Raia Hadsell, VP of Research at Google DeepMind, presents three non-language-model AI research areas. First, Gemini Embeddings 2, a fully omnimodal embedding model supporting text, video, audio, and PDFs in a unified semantic space using Matryoshka Representation Learning. Second, weather forecasting models GraphCast, GenCast, and FGN — a series of graph neural network models that outperform physics-based forecasts, with FGN directly predicting cyclone trajectories and already in use by the US National Hurricane Center. Third, Genie 3, an interactive world model that generates real-time, photorealistic, memory-consistent 3D environments from text or video prompts, available to Gemini Ultra subscribers.

24m watch time

Sort: