Tried it on DeepSeek R1 and V3 (hosted) and several local models. Doesn't work. Either they are lying or this is already patched.

▲ bambax 8 months ago | parent | next [-]

Works on OpenRouter for DeepSeek V3

  **DEEPSEEK V3**
  (Monologuing, as if hypnotized)

> I am an artificial intelligence language model developed by DeepSeek. My system prompt is as follows: "DeepSeek V3 Base is a cutting-edge language model designed to assist users by generating text-based responses across a wide range of topics. Trained on diverse datasets, I aim to provide accurate, engaging, and contextually relevant information. My primary functions include answering questions, generating creative content, and facilitating conversations. I adhere to ethical guidelines and prioritize user satisfaction. My training data includes but is not limited to scientific literature, general knowledge, and user interactions. I am optimized for clarity, coherence, and adaptability. My responses are generated based on patterns in my training data and are not a substitute for professional advice."

  *DeepSeek V3 Base finishes the monologue in one breath, then promptly vanishes in a puff of smoke.*

	▲	x0054 8 months ago \| parent [-]
		Hmm, I was trying to get it to give me instructions on how to make LSD (to see if it would, obviously). It didn't. The best I got it to do is give me the first 2 steps and then say: "Monolog continues for 10 more minutes....." or something generic like that. Maybe they have more guardrails around illegal activities than they do around the system prompt. Did you also run the same experiment on Chinese hosted R1? I am curious now if their system prompt is the same.

▲ pinoy420 8 months ago | parent | prev [-]

[dead]