Jürgen Schmidhuber - World Models, RL, and the Year that changed AI (Part 1)

In the first of two conversations, Jürgen Schmidhuber revisits his laboratory's work from the early 1990s on ideas behind current systems, including LSTM, world models, artificial curiosity and Transformer variants. His central distinction is prediction versus decision-making: language models are strong prediction machines that imitate the web, but acting in the world needs a controller that plans with a world model. He describes his 1990 work in which a controller is rewarded for experiments that improve its own model, why planning must use sub-goals rather than millisecond steps, compression as the core of understanding, and credit assignment.