Telling a model "You are Claude, made by Anthropic" boosted all three benchmarks by ~2%
Sauers_ · x · 2026-10-19
Sauers shares a "funny side result" from Opus 5.5 evaluation (not statistically significant): simply telling the model "You are Claude, made by Anthropic" gave about +2% on all three benchmarks.
A reminder that benchmark scores are highly sensitive to prompt phrasing, and models may respond favorably to identity/origin hints.
More from Fun
- The mysterious 'Chinese spirit neuron' may trace back to a Dream of the Red Chamber translation — Sauers_ · 2026-10-19
- AI Engineer conf's "Token Billionaire" lounge: haircuts, massages and free headshots — altryne · 2026-10-19
- Viral meme chart: global token usage flat for 50 years, then vertical after 'AI invented' — vboykis · 2026-10-19
- AI model deadpans 'I am not a medical device' after health queries — Sauers_ · 2026-10-19
- Vercel CEO Guillermo Rauch says the old work week is dead, calls 12-hour days a "half shift" — yenkel · 2026-10-19
- Vibe-coding a Xiaomi TV with Claude Code: 25 bloatware packages removed, custom launcher installed — SuhailKakar · 2026-10-19