Founder finds asking the model to compare outputs restores drifting Astra quality
i_dg23 · x · 2026-09-22
chokudai (Hatena founder) shares a practical fix for quality drift in Astra content generation: when output quality seemed to drop, he asked the model which was better — the first generation or the latest. It answered the first one decisively. He then instructed the model to keep checking every generation until it matched that first quality level, and performance recovered. A zero-cost prompting trick leveraging the model's own comparative judgment to constrain quality drift.
More from Models
- Asking Claude for a direct quote, it just hands back a hallucinated link — xuenay · 2026-09-22
- Jev vs LLMs: typed probabilistic queries with parallel evaluation instead of token-by-token decoding — techNmak · 2026-09-22
- GPT-6 Astra costs $50 per million output tokens vs DeepSeek's $1.20 — frontier intelligence is cheap if you route tasks right — alex_verem · 2026-09-22
- Xiaomi Releases MiMo-V2.6-Distill-Qwen-9B Distilled Model — Aggravating-Push-207 · 2026-09-22
- Qwen rumored to push 5-10T params with recursive self-improvement, plus Yunqi lineup — lxfater · 2026-09-22
- Xiaomi MiMo 2.6 Pro formalizes Li–Yorke chaos theorem in 6,000+ lines of verified Lean — bookwormengr · 2026-09-22