July 10, 2026
AI fight night: bots at dawn
Ask HN: What was the last task where only a frontier model could do it?
Coders are fighting over whether the fancy AI is truly special or just better at chaos
TLDR: A Hacker News thread asked for real examples where only the newest AI could handle a task, and the answers showed a split: cheaper models are fine with supervision, but many users still don’t trust them on bigger jobs alone. The comments mattered most, with people arguing over whether “good enough” means useful helper or genuinely reliable worker.
A simple question on Hacker News — when did only the most advanced AI actually save the day? — turned into a full-on comment section cage match. The original post challenged a popular claim: that free or open AI models, even if they lag behind the leaders by months, are already “good enough” for most real work. And the crowd’s answer was a giant, messy, very online it depends.
One camp basically said the hype is real: the top-tier models still pull ahead when jobs get long, complicated, or require staying on track without constant babysitting. That was the vibe from commenters who said cheaper models are fine for quick help, but start wobbling when you want them to handle bigger projects on their own. In plain English: they’ll help you hammer nails, but don’t ask them to build the whole house.
Then came the chaos goblins. One user described using AI to tear apart and rebuild 1990s DOS games in Rust, which sounds like a hobby designed in a laboratory to break robots. Another dropped a link to a game project and casually reported that even the premium model had a meltdown, with one version apparently bungling an algorithm so badly it had to be rescued from Git, prompting instant “the AI ate its own homework” energy.
The spiciest twist? Some commenters flipped the whole debate around: forget asking when only the best AI works — when can open-source bots do a project alone at all? That hot take gave the thread its mood: less victory lap, more suspicious side-eye. The community wasn’t just comparing tools; it was arguing over whether “good enough” means helpful assistant or trustworthy co-worker. And judging by the reactions, that question is still very much setting the comments on fire.
Key Points
- •The post questions the claim that open-weight models about six months behind frontier models are sufficient for most work.
- •It asks for concrete task examples from the last month rather than general impressions.
- •The requested comparisons specifically mention GLM, DeepSeek, Kimi, and Qwen versus Opus, Fable, and GPT.
- •The author also invites examples where the reverse happened, with a cheaper or open model succeeding where a frontier model did not.
- •A structured template is provided to capture task details, model failures, model successes, and whether a slightly older frontier model would have been enough.