousness has limits that classical computation cannot capture. Hassabis describes his working view while acknowledging that these questions are unresolved.

Fry asks what it feels like to lead at the frontier. Hassabis describes both excitement about scientific tools and a sense of responsibility for the risks and social changes they could bring.

He reflects on creativity: AI can accelerate prototyping, but people may worry about how their own skills and roles change. He sees toolmaking and human expression as central to how society adapts.

Hassabis distinguishes current, largely human-directed systems from more autonomous systems he expects to emerge. He says increasing autonomy raises concerns that need attention before deployment.

He closes by describing his goal as helping society steward advanced AI safely and collaboratively. The conversation ends with Fry’s reminder that the technical and institutional questions remain open.