Write Like It's 1866: LLMs Relearn Telegraphese (fiveminutesforward.com)

A recent study has explored the use of "cablese," a concise communication style inspired by 19th-century telegraphy, to enhance the efficiency of large language models (LLMs). By instructing models to respond in this terse register, researchers demonstrated significant savings in token usage, with models compressing outputs by 25% to 49% depending on the architecture. This approach allows models to convey the same information with fewer tokens while maintaining accuracy, thereby optimizing costs associated with LLM queries. The findings come from the "Telegraph Test," which assessed model performance across a 50-passage benchmark, revealing that models could process compressed records as effectively as plain text. The significance of this discovery lies in its potential to reduce operational costs for AI applications, as outputs written in cablese incur less expense than traditional worded responses. This compression technique harnesses latent capabilities in existing models, making it feasible without requiring new hardware or significant retraining. Furthermore, the flexibility of cablese, which remains human-readable and auditable, provides an advantage over other emergent communication protocols developed under token constraints. As AI continues to evolve, integrating such strategies could lead to substantial efficiencies in how information is processed and shared across systems.