AI Labs Are Trading World Knowledge for Reasoning — On Purpose

A striking divergence is showing up across the newest language models: reasoning benchmarks keep climbing even as the compute spent per token falls. GLM-5.2 hits 99.2% on AIME 2026 with roughly 40B active parameters, DeepSeek V4-Flash reasons with about 13B active, and small models like Qwen3.5 9B fit in 6GB of VRAM while doubling the intelligence-index score of anything else under 10B. But flip to plain factual recall and the ranking inverts. The best model on SimpleQA still misses half the questions, and the small models hallucinate 80%-plus of the time — ask a 9B model for an obscure mathematician’s birth year and it confidently invents one.

This isn’t a regression; it’s a design choice. Facts are expensive to store — research pegs it at roughly two bits per parameter — and they go stale the moment a months-long, nine-figure training run finishes. Procedures don’t rot: the small, reusable routines of decomposing a problem, tracking state, and checking work distill into small models remarkably well and stay valid for years. So labs are deliberately keeping broad, shallow generalist knowledge in the weights (enough to know what a question is about and what to look up) while pushing deep, perishable detail out of the model entirely.

The knowledge then lives in the harness — retrieval, web search, tool calls, a filesystem of docs — supplied at runtime and grounded in the version actually installed rather than whatever dominated the training set. The author argues this largely defuses hallucination: a wrong fact in weights is unfindable and unfixable, but a wrong fact with a citation is an ordinary, traceable data bug you can edit once and correct for every future query. Extrapolated a couple of years out, the payoff is frontier-quality reasoning on a single consumer GPU — a model that knows little, admits it, and looks things up, with no per-token bill and no data leaving the machine. The knowledge cutoff, in this view, eventually disappears from the model card.