LLM burnout is real

Alec Scollon wrote about LLM burnout and it landed. The thread on Hacker News hit 381 points in under a day. The core complaint is not that LLMs are bad. It is that building with them feels like debugging a conversation instead of engineering a system. You tune a prompt, it works, you ship it, then three weeks later it hallucinates in production because OpenAI changed something upstream. You have no stack trace. You have chat logs. This is not abstract anxiety. When you are training a model on labelled data, you know what broke. When you are orchestrating LLM calls with retrieval-augmented generation, you are guessing. The model’s reasoning is a black box wrapped in JSON. You can log the input and output but you cannot step through the decision. The comments split into two camps. Half are engineers who feel the same fatigue. The other half are people who never stopped writing deterministic code and think the first group is distracted by hype. Both are right. LLMs are useful. I have used them to write detection rules, summarise threat intel, and generate synthetic phishing datasets. They cut hours off research tasks. But integrating them into production feels like bolting a second brain onto a system that already had one. You end up maintaining two architectures: the one you control and the one you prompt. The burnout is not from using LLMs. It is from pretending they are the foundation when they work better as a tool. If your entire pipeline depends on a prompt staying stable, you built on sand.

---

Source: I think I have LLM burnout