The article argues that modern AI models are becoming “smarter” at reasoning while intentionally becoming less knowledgeable in their weights. Recent models such as GLM-5.2, Qwen3.5, and DeepSeek V4-Flash achieve impressive math and code benchmark results with far fewer active parameters than earlier frontier models like GPT-4. However, this improvement does not extend to factual recall: on benchmarks like SimpleQA, even the best models answer only about half the questions correctly, while smaller models hallucinate most of the time.
The author explains that this tradeoff is deliberate. Storing facts in model parameters is expensive because factual knowledge requires large capacity and quickly becomes outdated. By contrast, reasoning skills such as decomposing problems, tracking intermediate steps, and checking work can be compressed more efficiently and transferred through distillation and reinforcement learning. As a result, labs are optimizing models to be broad generalists with shallow knowledge rather than deep repositories of facts.
The article argues that the missing knowledge should come from external systems: retrieval tools, web search, documentation, and agent harnesses. This makes models more up to date, easier to correct, and less prone to unfixable hallucinations. In this vision, future frontier models may run on consumer GPUs, provide strong reasoning locally, and rely on runtime tools for current facts instead of storing them in weights.
Article timeline
Build this article’s timeline
Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.
Comments
Login to view and post comments.