Europe 2026
How Google DeepMind is researching the next Frontier of AI for Gemini — Raia Hadsell, VP of Research
Overview
This talk explores Google DeepMind's research into the next frontiers of AI, focusing on advancements beyond traditional language models. It highlights progress in creating more robust and versatile AI systems, including multimodal embedding models, advanced weather prediction systems, and interactive 3D world models. The core thesis is that significant breakthroughs lie in developing AI that can understand and interact with the world in more comprehensive and nuanced ways, moving towards more general intelligence.
Who should watch
- AI Engineers interested in multimodal AI and world models.
- Product Managers seeking to understand the future capabilities of AI.
- Researchers exploring novel AI architectures and applications.
- Builders looking for inspiration for new AI-powered products and experiences.
- Those interested in AI's potential beyond text generation, including scientific and creative applications.
Key takeaways
- Gemini embeddings 2 offers a fully omnimodal approach, unifying text, audio, video, and PDF data into a single vector representation for enhanced retrieval and agentic logic.
- Matryoshka Representation Learning (MRL) allows for flexible embedding dimensions, enabling efficient retrieval and greater expressiveness within the same network.
- GraphCast and GenCast demonstrate AI's capability to outperform physics-based models in weather prediction, offering faster and more accurate forecasts, even for complex phenomena like hurricanes.
- Functional Generative Networks (FGN) directly predict cyclones, integrating categorization and trajectory prediction for improved accuracy and operational use by meteorological centers.
- Genie, a world model, generates diverse, interactive, and high-quality 3D environments based on prompts, showcasing emergent properties like physics simulation and memory.
- Genie 3 allows for real-time interaction and in-world prompting, enabling dynamic changes to the environment and opening new possibilities for gaming, education, and interactive experiences.
- Research is focused on identifying and solving deep, fundamental problems rather than incremental improvements, aiming for significant downstream impact.
- The mission of DeepMind is to build AI responsibly for the benefit of humanity, emphasizing the importance of solving problems that are worth solving.
Notable quotes
*We want that in an artificial neural network for the same reason. We want fast retrieval recognition and comparison.*
*The motto the mission of DeepMind is to build AI responsibly for the benefit of humanity.*
*Predicting the weather, even though it is a very very hard problem using physics simulation of the atmosphere, is actually quite tractable for neural network models.*
Unofficial community note. Prefer the recording for nuance.