Is Information Digital? A Defense of Reality

Data-driven techniques, such as large-language models, have proven astonishingly powerful in recent years, but progress has been much slower with cyber-physical systems such as robots. Many people assume this is because there is not enough training data. In this talk, I explore a possible explanation that is much more fundamental. Specifically, I ask the question of whether there is a fundamental difference between acquisition of knowledge through observation and acquisition of knowledge through embodied interaction. Can you learn to ride a bicycle by watching others ride a bicycle? In previous work, I have used concepts from computer science (zero-knowledge proofs, bisimulation, etc.) to show that there are things you can learn from embodied interaction that cannot be learned by objective observation. In this talk, I use Shannon information theory to argue that objective observation falls far short of revealing everything about physical reality. There is information in the real world that cannot be represented digitally, and objective observation can never acquire more than a small subset of this information. In short, learning to ride a bicycle may require getting on a bicycle, even for a robot.