Frontier AI and the Future of Intelligence — Raia Hadsell, Google DeepMind

Three things that aren't language models[01:19]

Hadsell opens with a decades-at-a-glance biography — 90s philosophy of religion undergrad, a 2000s pivot into computer science and a PhD on convolutional and Siamese networks with Yann LeCun², then joining DeepMind in the early 2010s when it was "a small group of 30, 40 people" working on Atari, Go and StarCraft.¹¹ She now runs roughly 1,200 researchers and engineers across ten labs and is the UK's AI ambassador.

The organising idea of the talk is deliberately contrarian for a 2026 audience: "the theme of this talk overall is things that are not directly language models." She picks three — embeddings, weather, world models — to show that DeepMind's frontier work still reaches well beyond the chatbot stack.

The Jennifer Aniston cell[05:05]

To motivate embeddings, Hadsell borrows from neuroscience. In the human medial temporal lobe, a small group of neurons will activate for a specific concept — a particular person or place — regardless of modality: a written name, a photograph, a voice clip all light up the same cells.¹ The brain uses this invariance for fast retrieval, recognition and comparison.

Embedding models are the artificial analogue: a network trained with a contrastive loss to place the same concept at the same point in a vector space, no matter the input modality. Hadsell notes this is a direct descendant of the Siamese-network work she did under LeCun in the 2000s² — "sometimes we want to generate, sometimes we want to retrieve," and the retriever half has quietly become as important as the generator.

Gemini Embedding 2 and the Matryoshka trick[07:32]

Google's new Gemini Embedding 2 is the practical payoff.⁴ It is, in her framing, "fully omnimodal": a single vector that represents text up to 8K tokens, 128 seconds of video, 80 seconds of audio, or a full PDF. Because it's derived from Gemini itself, it inherits the base model's world knowledge rather than stapling together modality-specific encoders end-to-end.

The other half of the story is Matryoshka Representation Learning³: the same network produces nested embeddings — start a retrieval with 256 dimensions for cheap approximate recall, then widen out to more dimensions only when you need the precision. One model, many operating points.

Learning the weather from 40 years of data[09:52]

The second "not a language model" story starts with a question from the UK Met Office: can neural networks beat physics simulations at rainfall?¹² The answer, given forty years of global reanalysis data, was yes — and the team has iterated three times since.

GraphCast runs a spherical graph neural network over an icosahedral mesh that wraps the Earth, auto-regressively predicting roughly one hundred atmospheric variables — wind, temperature, humidity — up to fifteen days out.⁵ Hadsell uses Hurricane Lee as the canonical demo: GraphCast nailed the Nova Scotia landfall nine days in advance, versus six days for the gold-standard physics models — three days that matter when you're evacuating a coast.⁸

GenCast went further by making the model probabilistic, which matters because "the weather is fundamentally chaotic" and a point forecast throws away the tails you actually want to plan for. Against 1,320 gold-standard benchmarks GenCast is more accurate 97% of the time, and generates a 15-day forecast in eight minutes on a single chip instead of hours on a supercomputer.⁶

The current frontier is FGN — a Functional Generative Network that skips the "predict weather, then detect cyclones" pipeline and instead learns cyclone category, trajectory, wind speed and eye formation directly end-to-end. It's already running operationally at the US National Hurricane Center.⁷

From Genie 1 to Genie 3[14:33]

The final act is world models, which Hadsell traces back to DeepMind's long line of Atari, Go, StarCraft and MuJoCo work. The shift: stop training just the agent, start generating the environment itself.

Genie 1 produced a few seconds of a 2D platformer that responded to left/right input. Genie 2 lifted that into slow, interactive 3D. And in parallel, Veo 3 showed that photoreal video generation was solved — but non-interactive and not real-time.¹⁰ Genie 3 unifies the two: a playable, real-time, high-quality 3D world you can steer through a text prompt.⁹ Hadsell's live demo didn't cooperate with the conference Wi-Fi, but she described walking a muddy lane in Kent, skiing, and inhabiting an origami lizard world.

Her closing pitch is about what this unlocks. "Adversarially prompting your experience of a world" is a new genre of gaming — inject surprises into a friend's scene in real time — and the same capability reframes education as something you can step inside of rather than read about.