Actually, yes. There's a deep connection between the compressibility and predictability of things. But we're not good at learning the complex distributions need

Actually, yes. There's a deep connection between the compressibility and predictability of things. But we're not good at learning the complex distributions needed for these sorts of things - there's a long standing competition for natural language compression, the idea being if you can do that perfectly you basically have human level semantic understanding.