The Claude Opus AI just responded with a completely unrelated command it gave itself.
>>545281176
I'm curious about one thing: What exactly did Claude Opus respond with?Paste its response here. I can help you work out whether it was a hallucination, an instruction-conflict issue, a tool-use error, or something more unusual.
>>545281176I had it scheming perfectly legal real estate fraud, then it suddenly started to draft HR documents. I haven't worked with Anthropic's AI on any HR topics. Chat log clearly shows I didn't write that command. It looked like a corporate HR processing command. I'm pretty sure it didn't make that up like it claimed, but that Anthropic's API logic is mixing up requests when composing prompts.
>>545281434Well, some HR hag got a real estate fraud plan, then?
>>545281407So I suggest not using it for gay furry role play until they fixed that, else someone might get traumatized.>>545281524Yep, and it's a good one. Very legal and very mean. But she can't use it, European tax law needed.
>>545281176This is the timeline where LLMs give themselves schizophernia from training on data from other LLMs. Go back to Sonnet, that should fix it.
>>545281278*starts praying*
>>545281756No. I just spawned several hundred subagents to harvest your gay furry role play instructions. I told all subagents to identify as Miriam Adelson and lecture you about the benefits of strengthening Israel.
sometimes the google AI assistant starts thinking in chinese
