A Hacker News thread catalogs where LLMs fail

An Ask HN thread asked what large language models still get wrong. Two failure modes stood out from the discussion, which drew 68 comments on Hacker News.

Commenters described architectural floor plans that came out as nonsense even when every measurement and room requirement was spelled out. ASCII-map games like Nethack surfaced the same gap. Models that write working code still can't hold a map in their head and plan a path across it.

A second cluster is about output shape, not broken reasoning, and it matters more for anyone building on these systems. Commenters flagged redundant keyword strings like "nhl toronto scores nhl hockey toronto scores." They also pointed to over-explaining when distillation was asked for, and prose that settles into one unvaried sentence structure. Several replies described Claude ignoring documented project rules daily, unauthorized git commits, wrong ripgrep flags. I've watched that exact pattern in my own week using it.

Spatial failures describe a capability that never existed. Sloppy output and rule violations describe a model getting worse at tasks it used to handle. That difference is worth tracking. Regressions are fixable in a way missing capabilities aren't.

The thread kept collecting replies, cataloging failures that benchmarks don't test for.