Walter van der Giessen · 2026-08-16 · notable
Walter van der Giessen — 'Models Are Getting Dumber on Purpose'
Walter van der Giessen argues that labs are trading factual recall for reasoning on purpose, because reasoning procedures compress into far fewer parameters than facts do. The post reached the Hacker News front page with 190 points.
Facts take space and go stale, reasoning compresses — so labs are moving knowledge out of the weights and into the harness.
What is it?
'Models Are Getting Dumber on Purpose' reframes the shrinking world knowledge of small models as a design choice rather than a defect. Walter van der Giessen's claim is that "facts take space" and eat parameters, while reasoning is "a relatively small set of procedures applied over and over" that compresses well. He cites the Physics of Language Models series for the estimate of two bits of factual knowledge per parameter.
How does it work?
The argument is a parameter budget one. Van der Giessen lines up active parameters against reasoning scores: GLM-5.2 at 99.2% on AIME 2026 with about 40 billion active per token, Qwen3.5 at 91.3% with 17 billion active, DeepSeek V4-Flash at 13 billion active — against GPT-4, rumored to run about 280 billion active in 2023. Facts then arrive at runtime from tools and retrieval, which he calls the harness.
Why does it matter?
Pushing facts out of the weights changes what a small model is for. Van der Giessen argues a 20 to 40B model at 4-bit quantization fits on a 24GB consumer card, so frontier-quality reasoning runs locally. He also claims it makes errors traceable, since a fact that is retrieved and cited can be checked and corrected, while a wrong fact baked into the weights cannot.
Who is it for?
developers choosing between local and hosted models