Google DeepMind CEO Demis Hassabis on AI, Creativity, and a Golden Age of Science | All-In Summit

When a game designer learned to read the fabric of nature

Demis Hassabis's career stitches together unlikely threads: at once a teenage game designer, a neuroscientist fascinated by memory, and the CEO of a research engine now shaping products that billions of people touch daily. The throughline is curiosity about complex systems—whether the rules of a board game, the folding of a protein, or the choreography of a robot hand—and a stubborn belief that machine intelligence can accelerate human discovery rather than simply automate tasks.

Worlds built from words: the rise of interactive world models

What looks like science fiction on a demo reel is, in Hassabis’s telling, an emergent capability of models trained on massive video data. Genie 3, a new class of world model, generates interactive environments from a single text prompt and then responds to real-time control. The visuals are not 3D assets pre-authored by artists; they are pixels synthesized on demand, stitched together by a model that has internalized intuitive physics from millions of hours of footage and synthetic game data. The result reframes how creators and audiences might co-author experiences: worlds where a typed command can instantly add a character in a chicken suit, or a player’s turn of the camera reveals a previously non-existent corner of the environment.

What it means for entertainment and creativity

Rather than replacing artistic vision, these tools democratize the creative pipeline and turbocharge professionals. Directors and game designers can iterate narrative beats and visual effects orders of magnitude faster, while everyday creators can generate polished imagery and scenes without mastering complex toolchains. The future suggested is hybrid: blockbuster visionaries will still craft the major arcs, but millions will co-create and personalize experiences within a shared architecture of dynamic, text-driven worlds.

From video prediction to physical action: multimodal models and robotics

Hassabis views multimodality—systems that accept images, audio, video and text—as a prerequisite for an intelligence that understands the world the way humans do. That understanding is essential if models are to control robotic effectors, whether a pair of tabletop robot hands sorting objects or hypothetical general-purpose humanoids navigating human-built spaces. The company’s fine-tuned Gemini robotics models translate natural language instructions into motor commands, bridging perception and action.

Form follows function, but context favors humanoids

There are two pragmatic paths for robotics: form-specific solutions that excel at narrow, repetitive tasks, and humanoid robots designed to function in environments built for humans. Hassabis argues both will coexist: factories and labs will keep specialized manipulators, while broadly useful assistants might be humanoid simply because our world already has stairs, doorways, and shelves designed for human bodies.

The problem of creativity, and a benchmark for AGI

Hassabis returns repeatedly to one distinction between powerful present-day models and a true artificial general intelligence: creative intuition. He suggests a provocative test—give a system only the knowledge available before 1905 and see whether it can invent special relativity—to separate incremental problem solvers from agents capable of conceptual leaps. Current systems can optimize and innovate within boundaries, but analogical leaps, sudden reframings and the invention of new theoretical constructs remain elusive.

Hybrid engineering: when physics joins probabilistic learning

Progress in biology and chemistry has shown the limits of purely data-driven learning. AlphaFold succeeded because the architecture blended learned representations with hard constraints from physics and chemistry—bond angles and steric collisions are encoded so the model does not waste capacity learning impossible configurations. These hybrid systems are practical bridges: they combine probabilistic neural networks with deterministic constraints and planning layers, but the long-term aim remains to fold those handcrafted insights back into end-to-end learning where possible.

From structure to medicine: Isomorphic and computational drug design

AlphaFold’s structural predictions were a clearing of one major bottleneck. The next bottleneck is design: creating small molecules with therapeutic activity and acceptable safety profiles. Isomorphic, a DeepMind spin-out, applies the same principles to generate candidate compounds, aiming to collapse multi-year discovery timelines into months or weeks by predicting how molecules bind and behave. Early collaborations with pharma and planned preclinical programs point toward a near-term impact on oncology and immunology research.

Efficiency, scale, and the energy story

While model training draws headlines for power consumption, Hassabis emphasizes two opposing forces: continual efficiency gains in serving models and ever-larger experiments at the research frontier. Techniques such as distillation and optimized architectures have produced orders-of-magnitude improvements in cost and latency for deployment, even as research labs push scale to chase new capabilities. He believes AI will ultimately deliver returns for energy and climate systems—optimizing grids, designing materials, and accelerating innovation that offsets consumption.

A cautious forecast and a hopeful finale

On timelines, Hassabis is measured: the path to a truly general intelligence requires breakthroughs in creative reasoning, continual learning and consistent, robust performance across disciplines. He tentatively places those breakthroughs in the five- to ten-year window, not as a certainty but as a plausible trajectory if one or two key innovations emerge. Whether we call the coming decade a renaissance, a golden era of science, or simply rapid iterated progress, the underlying promise is clear: a set of tools that can amplify human ingenuity, help design novel medicines and energy solutions, and reimagine the stories we tell in interactive, personalized worlds. The most consequential question is not whether these technologies arrive, but how their arrival reshapes the practice of discovery and the architecture of everyday life.

key points