Walter van der Giessen documents a deliberate industry shift: frontier models are trading factual knowledge for reasoning capability. GLM-5.2 hits 99.2% on AIME 2026 with ~40B active parameters while GPT-4 (~280B) could barely solve AIME in 2023 — but ask the same small models a plain factual question and hallucination rates hit 80-82%.

The mechanism:

  • Reasoning compresses well — it’s a small set of repeated procedures (break into parts, track state, backtrack). Facts need ~2 bits per parameter and don’t compress.
  • Distillation and RL on verifiable tasks transfer reasoning into small models remarkably well. Phi-4 (14B) is good at math and bad at trivia — by design.
  • The knowledge that survives is shallow breadth: enough to understand what a question is about, not enough to answer it without tools.

The argument for why this is the right trade:

  • Facts rot, procedures don’t. API docs change, prices shift, people change jobs. Half of what a 2024 model believed about the JS ecosystem was stale before it shipped. Every fact baked into weights has a shelf life and the only refresh path is another training run.
  • The harness carries the knowledge. Retrieval, tool calls, web search supply runtime facts; the model just reasons about them. A coding agent greps node_modules instead of memorizing API surfaces — and grounds its answer in the version you actually have installed.
  • Hallucination becomes debuggable. Wrong facts from weights are unfindable and unfixable (can’t grep, can’t diff, fine-tune fixes break unrelated things). Wrong facts from retrieval have an address — you open the document, edit it, write a regression test.

The punchline: a 20-40B model at 4-bit fits on the 24GB consumer GPU that’s been in gaming PCs since 2022. “The model just gets handed the world’s current state at runtime, the same way a CPU gets handed a program.”