Demis Hassabis on Gemini 2.5, Genie 3, and the Next Steps Toward AGI

In the latest episode of the Google AI: Release Notes podcast, host Logan Kilpatrick sits down with Demis Hassabis, CEO of Google DeepMind, to explore how cutting-edge advancements in AI are bringing the field closer to artificial general intelligence (AGI). The conversation dives into Deep Think in Gemini 2.5, the transformative potential of world model capabilities in Genie 3, and the launch of the new Game Arena on Kaggle as a rigorous benchmark for evaluating AI progress. Hassabis explains that Deep Think in Gemini 2.5 represents a major leap in reasoning and planning, enabling the model to break down complex problems into steps, simulate outcomes, and refine its approach—much like a human would when solving difficult tasks. This capability is a key milestone in moving beyond pattern recognition toward true problem-solving intelligence. A central theme of the discussion is the development of world models—AI systems that learn to simulate and understand the physical and causal dynamics of the real world. Hassabis highlights Genie 3 as a breakthrough in this area, showcasing how the model can predict how objects interact, anticipate changes in environments, and reason about unseen scenarios. These world model capabilities allow AI to go beyond static data and instead build a dynamic, internal understanding of reality—essential for robots, autonomous systems, and general-purpose AI. The launch of the Game Arena on Kaggle is presented as a new benchmark for testing AI agents in complex, interactive environments. Unlike traditional metrics that focus on accuracy or speed, the Game Arena evaluates AI on its ability to adapt, strategize, and learn from experience in real-time, competitive settings. Hassabis sees this as a crucial step in pushing the industry toward AGI by creating shared, challenging environments where progress can be measured and compared. Throughout the conversation, Hassabis stresses that while AI has made remarkable strides, the path to AGI remains long and complex. He emphasizes the importance of building systems that can learn from fewer examples, generalize across domains, and understand cause and effect—capabilities that are now being tested and refined through tools like Genie 3 and platforms like the Game Arena. Listeners can access the full discussion on Apple Podcasts, Spotify, or by watching the video version below, offering a rare behind-the-scenes look at the vision driving one of the world’s most ambitious AI research organizations.