You can write docs in LLMese, but you don't have to (passo.uno)

A recent exploration into how large language models (LLMs) can communicate prompted an innovative experiment on whether documents can be effectively compressed into a terse format, termed "LLMese." The study involved creating a prototype, llmify, which tests various methods of condensing textual information while maintaining clarity and comprehension. By applying semantic rewrites and token pruning techniques to documents from Kubernetes and a fictional tech scenario, the research evaluated how well LLMs could process these compressed versions. Remarkably, semantic rewrites preserved meaning better than aggressive token pruning, proving that LLMs can respond accurately to questions based on the condensed documents. This investigation holds significant implications for the AI/ML community as it addresses the logistical challenge of optimizing LLM interactions without losing critical information. The findings suggest that while there's potential for efficient communication between LLMs through a structured dialect, it is crucial to prioritize human-centric writing first. Ultimately, the research underscores the point that while LLMs may not yet possess their own distinct vernacular, adapting human-written content into machine-readable formats can enhance model efficiency and performance. However, care must be taken to ensure essential information isn't sacrificed in the compression process.