I explain what I mean: those algorithms have no memory at all. Each request is made on a blank slate, so when you do a “conversation” with them, the chat program is actually including all the previous interactions (or a resume of them) plus all the relevant parts of the code, simulating a conversation with a human. So the user didn’t just ask “can you clear the cache” but actually asked the result of 600 messages + kilobytes of generated code + “can you clear the cache”, and this causes destructive hallucinations