The article says that their zstd decoder is 10x slower than the 'normal' one. I would be interested in seeing if that can be driven down to (almost) parity, with a bit more LLM time and perhaps human effort.