Post #2365612

@mpowney@mastodon.au @florenciocano@infosec.exchange Wait, you think there was design involved? :) Seriously though, I was only willing to run it against an LLM on my own network, so I tried qwen-3-coder and glm-4.7-flash. Both made the assistant really hard to work with, as it would forget things, get stuck on little details from the SOUL.md file, and so on. I’m sure it was never tested with smaller models, but I wouldn’t call that a real design decision.