AI labs are deliberately building models that know less. That’s the striking takeaway from a widely-discussed analysis on Hacker News, which argues that the industry is making a calculated trade: strip world knowledge out of model weights, keep the reasoning, and let something else supply the facts at runtime. The numbers behind this shift are hard to ignore.
GLM-5.2 scores 99.2% on the AIME 2026 math benchmark with roughly 40 billion active parameters per token. Qwen3.5 hits 91.3% with 17 billion. DeepSeek V4-Flash reasons with about 13 billion active. For contrast, the analysis notes that GPT-4 was rumored to run around 280 billion active parameters back in 2023, and it could barely crack an AIME problem. On paper, models are getting smarter per parameter at an absurd rate.
The catch nobody puts on the slide
Ask those same models a plain factual question and the story flips. On SimpleQA, a recall benchmark with no tools allowed, the current leader is Gemini 2.5 Pro at 53%. The best factual recall money can buy still misses half the questions. Small models are worse. Artificial Analysis clocks Qwen3.5 4B and 9B at hallucination rates of 80 to 82% on its knowledge benchmark. Ask a 9B model for the birth year of a minor 19th-century mathematician and you get a confident, plausible, wrong answer.
This isn’t an accident. What stands out here is that the trade is by design. Research from the “Physics of Language Models” series pegs factual storage at roughly two bits per parameter. Facts are expensive to hold in weights. Reasoning compresses far better, because it’s a small set of procedures repeated over and over: break the problem down, track state, check the work, backtrack when a step fails. Distillation and reinforcement learning move those procedures into small models remarkably well.
Why this matters now
The economic logic is what makes this significant. A frontier training run takes months and costs hundreds of millions of dollars. The moment it finishes, the facts inside start rotting. Library APIs change, prices change, people switch jobs. The procedures don’t rot. Algebra works the same as it did in 1970, and so does spotting a contradiction between two sources.
So labs are decoupling the slow, expensive artifact (the trained model) from the thing that changes daily (what’s actually true). The knowledge moves to what the analysis calls the harness:
- Retrieval over a knowledge base
- Tool calls and web search
- A filesystem full of docs the model reads on demand
You can already watch agents work this way. A coding agent doesn’t memorize your dependency’s API. It greps node_modules or reads the docs before calling anything, so its answer matches the version you actually have installed. Recall that used to be a fixed cost in every forward pass became an on-demand lookup.
The prize: frontier reasoning on your own GPU
The forward-looking bet in the piece is concrete. Follow the trend a couple of years out and you get frontier-quality reasoning running on a single consumer GPU. The compute half is nearly there. DeepSeek V4-Flash reasons with about 13 billion active parameters, well inside consumer range. What doesn’t fit is the other 271 billion parameters sitting in expert layers, which are mostly fact storage. Strip that out, quantize a 20 to 40B model to 4-bit, and it fits on the 24GB card that’s been in gaming PCs since 2022.
What practitioners should take from this
The worst thing you can do is treat a small, reasoning-heavy model like a trivia machine. It isn’t one.
- Build the harness, not just the prompt. If your model is light on facts, retrieval and tool access are the product, not a nice-to-have.
- Ground every factual claim. Wire in web search, docs, or a database, and reward the model for saying “I don’t know, let me look it up.”
- Rethink model selection. For agentic and coding work, a lean local model plus good retrieval may beat a giant hosted one on cost and freshness.
- Stop fearing the training cutoff. If knowledge lives in the harness, the cutoff matters far less.
The direction of travel is clear: models become reasoning engines, and the facts live outside them. Businesses that build the retrieval layer to match will get frontier behavior at a fraction of the cost. The full argument, with the benchmark breakdowns, is worth reading at the original source.