Looking at the `Educational values deemed by the filter` table on page 5, I can't help wondering whether compressibility might not be a usable proxy for educati

Looking at the `Educational values deemed by the filter` table on page 5, I can't help wondering whether compressibility might not be a usable proxy for educational value. High compressibility would imply low information content. High marginal compressibility when added to a large dataset would imply the information is already in there. This is basically Normalised Compression Distance all over again - it's only a rough proxy, but I guess the question is whether it could work at all, rather than is it as good as this result. If that worked as a proxy value, you could sidestep needing GPT-4 at all.