CongoSky Skunkworks
A compressor is a predictor: bytes it expects cost almost nothing, bytes that surprise it cost the most. Type below — a tiny order-2 model learns your text as it reads, and we colour each character by how surprised the model was (its information content, in bits). Predictable text glows green and compresses; random text stays gold and can't.
This is the live intuition behind the compression proof: an adaptive context model assigns each symbol a probability from what it has seen, and the ideal code length is −log₂(p) bits (Shannon). The real proof goes further — context-mixing across orders, a true arithmetic coder, and a benchmark where it beats xz by 26% on shared-domain text and honestly loses to LZ on big repetitive files. Same idea, measured.