GPT-6.1 invents Chinese-character dictionary encoding to compress Lean proof packets
kfountou · x · 2026-11-01
GPT-6.1 Sol spontaneously used Chinese characters and unusual punctuation to compress repeated Lean expressions in its workflow. Per the explanation screenshots, these weren't gibberish but dictionary references: repeated long expressions were replaced by single-character identifiers (e.g. 中 → @MeasureTheory.Measure), with full expressions appearing once in a dictionary.
- Chinese characters provide a large alphabet, enabling thousands of distinct references in very few characters; the odd punctuation abbreviates repeated JSON syntax and escaping
- The goal was fitting large Lean evidence packets into the reviewer's context without dropping information
- The author clarifies the caveat: fewer bytes doesn't automatically mean fewer tokens or better mathematical understanding; lossless reconstruction was verified and one packet showed 28.2% byte reduction, but that alone doesn't establish token savings
More from Fun
- 'Man + machine': a brief visual musing on human-AI fusion — apostraphi · 2026-11-01
- Dev builds a webcam-controlled pumpkin carving game that's surprisingly hard — jh3yy · 2026-11-01
- Voice model tone can be steered via on-the-fly text annotations — repligate · 2026-11-01
- A week of AI coding made me feel like an 'expediter', not a programmer — sloppenheimer · 2026-11-01
- Google scientist shares pirate AR filter Halloween moment: 'you're muted' — docmilanfar · 2026-11-01
- AI lab logo candy bowl goes viral: 'This would go so hard in SF' — evilrabbit_ · 2026-11-01