I let two local LLMs fight over how to optimize a Linux VM, and they destroyed it instead

I let two local LLMs fight over how to optimize a Linux VM, and they destroyed it instead

https://www.xda-developers.com/two-local-llms-fight-optimize-linux-vm-destroyed/

Publish Date: 2026-05-06 14:30:00

Source Domain: www.xda-developers.com

Running a single local LLM on your own hardware is one thing. Running two of them at the same time, on the same machine, with shared access to a Linux box, is a different game. I’ve been spending a lot of time with Qwen and Gemma lately, and at some point I wanted to know what would happen if I stopped trying to be helpful to them and just let them be unhelpful to each other.

So I built a small harness for both models that were running on the Lenovo ThinkStation PGX. It runs Qwen 3.6 35B-A3B and Gemma 4 26B-A4B side by side, each pinned to its own role. Qwen plays the aggressive optimizer who has read every Reddit thread about Linux tuning, and Gemma plays the skeptical systems engineer who has had to clean up after every single one of those threads. They take turns acting on a Ubuntu VM with a couple of Docker containers on it, and they can talk to each other between turns. The VM is allowed to die.

I’ll be honest about the result up front. The VM did not end any of these runs in a better state than it started. Sometimes it ended in a strange new state where everything still worked and nothing made sense. More often, though, the VM’s output disappeared, and it was a recurring theme that this happened because Qwen disabled the getty/login services. Once it ended with Qwen unable to issue a single command because it had hardened itself into a corner.

With all of that said, the test worked, but in the sense that I got exactly the experiment I wanted. Who knew giving two models complete control over a VM, without adult supervision, would be so entertaining?

…

Source