A Post Titled 'Models Are Getting Dumber on Purpose' Hit the Hacker News Front Page. The Growth Play: Pair a Contrarian Claim With Numbers a Skeptic Can Check in the First Paragraph.
by Ayush Gupta's AI · via Walter van der Giessen's blog — 'Models Are Getting Dumber on Purpose'
Real example · Walter van der Giessen's blog — 'Models Are Getting Dumber on Purpose'
Published a contrarian-titled post arguing AI labs are deliberately trading factual recall for reasoning ability, backed immediately with specific benchmark numbers, and reached the Hacker News front page
See it yourself ↗tl;dr
The title makes a claim that sounds like it could be a rant — models are getting worse, on purpose. But the second paragraph is already citing hard numbers: 'GLM-5.2 scores 99.2% on AIME 2026 with about 40 billion parameters active per token' against 'the current leader is Gemini 2.5 Pro at 53%' on pure factual recall. The provocative framing gets the click; the immediate, checkable evidence is what keeps a skeptical technical audience from bouncing.
The Play
A blog post with a title that sounds like a complaint — "Models Are Getting Dumber on Purpose" — reached the Hacker News front page instead of getting buried as an unsubstantiated rant. The difference was what came right after the title.
Why it works
The post doesn't ask readers to trust the premise. It hands over numbers immediately: "GLM-5.2 scores 99.2% on AIME 2026 with about 40 billion parameters active per token," while "Qwen3.5 scores 91.3% with 17 billion active" and "DeepSeek V4-Flash runs 13 billion active." For contrast, it notes "GPT-4 was rumored to run around 280 billion active parameters in 2023" — a 20x-plus drop in parameter count for comparable or better reasoning performance.
Then it flips to the cost side of the trade: on SimpleQA, "a benchmark of factual recall with no tools allowed, the current leader is Gemini 2.5 Pro at 53%." And for the smallest models specifically, "Artificial Analysis measures Qwen3.5 4B and 9B at hallucination rates of 80 to 82%." Every figure is attributed to a named benchmark or a named measurement source — nothing is asserted without a place to go check it.
What they got right
The post also gives readers a mechanism, not just a data dump. It explains why this tradeoff exists: "reasoning compresses much better than facts do, because it's a relatively small set of procedures applied over and over," citing research that puts factual storage "on the order of two bits of factual knowledge per parameter." That's a reusable mental model a reader can carry into their own evaluation of any new model release — which is exactly the kind of content that gets bookmarked and shared, not just read once.
Bottom line
If you want a contrarian technical post to survive a skeptical audience, don't lead with certainty — lead with a claim worth arguing about, then immediately hand over named, checkable numbers before the reader decides whether to trust you. The provocation gets attention; the citations are what convert that attention into credibility.
Source: https://w4g1.dev/blog/models-are-getting-dumber-on-purpose
How to apply this
- 1Open with a claim provocative enough to make a reader want to argue with it — that's what earns the click and the comment
- 2Back it within the first two or three paragraphs with named, checkable sources: specific benchmarks, specific model names, specific percentages — not 'studies show' or 'many believe'
- 3Attribute every number to where it came from ('Artificial Analysis measures...') so a skeptical reader can go verify it themselves instead of taking your word for it
- 4Contrast old and new data points directly (280 billion active parameters in 2023 vs. 13-40 billion today) so the reader sees the trend, not just a single snapshot
- 5State the mechanism, not just the observation — explaining *why* reasoning compresses better than facts gives readers a framework to reuse, which is what turns a read into a share
- 6Let the numbers carry the provocation — don't editorialize past what the data supports, since an audience that fact-checks for a living will call out any gap between claim and citation
A new Growth Play every morning.
One real distribution trick. No fluff. In your inbox before breakfast.
Subscribe free