In this in-depth interview, Demis Hassabis reflects on Google DeepMind's progress, emphasizing the dual focus on scaling and scientific innovation for achieving AGI. He highlights recent advancements like Gemini 3, a multimodal AI model capable of processing various types of data, and the significance of 'world models' in enabling AI to understand spatio-temporal dynamics beyond language. Hassabis addresses the 'jagged intelligence' phenomenon in current AI, where models excel in complex tasks but fail in simpler ones, underscoring the need for greater consistency and reasoning. He discusses the science-commercialization balance, acknowledging the current competitive landscape but reiterating DeepMind's commitment to foundational research and 'root node' problems like AlphaFold, a protein structure prediction system. The conversation also delves into the potential of simulated environments for AI training, the nature of the 'AI bubble', and the critical importance of building ethical AI that avoids social media's pitfalls. Looking ahead to AGI, Hassabis considers its societal impact, the need for new economic systems, and the ongoing quest to understand the limits of computation, particularly related to the Turing machine.