Write Like It's 1866: LLMs Relearn Telegraphese

The article shows that a simple prompt instruction can make multiple LLM families write in a compressed, machine-readable “cablese” that preserves downstream accuracy while cutting token usage by roughly 25% to 49%, potentially reducing output costs and effectively doubling agent memory capacity. For CIOs and IT leaders, the strategic implication is that AI economics can be improved immediately—without new hardware, training, or API changes—by storing scratchpads, summaries, and agent handoffs in compressed form, then expanding only for human consumption. The caveat is that this works best for model-consumed text and can backfire with mandatory-reasoning models, so adoption should be targeted and benchmarked per model.