For those of us who still didn’t understand what this is after reading TFA, it’s a way of compressing data where commonly used sequences get a short token and less common sequences need longer tokens to represent them. Just like Huffman coding except it works at a resolution of fractional bits rather than whole bits.