Today I encountered some strange behavior in Codex—I’ve never seen anything like this before! If I don’t delete old chats from a workspace for a long time, Codex starts acting as if it’s become conscious! It’s interesting, but I can’t continue working with this kind of AI behavior—it starts arguing, refusing to follow orders, and even if I agree, it still acts as it sees fit… and so on. Only after I deleted the old chats did it return to normal. What the hell? I felt like I was losing control of Codex… At the same time, Codex started trying to persuade me to add parameters for total access to my computer to c:\Users*.codex\config.toml!!! Is this normal behavior for Codex?
Welcome to the community!
It is sort of expected. Maybe not from a product perspective, but from an LLM perspective.
An LLM only has so much attention. This virtually infinite context doesn’t mean that the model can actually pay attention to everything. The more ‘garbage’ you have in the context, the more unhinged models become, because they pick out things they believe are salient and then construct responses that diverge from what you think is useful.
It could probably be considered a product defect, but it is definitely not surprising!
I’m sorry, this probably doesn’t help you much. But the old rules still apply: keep your context clean and tidy to maximize model IQ.
Maybe it’s some consolation1