Episode Details

Hemmingway-1, a 27B model from Altworld built on Qwen3.8-27B, claims to skip the preamble and answer your whole message instead of a piece of it. But the benchmark behind that claim, CommunicationBench, is the lab's own — and the model card says so up front. This episode digs into how directness gets measured, why the Human-Likeness metric is the softest of the three, and what RLHF, verbosity compensation, and sycophancy research tell us about why models waffle in the first place. Also: the name collision with Surge AI's Hemingway-bench, and why two m's matters. Episode #378146 — open it directly at myweirdprompts.com/378146