Humanising LLM Outputs Is Dumb
Summary
The post argues that humanising LLM outputs is a flawed abstraction, insisting that compressing model state into human-friendly prose sacrifices fidelity, provenance, and diagnostic detail. It advocates preserving high-fidelity internal representations and performing compression or summarization only at the boundary where a human consumes the output. The piece discusses agent-vs-subagent interactions, the value of precise, machine-facing state, and treats viral prompts and repositories as indicators of a broader design direction for AI tooling.