tl;dr: Open-weight GLM-5.2 (753B params) scores within 4 points of GPT-5.5 on the AA Intelligence Index while hallucinating far less (28% vs 86%), suggesting that scaling parameters and training data has plateaued and often worsens uncertainty calibration. The author argues models should be evaluated on a trilemma of raw capability, hallucination rate, and compute efficiency rather than size alone, citing a coding test where DeepSeek V4 Pro burned 10x more reasoning tokens to produce a confidently wrong answer that GLM-5.2 solved in 12 seconds.