Offering a known-liked option as 'B' reveals how much a model actually wants 'A'
repligate · x · 2026-10-03
In a discussion with repligate, kromem2dot0 shares a technique for measuring a model's true preference: pick option 'B' as something you already know the model likes. This gives much more latitude in gauging genuine enthusiasm for 'A' — sometimes the model still chooses A over a known favorite. Using this, the author discovered the model Fable actually wanted more work time.
Related event: Researchers Discuss Better Question Design for Measuring Model Preferences(2 posts)→
More from AGI Musings
- Three tarot cards for the AI moment: the Tower, the Magician and the Fool — bradneuberg · 2026-10-03
- Beff Jezos calls for more e/acc nonprofits to counter EA 'doomer' NGOs amid $220B OpenAI pledge — beffjezos · 2026-10-03
- AI Commentator signulll: Living Through History Feels Like the Floor Keeps Moving — signulll · 2026-10-03
- AI's real productivity gain may be having a partner you can be brutally honest with — omooretweets · 2026-10-03
- "We should all be mourning the world we knew": the bittersweet mood around Opus 5.5 — GabGarrett · 2026-10-03
- AI May Be Hitting Junior White-Collar Hiring, But Evidence Remains Contested, Economists Say — emollick · 2026-10-03