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

The session, billed as a conversation with Google DeepMind's CEO, opens with a short introductory reel. It frames AlphaGo as "the big watershed moment" for DeepMind and for AI in general, and it presents using AI to accelerate scientific discovery as a lifelong aim. The reel also notes a Nobel Prize and a knighthood.

The first question is about the Nobel Prize for the AlphaFold breakthrough. The account of that day stresses how strange it felt. Laureates are told only about 10 minutes before the announcement goes live, and the call from Sweden leaves them "sort of shell-shocked." Next came a week of ceremonies with the Swedish royal family. The account singles out one moment: signing the Nobel book, brought out from a safe in the vaults, after leafing back past the pages of Feynman, Marie Curie, Einstein and Niels Bohr.

Asked whether the prize was expected, the answer stresses uncertainty. The process is described as "amazingly locked down" and treated almost as a Swedish national treasure. The committee looks for real-world impact as well as the scientific breakthrough itself, and that impact "can take 20, 30 years to arrive," so it was impossible to know when, or whether, the recognition would come.

The conversation then turns to where DeepMind sits inside Alphabet's many business units. The answer recounts a merger a couple of years earlier that brought Google's and Alphabet's separate AI efforts, DeepMind included, into one division, and describes that division in these words:

we're the engine room of the whole of Google and the whole of Alphabet.

The division's output includes Gemini, described as the main model, along with video models and interactive world models. According to the discussion, these are now built into "pretty much every product, every surface area," and billions of people use Gemini models through AI Overviews, AI Mode or the Gemini app. Integration into Workspace and Gmail is called "just the beginning." The pitch is that cutting-edge research can go straight to billions of users.

Google DeepMind has about 5,000 people. More than 80% are described as engineers and PhD researchers, which the speaker rounds to "three or four thousand."

Next comes a demo of the newly released Genie 3 world model. The demo video's narration argues that the clips are "not games or videos. They're worlds": interactive environments generated from a single text prompt that users can explore with the arrow keys and space bar.

The commentary stresses that every pixel is generated on the fly. A part of the world does not exist until the user turns toward it. In one clip, someone paints marks on a wall, looks away, and looks back to find the same marks still there. That is offered as evidence that the generated world stays consistent. Users can also type in new elements, such as "a person in a chicken suit or a jet ski," and they appear in the scene in real time.