the idea that AI and data compression are isomorphic is really interesting and makes me think of Ashby's law of requisite variety, where the number of possible states (this could also be rendered in information-theoretic terms of entropy) of an agent operating in an environment must be at least as high as the number of states of the environment itself. Thus, to navigate a given environment is to be able to represent that environment internally. Reducing the byte-size of that representation then means that you are representing the environment more tersely, and in my experience (there's probably a theoretical explanation for this but I've only seen it experientially) terser-but-just-as-accurate representations are usually better models.