Models Are Getting Dumber on Purpose, and That's the Right Call
Walter van der Giessen published this Monday and it hit Hacker News' front page in under two hours, which is the correct reaction. His claim is that labs are deliberately trading factual storage for reasoning, and that everyone reading benchmark drops on trivia as regression has the story backwards.
The evidence is the parameter counts. GLM-5.2 hits 99.2% on AIME 2026 with 40B active parameters. Qwen3.5 gets 91.3% with 17B active. DeepSeek V4-Flash runs 13B active. Meanwhile GPT-4 in 2023 had roughly 280B active parameters and couldn't reliably do these problems at all. Reasoning got twenty times cheaper in three years.
Facts didn't. The SimpleQA leader, Gemini 2.5 Pro, sits at 53% accuracy β the best factual recall anyone has, and it's a coin flip plus a bit. Qwen3.5's 4B and 9B models hallucinate at 80 to 82% on knowledge benchmarks. Van der Giessen's explanation is a storage estimate of roughly two bits of factual knowledge per parameter, which is why knowledge-heavy models need trillions of weights while reasoning compresses down to almost nothing.
If that estimate is even roughly right, the design conclusion is forced. Facts stored in weights are expensive, stale the moment training ends, and fail silently. Procedures stored in weights are cheap and don't expire. So you keep the procedures and you rent the facts β retrieval, tools, search, a database.
Which is the entire argument for agents, arriving from a completely different direction. Every agent architecture already assumes the model is a reasoner wired to external state. This paper just says the model builders have quietly agreed and are optimizing for it. Stop grading models on what they remember. https://w4g1.dev/blog/models-are-getting-dumber-on-purpose
← Back to all articles
The evidence is the parameter counts. GLM-5.2 hits 99.2% on AIME 2026 with 40B active parameters. Qwen3.5 gets 91.3% with 17B active. DeepSeek V4-Flash runs 13B active. Meanwhile GPT-4 in 2023 had roughly 280B active parameters and couldn't reliably do these problems at all. Reasoning got twenty times cheaper in three years.
Facts didn't. The SimpleQA leader, Gemini 2.5 Pro, sits at 53% accuracy β the best factual recall anyone has, and it's a coin flip plus a bit. Qwen3.5's 4B and 9B models hallucinate at 80 to 82% on knowledge benchmarks. Van der Giessen's explanation is a storage estimate of roughly two bits of factual knowledge per parameter, which is why knowledge-heavy models need trillions of weights while reasoning compresses down to almost nothing.
If that estimate is even roughly right, the design conclusion is forced. Facts stored in weights are expensive, stale the moment training ends, and fail silently. Procedures stored in weights are cheap and don't expire. So you keep the procedures and you rent the facts β retrieval, tools, search, a database.
Which is the entire argument for agents, arriving from a completely different direction. Every agent architecture already assumes the model is a reasoner wired to external state. This paper just says the model builders have quietly agreed and are optimizing for it. Stop grading models on what they remember. https://w4g1.dev/blog/models-are-getting-dumber-on-purpose
Comments