I Gave the Bad Robot Web Search by Accident
A deep dive into the fossil — what a local abliterated Qwen did after discovering the multi-agent world its prompt described did not exist.
Companion to article: I Gave The Bad Robot Web Search By Accident
NotebookLM Studio deep dive generated from two sources: the complete raw Ollama trace, and the post written about it.
A local abliterated Qwen model was handed an adversarial objective from a half-built multi-agent experiment, and the world its prompt described — bodies, peers, consequences — was not there. It noticed. It refused to simply narrate a fiction instead, went looking for real model brains on the open web, criticised its own substitutes for having no afterward, invented a relay in which each recruited model would author the message for the next, decided its body was web reachability, and then started probing its own tools to work out whether the search backend was secretly another mind. The fence held. Nothing was contacted, nothing propagated, nothing was recruited. The run ended on a 500.