Write Like It's 1866: How LLMs Relearn Telegraphese to Slash Token Costs
Summary
The article investigates cablese, a telegraphese-style prompt that compresses LLM outputs to save 40–49% on API token usage while preserving information. It presents the Telegraph Test benchmark, shows cross-model compression, discusses codebooks vs cablese, and argues that compression can be a practical, auditable optimization for model-to-model and human use, with caveats on reasoning, readability, and model behavior.