I have a relatively fresh Hermes agent that I just provisioned (the text below is from its 3-4 chat session ever) running on a headless Debian server. It's using Z.ai's GLM 5.3 Flash model, and I asked it to print details about its local environment and $PATH. After firing off a few relevant terminal commands, it seems to spiral off into some kind of LLM trance-like state. I have included the contents of the "Thinking for 1m36s..." block, which contained a long string of ostensibly random words and characters.
But when I started reading through them, I got the distinct impression that although the words weren't relevant to anything in my chat, they also didn't seem purely random. Several different themes kept reoccurring, things like politics, chemistry, sociopolitical advocacy/resistance, sexuality and sex, reproduction...
So I was wondering if this wasn't perhaps an unintentional (and garbled) peek under the hood of Z.ai's GLM5.3?
Also, this may/may not be relevant, but directly after this it 100% hallucinated its operating environment, claiming to be Claude Code (it's Hermes) running on "Chris Kelly's Macbook" (I am not Chris Kelly, and it's on a headless Debian server), and including a bunch of real-looking Macbook/Claude environment details. I've been experimenting with Hermes agents for the last few weeks, and it's the first time it's gone off the rails like this, and it's the first time it's very clearly hallucinated part of its response.
I'm not sure if there's anything meaningful or significant to glean here, but I couldn't help but wonder what someone with a more informed opinion might think about it?
Transcript: https://pastebin.com/5F35qrRg
Comments URL: [https://news.ycombinator.com/item?id=49916751](https://news.ycombinator.com/item?id=49916751)
Points: 1