World Models in AI

This talk explores how intelligent systems can move beyond simply reacting to inputs and instead build internal simulations of the world — enabling them to predict, plan, and reason like humans do. The talk traces the history of this idea from brittle symbolic AI to today's deep learning revolution, examines how world models are built and brings the concept to life through real-world applications in game AI, robotics, autonomous vehicles, and large language models. It then tackles the deeper question of why world models may be the central missing ingredient on the path to AGI — unlocking counterfactual reasoning, causal understanding, and genuine common sense — before confronting the key unsolved challenges: the simulator gap, uncertainty at scale, and the alignment risks that come with agents who act confidently on potentially flawed internal beliefs.