big world hypothesis
Richard Sutton's argument that any learning system is always far smaller than the world it operates in, and should be designed accordingly.
The idea is simple and slightly humbling. Whatever you build, the world it has to cope with is vastly larger and more complicated than the system itself. A model can therefore never hold a complete picture of its environment, only a compressed and partial one, and the honest design response is to build systems that keep learning from the world rather than ones that try to finish learning in advance.
Richard Sutton uses it as the basis for a pointed criticism of synthetic data, training material a model generates for itself. If the world is endlessly complex and the model only ever captured a sliver of it, then data the model produces comes from that same sliver. You get more of what it already knows rather than more of what it does not. Others argue synthetic data demonstrably works for teaching specific checkable skills like code and reasoning steps, which is a different claim than whether it can produce anything genuinely new.