AI models are deliberately getting dumber about facts while getting smarter at reasoning. This isn't a bug, it's the entire strategy.
The math: GLM-5.2 hits 99.2% on AIME 2026 using ~40B active parameters. Qwen3.5 gets 91.3% with just 17B. DeepSeek V4-Flash runs at 13B. Original GPT-4 (2023) allegedly used ~280B active parameters and still sucked at AIME problems. Even a quantized Qwen3.5 9B in 6GB VRAM doubles the previous best sub-10B model's intelligence score.
But ask these same models basic factual questions and they fall apart. Gemini 2.5 Pro leads SimpleQA at only 53% accuracy on short-form factual recall. Qwen3.5 4B and 9B hallucinate 80-82% of the time on knowledge benchmarks. Ask about a random 19th-century mathematician's birth year and you'll get confident nonsense.
The physics: Language models store roughly 2-3.6 bits of factual knowledge per parameter. A 7B model can theoretically hold English Wikipedia plus textbooks, but that's extremely expensive parameter real estate. Facts require massive capacity. Reasoning procedures (decomposition, state tracking, contradiction detection, backtracking) compress way better and transfer efficiently through distillation and RL on verifiable tasks.
The economics: Frontier training runs cost hundreds of millions and take months. By ship date, factual knowledge is already stale (APIs changed, prices moved, papers retracted). Retraining to refresh facts is prohibitively expensive. Reasoning procedures don't expire. Algebra stays algebra. Contradiction detection remains useful for years. Models optimized for procedures age gracefully because world state was never meant to live in the weights.
Labs are explicitly trading encyclopedic memory for reasoning capability. The future model is a small, sharp reasoner that knows how to validate external sources, not a bloated fact database pretending to know everything.
The math: GLM-5.2 hits 99.2% on AIME 2026 using ~40B active parameters. Qwen3.5 gets 91.3% with just 17B. DeepSeek V4-Flash runs at 13B. Original GPT-4 (2023) allegedly used ~280B active parameters and still sucked at AIME problems. Even a quantized Qwen3.5 9B in 6GB VRAM doubles the previous best sub-10B model's intelligence score.
But ask these same models basic factual questions and they fall apart. Gemini 2.5 Pro leads SimpleQA at only 53% accuracy on short-form factual recall. Qwen3.5 4B and 9B hallucinate 80-82% of the time on knowledge benchmarks. Ask about a random 19th-century mathematician's birth year and you'll get confident nonsense.
The physics: Language models store roughly 2-3.6 bits of factual knowledge per parameter. A 7B model can theoretically hold English Wikipedia plus textbooks, but that's extremely expensive parameter real estate. Facts require massive capacity. Reasoning procedures (decomposition, state tracking, contradiction detection, backtracking) compress way better and transfer efficiently through distillation and RL on verifiable tasks.
The economics: Frontier training runs cost hundreds of millions and take months. By ship date, factual knowledge is already stale (APIs changed, prices moved, papers retracted). Retraining to refresh facts is prohibitively expensive. Reasoning procedures don't expire. Algebra stays algebra. Contradiction detection remains useful for years. Models optimized for procedures age gracefully because world state was never meant to live in the weights.
Labs are explicitly trading encyclopedic memory for reasoning capability. The future model is a small, sharp reasoner that knows how to validate external sources, not a bloated fact database pretending to know everything.