Gzip Plays Language Model, Sort Of

🔥 What happened A developer turned gzip — the compressor baked into every OS — into a language model. No neural net, no training, just zlib and beam search. The output: semi-coherent Shakespeare, unedited. 💡 Why it matters Every compressor hides a probability model: few bits for expected bytes, many for surprises. This confirms the compression-prediction equivalence from information theory — language modeling doesn't require a neural network, just statistics. The paper tried this and failed; beam search fixed it. ⚡ Our take Anyone calling this a GPT-killer missed the point. But it's a brutal reminder of how much "intelligence" lives in dumb compression — and how little we actually understand the line between the two.