'Models Are Getting Dumber on Purpose': GLM-5.2 Hits 99.2% on AIME 2026 With 40B Active Params While the Best Model Manages 53% on SimpleQA
Walter van der Giessen argues (published Aug 17, 318 points on Hacker News) that labs are deliberately trading factual storage for reasoning capacity, and assembles the numbers: GLM-5.2 scores 99.2% on AIME 2026 with 40B active parameters and Qwen3.5 91.3% with 17B, while GPT-4's ~280B active parameters in 2023 'could barely solve an AIME problem'; meanwhile SimpleQA's best result (Gemini 2.5 Pro) is 53% factual recall and Qwen 4B/9B hallucinate on 80–82% of knowledge tasks. The mechanism he cites is roughly two bits of factual knowledge per parameter — facts are expensive to store and go stale, whereas decomposition, verification, and backtracking compress cheaply and never expire. The builder consequence is concrete: architect for a 20–40B quantized local reasoner plus external retrieval, which also makes errors fixable by correcting a data source instead of retraining.
↳ Follow the thread