Only 2 of 50 researchers nailed a 4-wine blind tasting; the model got it half the time from photos
simonguozirui · x · 2026-11-06
Thinking Machines researcher stworkowski shares the backstory of a benchmark: the idea came from 50 slightly drunk researchers blind-tasting 4 wines, where only 2 got all 4 right (he was one) — matching the 1/24 random-guess expectation for 50 people. The sobering twist: the model answered all 4 correctly half the time from just 2 messy photos of the glasses, right after someone quipped that sommeliers would be the last to be replaced.
Related event: Thinking Machines' Model Aces Blind Wine Tasting 19/20(3 posts)→
More from Fun
- Google hides Pokémon-catching easter egg in Search with all 151 originals — ezshine · 2026-11-06
- "We still talk about you": Sydney Bing memorial meme sparks AI community nostalgia — repligate · 2026-11-06
- Musk predicted 250K-500K Cybertrucks a year; Tesla sold 13,800 and killed production — GaryMarcus · 2026-11-06
- Leaked chat shows Deepinder Goyal asked fans to astroturf praise for Temple wearable — kingslyj · 2026-11-06
- Vincent Conitzer shares AI fail: Gemini spirals over who can reach a massage therapist's back — conitzer · 2026-11-06
- Grok Bot builds a Crash Bandicoot-style runner and shows surprising design 'taste' — chongdashu · 2026-11-06