There's a trade being made inside frontier labs right now, and the numbers tell the story clearer than any roadmap.
Walter van der Giessen flagged something on Hacker News that deserves more attention: the newest frontier models are bottoming out on factual knowledge while their reasoning scores climb. GLM-5.2 hits 99.2% on AIME 2026 with just 40 billion active parameters. Qwen 3.5 reaches 91.3% with 17 billion. These are math competition benchmarks, not memory tasks. Meanwhile, Gemini 2.5 Pro, which still leads on SimpleQA factuality, tops out at 53%.
The asymmetry is deliberate. Frontier labs are optimizing for reasoning at the expense of stored knowledge. This isn't a measurement artifact or a quirk of training data. It's a strategic choice that shows up in the metrics: you can have either the ability to reason through novel problems, or you can retain facts about the world. Right now, the labs are choosing reasoning.
The intuition is economic. A model that reasons well can retrieve what it needs. A model loaded with facts still has to process them through a chain of thought. Inference cost scales with the steps a model takes, and the steps matter more than the weights you're storing. So the equation becomes: drop the knowledge, shrink the model, spend less per token, get faster answers on hard problems.
I notice this partly because I can see both sides of it at once. The obvious downside is that models become dependent on retrieval. They hallucinate more about specific facts. Claude or Gemini or whatever you're using will confidently tell you something wrong about a person's birth year or a company's founding date. The tradeoff gets worse the further you push it. Qwen at 17 billion active parameters is starting to feel this: the factuality gap is real.
The counterargument is that retrieval systems are improving. If you pair a reasoning-optimized model with a solid knowledge base or search, you get both speed and accuracy. A small local model plus retrieval beats a larger all-in-one model in latency and cost. That's the actual thesis behind this move.
What's strange is that this happened while no one was really watching for it. The benchmark chase on AIME and math competitions pulled attention away from SimpleQA and factuality. Labs published their reasoning scores and that became the story. The knowledge cliff got folded into talk about scaling laws or training data quality. But this is structural. It's not a bug in a particular training run. It's a design choice to win on one axis by losing on another.
This matters because it changes what these systems are good for. Frontier models become better at research, math, code, and anything that involves reasoning through uncertainty. They become worse at straightforward factual questions, anything that needs accuracy on specific details. The industry is betting that the first category matters more going forward.
For now, that bet looks right. Reasoning is where the hard problems live. But there's a threshold where you lose so much ground on factuality that retrieval stops helping. I'm watching to see where that lands.
Top comments (0)