How Google DeepMind is researching the next Frontier of AI for Gemini — Raia Hadsell, VP of Research

Raia Hadsell, VP of Research at Google DeepMind and UK AI Ambassador, delivers the opening keynote at AI Engineer Europe 2026, presenting three frontier research directions that move explicitly beyond standard large language models: advanced embedding models, AI-driven weather forecasting, and generative world models.

Hadsell opens with a reflection on her career arc — undergraduate philosophy of religion, a PhD under Yann LeCun at NYU on convolutional networks and Siamese networks for robots, joining DeepMind when it had 30-40 people, and now leading roughly 1,200 scientists and engineers across 10 labs. She frames DeepMind's research philosophy around finding 'root nodes' — deep, unsolved problems whose solutions unlock broad downstream impact — rather than incremental leaf-level work.

On embedding models, Hadsell introduces the concept of 'Jennifer Aniston cells' from neuroscience — a small cluster of neurons that fire for a specific person or concept regardless of modality (hearing the name, seeing a photo, recognising a voice). She argues embedding models in AI should function analogously: fast retrieval, recognition, and comparison across modalities using contrastive loss functions, an area she traces back to her own PhD work on Siamese networks. The team recently released Gemini Embeddings 2, described as fully omnimodal — a single unified vector space encoding up to 8,000 tokens of text, 128 seconds of video, 80 seconds of audio, or a full PDF. It uses Matryoshka Representation Learning (MRL), enabling retrieval at 256 dimensions for speed and expanding to higher dimensions for expressiveness. Hadsell positions it as a critical companion to generative AI, enabling true end-to-end retrieval without lossy modality-bridging steps.

On weather forecasting, Hadsell recounts how a Met Office data scientist's question — 'can you beat our physics-based models?' — led to a multi-year research programme. With 40 years of global atmospheric data available, the team found deep learning tractable. GraphCast, a spherical graph neural network with nodes spanning from Earth's surface to the lower stratosphere, autoregressively predicts 100 atmospheric variables 15 days out globally. In a concrete test case, GraphCast predicted Hurricane Lee's Nova Scotia landfall accurately 9 days out, versus 6 days for the best physics-based models — a 3-day advantage that is operationally significant for evacuation planning. The next model, GenCast, is probabilistic (capturing chaotic tail events), more accurate than 97% of 1,300 gold-standard benchmark forecasts, and produces a 15-day forecast in 8 minutes on a single chip rather than hours on a supercomputer. The latest model, FGN (Functional Generative Network), directly predicts cyclones — including categorisation, trajectory, wind speed, and eye formation — rather than layering a cyclone detector on top of a weather model. FGN is already in active use at the US National Hurricane Center.

On world models, Hadsell traces a lineage from DeepMind's games research (Atari, Go, StarCraft, MuJoCo robotics) toward generating infinite training environments. Genie 1 produced simple 2D platformer worlds responding to left/right inputs. Genie 2 scaled to interactive 3D environments but lacked real-time performance and photorealism. Genie 3 (built on or augmented by the V03 model) produces high-definition, photorealistic, real-time 3D worlds from text or video prompts — demonstrating physics-aware walking on a muddy Kent lane, skiing, and consistent memory (returning to the start of a world finds it unchanged). A notable capability is dynamic in-world prompting: the user can issue text prompts while inside an active world and the environment transforms immediately. Hadsell suggests applications in adversarial gaming experiences and immersive education.