Inside Google DeepMind: AGI, Robotics, & World Models Explained - Demis Hassabis

Google DeepMind is now Alphabet’s centralized AI engine, not a standalone research lab. Hassabis says the merged division has about 5,000 people, over 80% engineers and PhD researchers, and its models run across nearly every Google surface, reaching billions of users.

Genie 3 points beyond video generation: it creates interactive worlds pixel-by-pixel as users move through them, without a conventional 3D rendering engine. Trained on video plus synthetic game data, it can maintain world consistency for a minute or two, making it a test bed for physical-world understanding.

DeepMind is pursuing an “Android for robotics” strategy alongside vertically integrated robot systems. Hassabis expects a major robotics “wow moment” within a couple of years and eventually millions of robots, but says scaling hardware too early risks locking factories into designs that may be obsolete six months later.

Hassabis rejects claims that today’s models are already “PhD intelligences” and puts true AGI roughly five to ten years away. He says current systems lack consistent general reasoning, continual learning and human-like creative leaps, and probably still need one or two fundamental breakthroughs rather than scaling alone.

Isomorphic Labs is trying to compress drug discovery from years—or a decade—to weeks or days within the next ten years. Hassabis says it has partnerships with Eli Lilly and Novartis, works with MD Anderson, and expects its own drug programs to enter preclinical development next year.

For scientific AI, DeepMind expects hybrid models to remain necessary for roughly the next five years. In biology and chemistry, sparse data means learned systems still need scientific constraints; AlphaFold and AlphaGo combined neural learning with engineered rules or search before later systems absorbed more of that structure end-to-end.

AI efficiency is improving dramatically without reducing total power demand. Hassabis says models have become roughly 10× to 100× more efficient at the same performance over two years, but frontier training keeps expanding, so cheaper inference and rising total compute demand are happening simultaneously.

DeepMind is already co-designing creative AI with professional filmmakers, including Darren Aronofsky, rather than treating creators only as end users. Hassabis says elite creatives can become 10× to 100× more productive with these tools, while skill, taste and storytelling still separate top professionals from casual users.