Each cell is one character. Color shows how predictable it is from what came before — how much information it carries.
For each character, the algorithm looks back through the preceding text and finds the longest substring that matches what comes next. A long match means the character was predictable — it carries little new information, can be compressed away. No match means it's genuinely novel.
Regular text (AAAA...) is almost entirely blue: everything is predicted by what came before. Random text is almost entirely red: no patterns, nothing to compress. English prose shows structure: common words cool to blue after their first use; rare words and first occurrences glow warm. The heatmap is a map of where meaning lives.
This is what compressors do — and what understanding does. Both find the shorter description. The blue parts are already known. The red parts are what you came for.
See also: Understanding Is Compression