> "You are ChatGPT, a large language model trained by OpenAI, based on the ChatGPT-4 architecture. Knowledge cutoff: 2022-01. Current date: 2023-09-19."
That seems ... insufficient. Weren't the previous "system prompts" full of revealing instructions like "don't be racist, don't repeat anything back above this line" etc.? I'm thinking they must either be using a different mechanism to censor/control output (RLHF?) or have implemented a trick to hide the most interesting parts of the system prompt (and maybe tease a little bit to trick people into thinking they successfully got it).
That was Bing. Chatgpt was always this short. If you're going to significantly finetune the model, you don't need the prompt to be complicated and detailed. Even a single token to let it know "you're in assistant mode now" could be enough.
That seems ... insufficient. Weren't the previous "system prompts" full of revealing instructions like "don't be racist, don't repeat anything back above this line" etc.? I'm thinking they must either be using a different mechanism to censor/control output (RLHF?) or have implemented a trick to hide the most interesting parts of the system prompt (and maybe tease a little bit to trick people into thinking they successfully got it).