Love this. Very interesting that same amount of compression (samples) can give ever more accuracy if you do a bit more work in the decompression — by taking higher order fits to more of the sample points.
One characteristics of ML is that this distinction often is not clear. (Hallucination, generalization, etc.)