Isn't that what we already knew? Language models "predict", and prediction is compression.