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, presents three non-LLM frontiers: omnimodal embeddings, AI weather forecasting, and real-time interactive world models.

• Gemini Embeddings 2 is fully omnimodal: one vector encodes 8K tokens of text, 128s of video, 80s of audio, and a full PDF.
• GenCast beat 1,300 gold-standard weather benchmarks 97% of the time and produces a 15-day forecast in 8 minutes on a single chip vs. hours on a supercomputer.
• GraphCast predicted Hurricane Lee landfall accurately 9 days out; best physics-based models were only accurate to 6 days — a 3-day edge on a major hurricane.
• FGN (functional generative network) predicts cyclone trajectory and eye formation directly, skipping post-processing; already in use at the US National Hurricane Center.
• Genie 3 generates real-time photorealistic 3D interactive worlds with persistent memory — run a minute away, return, and the environment is identical.
• Genie 3 worlds can be re-prompted mid-experience, enabling adversarial or educational real-time environment manipulation.
• Hadsell framed embedding models as an underappreciated companion to generative AI: fast retrieval and cross-modal comparison that LLMs alone cannot provide.

2026-04-18 · Watch on YouTube