DeepSeek 'Harness Paper' Roasted for Lacking Actual Evals, Filled with Hardcore Math
teortaxesTex · x · 2026-08-14
Developer @teortaxesTex mocked a paper circulating as the 'DeepSeek harness paper,' noting that it ironically never mentions DeepSeek or its harness, nor provides any actual evaluations.
Instead, the paper reads like a vision statement that semi-formalizes concepts named 'Cordis' and 'Koishi' (possibly a Chinese chatbot app) using unreasonably hardcore math, leaving readers thoroughly confused.
More from Fun
- Sabine Hossenfelder fires back: academia's impact metric turned physics into a paper mill — skdh · 2026-10-16
- Meme: Clicking 'Yes to all' on Claude Code without reading anything — jnack · 2026-10-16
- Minecraft AI mod turns region selection + prompts into autonomous building agent — Acclynn · 2026-10-16
- Throwaway prop code in movies is getting eerily accurate, and AI is why — mike64_t · 2026-10-16
- Distillation copies logits, not hidden CoT — open-sourcing 1.3 leaks IP, not reasoning data — teortaxesTex · 2026-10-16
- Scott Aaronson's 9-year-old son: "I don't think problems should be independent of the axioms" — basedjensen · 2026-10-16