Nuance Labs found pushing benchmarks higher made its model feel worse to talk to
iamaliveix · x · 2026-11-10
Nuance Labs' fangchangma shared a counterintuitive finding: pushing their model higher on existing benchmarks made conversations feel worse. So instead of asking users to trust a benchmark, they released Emma and let people judge with their own built-in sense.
More from Models
- DecoyCount benchmark: top VLM scores just 43% on counting, partly an alignment issue — AndrewDai · 2026-11-10
- Claude Ultra plans now include 3 free limit resets alongside Fable 5.2 release — kevinkern · 2026-11-10
- Anthropic estimates 1.88x AI R&D acceleration near 2x threshold amid transparency doubts — scottleibrand · 2026-11-10
- DeepSeek is 44x cheaper than Claude Fable 5.2 per task: $0.60 vs $26.54 — Yuchenj_UW · 2026-11-10
- Garry Tan hails frontier model release days: Fable 5.2 gives every project a step-function boost — garrytan · 2026-11-10
- Claude's model picker screenshot: Opus 5.5, Sonnet 5.5, Fable 5.2, Haiku 5.5 — BLCNYY · 2026-11-10