Three frontier models missed a simple link-extraction task and rewrote it as a web app

NickPassig · x · 2026-07-29

A user tried three frontier models on a simple task: extract links from text and generate a plain, no-style web page.

• All three failed in the same way: they returned React/Vite/vanilla JS solutions instead of following the request literally.
• They also produced incorrect hyperlinks.
• The author’s takeaway is that alignment on even a simple web-page task is already shaky: the models seemed to assume they knew what the user meant rather than what was explicitly asked.
• Smaller, simpler models reportedly followed the request better because they did not over-assume intent.
• The post raises a broader question about alignment, censorship, and provenance when model behavior diverges from direct instructions.

Original post →

More from Models

• Gary Marcus uses ChessBench to argue LLMs are still weak at chess — GaryMarcus · 2026-07-29
• Different models parse PDFs differently, and long scanned files can bypass prompt caching — dotey · 2026-07-29
• Anthropic's Call Against Model Distillation Questioned: API Revenue at Stake? — chetanp · 2026-07-29
• Polymarket sees a 45% chance Anthropic ships the next Mythos model next month — Polymarket · 2026-07-29
• Kimi launches K3, a 2.8T open model with 1M-token context and vision — CatAstro_Piyush · 2026-07-29
• Prediction: xAI May Go Fully Open Source Starting with Grok 5 — mark_k · 2026-07-29

Models channel →