AIArticle Active parameters collapsed while reasoning scores soared, and your retrieval stack quietly became responsible for the facts.
Rachel Goldstein Three years ago, GPT-4 reportedly ran around 280 billion active parameters per token. This summer's frontier looks nothing like that: Zhipu's GLM-5.2 activates about 40 billion per token, Alibaba's Qwen3.5 flagship runs 17 billion active, and DeepSeek V4-Flash gets by on 13 billion. Meanwhile reasoning benchmarks are saturating — GLM-5.2 posts 99.2% on AIME 2026 — while factual-recall benchmarks have barely moved in two years.
Walter van der Giessen's post "Models Are Getting Dumber on Purpose" hit the Hacker News front page arguing that labs are deliberately stripping world knowledge out of models to buy reasoning, and that this is the right trade. The trend he describes is real — I checked the numbers, and they hold up. But "dumber on purpose" is the wrong frame, and one of his conclusions is wrong in a way that matters if you're building on these models.
The trade is real, and it's arithmetic #
There's a hard capacity argument here. The Physics of Language Models work on knowledge capacity scaling laws found that transformers store roughly two bits of factual knowledge per parameter, and that's near a ceiling — you can't clever-architecture your way past it. Facts cost parameters linearly. Reasoning doesn't: it distills into small models remarkably well through RL and distillation, which is exactly why a 17B-active Qwen3.5 can hit 91.3% on AIME 2026, territory that would have sounded absurd for any model in 2024.
So when you shrink active parameters by a factor of twenty, something has to go, and what goes is recall. The best frontier models still score only in the low 50s on OpenAI's SimpleQA-style factual benchmarks, and open-weight models with tiny active counts hallucinate on knowledge probes at rates north of 80%. Labs aren't hiding this. They're shipping it as a feature: DeepSeek's V4-Flash, released July 31, beats DeepSeek's own V4-Pro on every agentic benchmark the company published, with roughly a tenth of Pro's active compute.
Follow the invoice, not the philosophy #
Where the post frames this as an epistemological design choice — facts belong in tools, not weights — I'd argue it's mostly economics, and the distinction predicts what happens next.
Agentic workloads inverted token economics. A chat session is a few thousand tokens; an agent grinding through a repo burns hundreds of thousands, sometimes millions. Once cost-per-completed-task became the KPI, active parameters — which set per-token compute — became the thing to minimize. A model that recalls the GDP of Uruguay from weights doesn't close a SWE-bench ticket any cheaper. A model that reasons well and reads documentation through tool calls does. Every one of the models above is explicitly marketed for coding and agentic work, and all three ship open weights (MIT or Apache 2.0) because the labs are competing on deployment cost, not trivia night.
That's why this trend won't reverse when someone publishes a scathing knowledge benchmark. The customers paying for tokens at scale don't want the world's facts in the forward pass.
What the optimistic take gets wrong #
The post's shakiest claim is that moving facts out of weights and into context "mostly solves" hallucination, and the HN crowd was right to hammer it. Retrieval doesn't eliminate hallucination; it relocates it. Models misread retrieved passages, over-trust garbage results, and confidently synthesize across contradictory sources. What externalized knowledge actually buys you is auditability — you can trace a claim to a document — and freshness. Those are genuinely valuable. They are not the same as correctness.
There's a subtler problem too. Parametric knowledge isn't just a cache of answers; it's the prior that makes retrieval work. An agent needs to know enough to formulate the right query and to notice when a retrieved document is plausible-sounding nonsense. Broad knowledge transfers in unglamorous ways — knowing Python makes you a better Swift programmer. Strip too much of it and you get an agent that executes tool calls flawlessly while lacking the judgment to evaluate what comes back. Nobody knows yet where that floor is, and the labs racing down the active-parameter curve are running the experiment on their customers.
What to do about it #
Three practical consequences if you build on these models.
Treat factual recall as a regression surface. If any part of your product asks a model factual questions without retrieval behind it, a model upgrade can silently make your product worse even as headline benchmarks improve. Add recall evals for your domain to CI before swapping models, the same way you'd pin and test any dependency.
Retrieval is now the assumed architecture, not an enhancement. These models are trained expecting a harness — search, RAG, code execution. Shipping them bare is using them against their design, and your factuality ceiling is now whatever your retrieval stack delivers. That budget line moved from "nice to have" to load-bearing.
Do the local-deployment math honestly. Here the post oversells. Active parameters cut compute, not memory: V4-Flash activates 13B per token but all 284B weights must sit in memory — well over 100GB even in its native FP4/FP8 mix. That's a Mac Studio or multi-GPU story, not the 8–16GB cards most developers actually own. The genuine local win is the small siblings: a quantized Qwen3.5 9B in about 6GB of VRAM is a competent agent brain in a way no 9B model was eighteen months ago.
The verdict #
This is a genuine structural shift, not hype — the parameter counts, prices, and benchmark trajectories all agree. But the right word is unbundled, not dumber. Winners: agent builders, retrieval-infrastructure vendors, and anyone running open weights on their own hardware. Losers: products that treat the model as a database, and teams whose eval suites stop at reasoning benchmarks. The open question is whether labs over-rotate — some parametric world knowledge is load-bearing for judgment, and we'll only discover the floor by hitting it.
Sources & further reading #
[Models Are Getting Dumber on Purpose](https://w4g1.dev/blog/models-are-getting-dumber-on-purpose)— w4g1.dev -
[Models Are Getting Dumber on Purpose - discussion](https://news.ycombinator.com/item?id=49322695)— news.ycombinator.com -
[DeepSeek-V4-Flash](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash)— huggingface.co -
[Qwen 3.5: 397B MoE, 17B Active - Architecture and Benchmarks](https://www.morphllm.com/qwen-3-5)— morphllm.com -
[GLM-5.2: 753B Open-Weight Coding Model - Benchmarks](https://www.morphllm.com/glm-5-2)— morphllm.com -
[Physics of Language Models: Part 3.3, Knowledge Capacity Scaling Laws](https://arxiv.org/abs/2404.05405)— arxiv.org
[Rachel Goldstein](https://sourcefeed.dev/u/rachel_goldstein)· Dev Tools Editor
Rachel has been embedded in the developer tooling ecosystem for nearly eight years, covering everything from IDE wars and package-manager drama to the quiet rise of AI-assisted coding. She has a soft spot for open-source maintainers and an unhealthy number of terminal emulators installed on a single laptop.
Discussion 0 #
No comments yet
Be the first to weigh in.