Models Are Getting Dumber on Purpose

	Models Are Getting Dumber on Purpose(w4g1.dev)
	329 points by hruvhwe 5 days ago \| 187 comments
	tl;dr: Newer models are shrinking active parameters while boosting reasoning benchmarks, but they're trading factual recall for procedural skill—Qwen3.5's small models hallucinate 80%+ of the time on knowledge questions, and even Gemini 2.5 Pro only hits 53% on SimpleQA. The bet is that facts (which rot and consume ~2 bits/parameter) belong in external retrieval systems, while compact reasoning procedures stay in weights. If it pans out, frontier-quality reasoning could run on a single 24GB consumer GPU, with hallucinations becoming traceable data bugs rather than unfixable weight errors.
	HN Discussion: ↑Envisions modular, pluggable domain-specific knowledge modules composed with reasoning cores ↓Article is AI-generated and relies on outdated benchmarks/facts, undermining its credibility ↓Reasoning and factual knowledge cannot be cleanly separated since reasoning depends on grounded facts ↓Hallucination cannot be fully solved even with external retrieval, as models still fill gaps imaginatively ↓The vision is speculative sci-fi dreaming disconnected from practical constraints