It's not that we can do it for abitary compressive encodings, it's that we can do it for some. A machine trained only on non-compressive encodings, cannot take

It's not that we can do it for abitary compressive encodings, it's that we can do it for some. A machine trained only on non-compressive encodings, cannot take a compressed representation and produce facts about it. A human /can/. (NB. via the fact that intelligence isn't just 'generalized interpolaton' but a contenful process of reasoning, skill, curiosity etc. that comes to bare on generating descriptions in a novel mode).