AI Models Are Trading Facts for Reasoning Skills
Original: Models Are Getting Dumber on Purpose
Why This Matters
Signals a structural shift in AI architecture strategy: smaller, reasoning-focused models paired with retrieval tools may replace knowledge-dense frontier models.
Modern AI models are deliberately reducing stored world knowledge in favor of reasoning capabilities. GLM-5.2 scores 99.2% on AIME 2026 with ~40B active parameters, while GPT-4 ran ~280B parameters in 2023 and could barely solve AIME. Meanwhile, factual recall leader Gemini 2.5 Pro scores only 53% on SimpleQA.
A technical analysis by Walter van der Giessen (published August 17, 2026) argues that AI labs are deliberately engineering smaller, more reasoning-capable models at the cost of factual recall. GLM-5.2 achieves 99.2% on AIME 2026 with approximately 40 billion active parameters per token. Qwen3.5 scores 91.3% with 17 billion active parameters, and DeepSeek V4-Flash runs on 13 billion — a stark contrast to GPT-4's rumored 280 billion active parameters in 2023, which could barely solve AIME problems. The tradeoff is measurable on the other side: on SimpleQA, the best factual recall model (Gemini 2.5 Pro) scores only 53%, while Qwen3.5 4B and 9B show hallucination rates of 80–82% on knowledge benchmarks. The post cites the 'Physics of Language Models' research series, which estimates approximately two bits of factual knowledge stored per parameter, making broad factual depth expensive to maintain in weights. Reasoning procedures, by contrast, compress well and transfer effectively via distillation and reinforcement learning on verifiable tasks. The author notes facts stored in model weights also 'rot' as APIs, prices, and events change, while reasoning procedures remain stable over time. Models retain breadth of knowledge — enough to understand topics and judge source plausibility — while offloading depth to retrieval systems.