Jürgen Schmidhuber: World Models, RL and Year That Changed AI (www.the-information-bottleneck.com)

In a recent episode of "The Information Bottleneck," AI pioneer Jürgen Schmidhuber discusses his foundational contributions to modern artificial intelligence, including world models and reinforcement learning (RL). Reflecting on his lab's groundbreaking work in the early '90s, Schmidhuber emphasizes the significance of creating algorithms when computational resources were vastly limited. He argues that while large language models (LLMs) excel at prediction by synthesizing web data, they lack the decision-making capabilities essential for real-world interactions, which require a "controller" that employs a world model for planning. Schmidhuber highlights the distinction between prediction and decision-making as crucial for advancing AI beyond mere imitation of human data. He critiques the current portrayal of LLMs as comprehensive AI solutions, reinforcing the need for development in RL that incorporates planning through world models and goal-setting. With a prediction model in place, agents can simulate future actions and learn effectively, ultimately bridging the gap toward advanced robotic capabilities and possibly AGI. The conversation serves as a reminder of the foundational insights from decades past that are vital in addressing today's AI challenges and shaping future innovations.