Write Like It's 1866: LLMs Relearn Telegraphese
4 hours ago
- Cablese, a telegraph-style compressed language, can reduce LLM output tokens by 40–49% with minimal accuracy loss when read by other models (recovery ratios 0.99–1.10).
- The technique requires only a one-sentence instruction and no training; it works across multiple model families, leveraging latent training data from historical telegraphy.
- Compression is most effective after content is finalized (e.g., agent memory, handoffs) and for machine-consumed text, not during model composition.
- Mandatory-reasoning models like GPT-5-mini incur higher costs with cablese; models with controllable reasoning benefit fully.
- Cablese is human-readable and auditable, unlike emergent protocols that are dense but unstable and opaque.
- The Telegraph Test provides a reproducible benchmark measuring a model's compressibility, legibility (cross-model reading), and cost efficiency.
- Historical practices (telegraph cablese, codebooks) directly inspired the technique, as token-based billing mirrors per-word telegraph costs.
- Cablese compression outperforms simple codebook substitution (~10% savings), and decoded/stored records retain nearly all plaintext accuracy.