Frontier AI and the Future of Intelligence — Raia Hadsell, VP of Research, Google DeepMind
TL;DR
Raia Hadsell frames frontier AI as bigger than chatbots — at Google DeepMind she says the real job is finding “root nodes,” not chasing leaves, meaning foundational problems that reshape artificial, human, and robotic intelligence together.
Gemini Embeddings 2 is her sleeper pick — she highlights DeepMind’s new fully omnimodal embedding model as a critical companion to generative AI, able to encode text up to 8K tokens, 128 seconds of video, 80 seconds of audio, and full PDFs into one semantic space.
She uses the “Jennifer Aniston cell” from neuroscience to explain retrieval — the point is that robust concept representations should fire across modalities, so a model can recognize the same thing from text, image, video, or voice and retrieve it fast.
DeepMind’s weather models beat physics systems where it matters — GraphCast predicted Hurricane Lee’s Nova Scotia landfall 9 days out versus 6 days for top traditional models, and Hadsell stresses that “three days is really important” when a major storm is coming ashore.
GenCast pushed weather forecasting from impressive to operational — she says the probabilistic model beat gold-standard forecasts 97% of the time across 1,300 benchmarks and produced 15-day forecasts in 8 minutes on a single chip instead of hours on a supercomputer.
Genie 3 is her vision of world models becoming a new medium — after showing interactive prompts like “a muddy lane in Kent” and “an origami lizard in an origami world,” she argues real-time, editable simulated worlds could matter not just for gaming but for education.
The Breakdown
From philosopher to DeepMind veteran
Raia Hadsell opens with a quick, funny personal arc: philosophy of religion in the 1990s, convolutional nets for robots in New York with Yann LeCun in the 2000s, then joining a “small group of curious, scrappy individuals” at DeepMind when it was just 30 to 40 people. Now she helps lead roughly 1,200 scientists and engineers across 10 labs, and she uses that arc to set up a broader point: frontier AI is about shaping the future of intelligence, not just shipping the next model.
“Find the root nodes, not the leaves”
Her core research philosophy is unusually crisp: don’t waste time on leaves; go after the deepest unsolved problems that unlock downstream impact. She ties that directly to DeepMind’s mission — building AI responsibly for the benefit of humanity — and says the work has to involve the wider world, from government to academia to industry, not just one lab acting alone.
Why embeddings still matter in the age of generators
Hadsell deliberately spends her first technical section on something that is “not directly language models”: embeddings. She borrows the neuroscience idea of the “Jennifer Aniston cell,” where the brain seems to encode a person or concept robustly across modalities, and says that’s exactly what you want in AI for fast retrieval, recognition, and comparison.
Gemini Embeddings 2 as an omnimodal retrieval layer
She describes Gemini Embeddings 2 as a companion to generative AI rather than a side show: “sometimes we want to generate, sometimes we want to retrieve.” The model can map text, audio, video, and documents into a single vector space — including 8K tokens of text, 128 seconds of video, 80 seconds of audio, and a full PDF — and uses Matryoshka Representation Learning so developers can start with smaller embeddings like 256 dimensions and scale up when they need more expressiveness.
Weather forecasting as an AI-native science win
The next example is even more removed from LLM discourse: weather. Hadsell recounts how a Met Office scientist basically challenged DeepMind to beat physics-based rainfall prediction, and that turned into GraphCast, a spherical graph neural network trained on 40 years of global weather data to forecast up to 15 days across Earth.
From GraphCast to GenCast to cyclone-specific models
She gives the vivid example of Hurricane Lee: GraphCast correctly called its Nova Scotia landfall 9 days ahead, versus 6 days for the best traditional systems. Then she walks through GenCast, a probabilistic successor that beat gold-standard forecasts 97% of the time across 1,300 benchmarks and ran a 15-day forecast in 8 minutes on a single chip, before landing on FGN, which predicts cyclones directly and is already being used by the US National Hurricane Center.
Genie and the jump from generated video to playable worlds
Her final section is about world models, tracing a line from Atari, Go, StarCraft, and robotics simulators to Genie 1’s simple 2D platformers and Genie 2’s slower 3D worlds. The real excitement is Genie 3: a system aiming for high-quality, real-time, interactive environments, though the demo briefly gets derailed by slide and Wi-Fi issues in a very human conference moment.
The demos: Kent mud, origami lizards, and promptable reality
Once the videos work, the room gets the payoff: a first-person muddy lane in Kent where the user has a visible body, skiing scenes, an artist’s hand-built clip turned into a navigable world, and an “origami lizard in an origami world” used to show memory and consistency. Hadsell’s most playful idea comes at the end, when she changes a Camden Canal scene live with prompts and jokes about “adversarially prompting” someone else’s game — a glimpse of world models as a new entertainment and education medium, not just another model benchmark.