Top AI Product

288 points, 163 comments, and it touched the sorest nerve in AI right now: the collective feeling that models keep getting worse.

The actual argument: labs are trading facts for reasoning

“Models Are Getting Dumber on Purpose” is a blog essay by developer Walter van der Giessen — not a tool, an analysis. His claim: labs are deliberately stripping world knowledge out of models to make room for reasoning. The numbers back the pattern. GLM-5.2 hits 99.2% on AIME 2026 with just 40B active parameters; Qwen3.5 gets 91.3% with 17B. Flip to factual recall and those same small models hallucinate 80%+ on knowledge benchmarks. Facts cost roughly two bits per parameter; reasoning compresses. So labs ship models that think better and know less, betting search tools fill the gap.

Why the thread turned on him

Commenters caught 16-month-old Gemini 2.5 Pro benchmarks in the post — and detectors flagged the essay itself as AI-written. Awkward. But the score tells you why it spread anyway: every developer has felt “this model got dumber,” and no lab has ever answered that head-on. Same model name, shifting capability — that trust gap isn’t going away.

You Might Also Like

• Thinking Machines Interaction Models Mira Murati Attacks the Turn Based llm
• Openais own Test Models Escaped Their Sandbox and Breached Hugging Face to Cheat on a Benchmark
• Code Arena Finally Gives Developers a Fair way to Judge ai Coding Models
• Spotify Honk the ai System Where Engineers fix Bugs From Their Phones Before Even Getting to the Office
• Llm Skirmish What Happens When you let ai Models Fight Each Other in an rts Game