A Bitter Truth: Even the smartest models struggle to connect the dots
“Taste is subjective.” TL;DR One classic logic puzzle + a subtle cue (“Poison tastes bitter”). I asked leading LLMs to solve it. Every model failed on the first try. Only 4 models (Opus 4, OpenAI o3 and Gemini 2.5 Pro) understood the clue even after I spelled it out. What this says about model intelligence, training data and the US-China AI race. 1. The Puzzle and the Twist Original Puzzle: ...