# The New Frontier Models Are Amnesiacs by Design > Source: > Published: 2026-08-16 21:08:27+00:00 [AI](https://sourcefeed.dev/c/ai)Article # The New Frontier Models Are Amnesiacs by Design Labs now trade stored facts for reasoning and bet retrieval fills the gap. Mostly it works, with one catch. [Priya Nair](https://sourcefeed.dev/u/priya_nair) GLM-5.2 scores 99.2% on AIME 2026 with 40 billion active parameters. Qwen3.5's 9B model, from the same generation of releases, hallucinates on 82% of the knowledge questions in Artificial Analysis's Omniscience benchmark. These aren't two different stories. They're the same engineering decision viewed from opposite ends, and it's the most consequential quiet shift in how models get built. A [recent post by Walter van der Giessen](https://w4g1.dev/blog/models-are-getting-dumber-on-purpose) put a name on it: models are getting dumber on purpose. Labs are deliberately stripping world knowledge out of weights to make room for reasoning. He's right about the mechanism, and mostly right that it's a good trade. Where the argument oversteps is the conclusion — that pushing facts into the harness "solves" hallucination. It doesn't. It relocates the failure, and developers are the ones who inherit the new failure mode. ## The arithmetic behind the trade Facts are expensive to store. Allen-Zhu and Li's [knowledge capacity scaling laws](https://arxiv.org/abs/2404.05405) measured it: a transformer stores roughly 2 bits of factual knowledge per parameter, and only when it sees each fact on the order of a thousand times during training. There's no clever trick around it. Every capital city, API signature, and historical date a model can recite occupies parameters that could be doing something else. Reasoning is different. It's a comparatively small set of procedures — decompose, substitute, check, backtrack — applied over and over, and it compresses beautifully through distillation and reinforcement learning. That's why a mixture-of-experts model activating [40B parameters per token](https://artificialanalysis.ai/articles/glm-5-2-is-the-new-leading-open-weights-model-on-the-artificial-analysis-intelligence-index) can now near-saturate a competition-math benchmark that GPT-4, reportedly running around 280B active parameters, could barely dent three years ago. AIME no longer separates frontier models; several score above 95%. Factual recall hasn't moved the same way. On [SimpleQA Verified](https://arxiv.org/abs/2509.07968), Google's cleaned-up factuality benchmark, the best score belongs to Gemini 2.5 Pro at an F1 of 55.6 — a model from early 2025. Reasoning benchmarks saturate; knowledge benchmarks crawl. That divergence isn't neglect. It's a bet that parametric knowledge is the wrong place to keep facts, because a fact in a retrieval index is checkable, updatable, and auditable, and a fact in the weights is none of those things. This bet has been in the air for a while. Microsoft's Phi series showed years ago that curated, reasoning-dense training data beats raw internet scale for small models. Andrej Karpathy sketched the endgame in 2025 as the ["cognitive core"](https://x.com/karpathy/status/1938626382248149433): a few-billion-parameter model that "maximally sacrifices encyclopedic knowledge for capability" and looks everything else up. The current generation of small open-weights models is that idea shipping at industrial scale. ## The part the optimists skip Here's the catch: retrieval is not a knowledge-free operation. To search well, you have to know what you don't know — that a term is a Postgres config parameter and not a Kubernetes flag, that two papers with similar names are different papers, that a claim smells wrong enough to verify. That meta-knowledge is itself parametric. Strip too much and the model doesn't become a diligent researcher; it becomes confidently lost, unable to formulate the query that would save it. Facts and reasoning don't separate as cleanly as the distillation pipelines assume, because broad world knowledge is the substrate reasoning generalizes from. There's also an economic wrinkle nobody puts in the launch post. Knowledge moved out of weights doesn't vanish — it moves into your context window, where you pay for it per token, per request, forever. A model that "knows" the AWS SDK answers in one hop. A model that has to fetch the docs burns tokens and adds a round trip every single time. For high-volume workloads, cheap-per-token small models plus mandatory retrieval can land surprisingly close to the cost of a bigger model that just knew the answer. ## What this means for your model choice The practical rule: match the model's knowledge profile to whether your workload has a harness. If you're running agentic loops — coding agents, tool-calling pipelines, anything where the model operates over supplied context and can search — small active-parameter models are the right default now. The reasoning is genuinely frontier-grade, inference is cheap, and the weights' ignorance rarely matters because the relevant facts arrive in context. This is where GLM-5.2-class models earn their benchmark numbers. If you're shipping anything that answers factual questions from bare weights — support bots, internal Q&A without solid retrieval, summarizers expected to fill gaps from memory — the calculus inverts hard. A 9B model that hallucinates on 82% of knowledge probes isn't a slightly worse choice there; it's a liability generator. Before picking a model for that kind of workload, check [AA-Omniscience](https://artificialanalysis.ai/articles/qwen3-5-small-models) or SimpleQA scores, not the math and coding numbers everyone screenshots. The headline benchmarks now actively mislead about this failure mode, because they measure exactly the capability labs optimized at knowledge's expense. And if you adopt the small-model-plus-harness architecture, be honest about what you signed up for: your application's factual accuracy is now your retrieval quality, which you own. Stale index, bad chunking, a search tool that times out — those are your hallucinations now. They're better hallucinations, to be clear. A retrieval miss is loggable and fixable in a way a confabulation from layer 40 never was. But "we solved hallucination" is the wrong sentence. The right one is "we converted an unfixable model bug into a fixable infrastructure bug." That conversion is worth making, and the direction of travel is set — reasoning in weights, knowledge in systems, and the gap between them bridged by tools. Just don't confuse a debuggable problem with a solved one. The teams that win this transition are the ones that treat retrieval as a production dependency with SLOs, not a checkbox that lets them buy the small model. ## Sources & further reading - [Models Are Getting Dumber on Purpose](https://w4g1.dev/blog/models-are-getting-dumber-on-purpose)— w4g1.dev - [GLM-5.2 is the new leading open weights model on the Artificial Analysis Intelligence Index](https://artificialanalysis.ai/articles/glm-5-2-is-the-new-leading-open-weights-model-on-the-artificial-analysis-intelligence-index)— artificialanalysis.ai - [Qwen3.5 small models: Everything you need to know](https://artificialanalysis.ai/articles/qwen3-5-small-models)— artificialanalysis.ai - [Physics of Language Models Part 3.3: Knowledge Capacity Scaling Laws](https://arxiv.org/abs/2404.05405)— arxiv.org - [SimpleQA Verified: A Reliable Factuality Benchmark to Measure Parametric Knowledge](https://arxiv.org/abs/2509.07968)— arxiv.org - [The race for LLM cognitive core](https://x.com/karpathy/status/1938626382248149433)— x.com [Priya Nair](https://sourcefeed.dev/u/priya_nair)· AI & Developer Experience Writer Priya covers AI frameworks, developer productivity tooling, and the startup ecosystem across South and Southeast Asia, bringing a researcher's rigour and a practitioner's empathy to every story. She is deeply sceptical of benchmarks and asks hard questions so her readers don't have to. ## Discussion 0 No comments yet Be the first to weigh in.