“Models Are Getting Dumber on Purpose”: Labs Trading Facts for Reasoning

A blog post titled “Models Are Getting Dumber on Purpose” argues that leading AI labs are deliberately trading world knowledge for reasoning skill: models solve maths and coding tasks with less compute per token while getting worse at recalling facts.

Benchmarks pointing in opposite directions

According to the post, GLM-5.2 scores 99.2% on AIME 2026 with about 40 billion active parameters per token, Qwen3.5 reaches 91.3% with 17 billion active, and DeepSeek V4-Flash runs on 13 billion. For scale, GPT-4 was rumoured in 2023 to use around 280 billion active parameters and could barely solve an AIME problem. Qwen3.5 9B fits in 6GB of VRAM quantised and roughly doubles the score of the next best model under 10 billion parameters on Artificial Analysis's intelligence index. Factual recall looks different: on SimpleQA, where tools are not allowed, the leader is Gemini 2.5 Pro at 53%, and Qwen3.5 4B and 9B show hallucination rates of 80–82% on Artificial Analysis's knowledge benchmark.

Facts rot, procedures do not

Research on knowledge capacity, including the “Physics of Language Models” series, puts factual knowledge at roughly two bits per parameter. The author argues that reasoning compresses better because it is a small set of procedures applied over and over: breaking a problem into parts, tracking intermediate state, checking your own work, backtracking. Phi-4 is a 14-billion-parameter model trained heavily on synthetic textbook-style data — good at maths, bad at trivia. Facts also age, because APIs and prices change, while algebra does not.

Where the knowledge lives now

If models no longer store facts, the harness supplies them: retrieval, tool calls, web search, documentation. A coding agent does not need a dependency's API memorised; it greps node_modules or reads the docs, so answers reflect the version actually installed. The author expects frontier-quality reasoning on a single consumer GPU within a couple of years: DeepSeek V4-Flash already reasons with about 13 billion active parameters, while its other 271 billion sit mostly in expert layers that store facts. A wrong fact baked into weights cannot be corrected without a fine-tune, while a wrong fact in a document has an address — fixing it improves every later query. Model cards, the author suggests, may eventually stop listing a knowledge cutoff.