Write Like It's 1866: LLMs Relearn Telegraphese

Commenters discussed the observation that LLMs are adopting a “telegraphese” or cablese style to compress outputs and save tokens. Some found the phenomenon intriguing, pointing to historical commercial telegraph codes and to research like BabelTele that allegedly shows models can encode compact, nonstandard forms (emoji, fragments) with high semantic fidelity. Several people (andai, z2, Theory42) argued this behavior is already present in agent-to-agent messaging and model training, and that explicit prompts to write like a telegram can cut output tokens dramatically. Others noted practical uses for agent internals and system files (AGENTS.md, TOOLS.md), where settled, machine-oriented phrasing could free context budget.

Opinion split on whether the effect is meaningful or an artifact. Critics (OtherShrezzing, hnd9q09qk4) said exact-match grading and brittle benchmarks inflate the impression of gain, and criley2 pointed to Anthropic’s “mannered prose” framing and recommended instructing models to avoid that style. Several commenters mocked the idea as unserious or stylistic slop (novideonoradio, jubilanti) and complained about the article’s own AI-laden prose and imagery (klaff, soleman). Others admitted they personally use physical-object metaphors and worried this would trigger false AI-writing accusations (Hugsbox). Ultimately the divide was between those who see token-efficient telegraphese as a useful, baked-in optimization for machine communication and those who see it as a noisy benchmark artifact and an aesthetic or evaluation problem.