Reddit
"Unpopular opinion: Qwen 3.8 is hard to understand" hits 125 upvotes and 115 comments over set-theory notation in prose
A practitioner posted examples of Qwen 3.8 27B and Flash-Next writing explanations as set expressions, including a tool-availability rule written as persona tool allowlist ∩ session tool surface ∩ tools not deny-classified, rather than as a sentence. Their hypothesis is that this is what optimizing tokens-per-unit-of-intelligence produces: maximally dense output that a human then has to decompress. It is a concrete cost of the efficiency race that benchmarks do not measure, and it matters if the model's output goes anywhere near a human reader.
Source
↳ Follow the thread