Hacker News Reader

Comment by ctkhn | Hacker News Reader

What extent does that happen for you? I have got very good results with self hosted qwen3.8 flash next at q4 and even with qwen3.635b on a laptop. I don't keep it running forever on every single repo, its ok to leave notes in agents.md and let it search that instead of the entire history staying in context.

View on HN

The Qwen models, especially when quantized, can be really bad about just trying things and seeing what works. If you’re not watching all the tool calls you may not see it, but it’s kind of scary to watch them just bump into wrong decisions and backtrack.

They also have a bad habit of accidentally building URLs that hit Alibaba infrastructure, likely because their training environment had them use those URLs. If you haven’t watched the outgoing network requests you might be very surprised at what your Qwen agents do sometimes.

> [Humans] can be really bad about just trying things and seeing what works. If you’re not watching all the tool calls you may not see it, but it’s kind of scary to watch them just bump into wrong decisions and backtrack.

I’m not talking coding agents. More so agent harnesses and using LLMs for decision making

heh I got a completely stupid error from gemini about some apache configs, when I pointed it out, the llm apologised, said the line(s?) should look like this, then posted the same two lines uneditted haha

Reply on news.ycombinator.com