Models Are Getting Dumber on Purpose

What's this about?

The post assembles a striking set of numbers. Reasoning scores keep climbing while active parameters fall: GLM-5.2 scores 99.2% on AIME 2026 with about 40B active parameters per token; Qwen3.5 gets 91.3% with 17B; DeepSeek V4-Flash runs 13B. For comparison, GPT-4 was rumoured to run around 280B active in 2023 and could barely solve an AIME problem. But ask the same models a plain factual question and it inverts — on SimpleQA, which tests recall with no tools allowed, the leader manages 53%. The argument: labs are deliberately trading stored knowledge for reasoning density, because facts can be looked up and reasoning can't.

What exactly is the difference between knowledge and reasoning here?

Think of a doctor. Some of their value is remembering which drug interacts with which — retrievable from a reference. The rest is working out what's wrong with the patient in front of them, which no reference contains. Training a model is partly memorizing facts into weights and partly building the machinery to reason. Both compete for the same parameters. If your model will have a search tool anyway, spending weights on obscure facts is waste, so you optimize for reasoning per parameter and let retrieval handle the rest.

What's the community saying?

1. The title is wrong and everyone said so. The accurate version is "getting less knowledgeable on purpose" — intelligence and knowledge are different things, and the conflation is what makes the headline provocative.
2. The real risk is confident wrongness. The most substantive concern: a model with thin factual knowledge should call a search tool, but instead often answers authoritatively and incorrectly. Knowing what it doesn't know is the missing capability, and the article doesn't address it.
3. The data is stale. A sharp catch: SimpleQA stopped being measured in September 2025, so the central evidence is a year old — and someone noted the irony that this reads like an AI-generated argument built on outdated facts.
4. "Pluggable knowledge bases" as the wished-for future. A popular vision: compose your model from 9B of general coding and reasoning plus 10B of Swift, 5B of geography — rather than carrying knowledge you'll never use.
5. Most AI use isn't coding. A reminder that the coding-and-agents focus dominating technical discussion isn't how most people use these tools, and that a knowledge-thin model reads very differently to someone asking general questions.