March 30, 2026
Vibe-coded and mildly cursed
Ask HN: Where have you found the coding limits of current models?
Dev Drama: AI code hits a wall—too chatty, bad at bugs, needs a babysitter
TLDR: An Ask HN crowd says AI is fine for small, testable tasks but stumbles when left to architect systems, often patching bugs with random delays and over-verbose code. The debate splits between “keep a human in the loop” and “why can’t it write clean, consistent designs?”—a reality check for anyone betting on auto-coding.
The latest Hacker News thread asks where AI coding tools break—and the comments turned into a roast with receipts. The loudest take: AI can write code, but it can’t be trusted to steer the ship. One dev says models don’t “push back” on bad ideas and even quietly change important values. Another reports a spooky fix for tricky threading bugs: models sprinkle in random delays to “let things settle,” AKA sleepy-time patches for problems that need real engineering.
Plenty of folks swear by AI for small, well-defined tasks—with tests, linting, and a human editor with “taste.” But when users ask it to build bigger systems, the vibe turns chaotic: unnecessary abstractions, duplicate code, and weird inventions that no one asked for. There’s mini-drama around style too: some say models drown everything in verbose code, even when begged for compact solutions.
Model face-off? One commenter says Anthropic’s Sonnet stumbled on a common list-merging task, while Opus 4.6 pulled off the trick—suggesting progress, but not peace of mind. Others complain AI gets lost in layers of object-oriented design, like it understands the idea but can’t create it cleanly. The meme of the day: “vibecoded systems”—fast, flashy, and held together by vibes and sleep calls. The crowd verdict: great assistant, not your CTO.
Key Points
- •An Ask HN post seeks examples of where current coding models reach practical limits.
- •The request highlights dimensions like lines of code, systems integration, and databases.
- •The goal is to gather real-world cases that show capability boundaries and failure modes.
- •The thread is hosted on Hacker News as an open community question.
- •At capture time, it had 26 points and 39 comments, indicating active engagement.