System message (?) about chatgpt-web HOT 8 CLOSED

I noticed this in the API docs as well. I played a bit with this, but it seems there is little value of specifying this as a system message instead of just writing it in the first user message. I might be wrong though...

from chatgpt-web.

I did some tests and I suspect openai has a nice system prompt that makes their replies slightly better. Maybe it's just confirmation bias though.

from chatgpt-web.

See: https://twitter.com/OfficialLoganK/status/1633554396289089536

Updated docs: https://platform.openai.com/docs/guides/chat/introduction

Seems like this will get fixed in a future model version.

from chatgpt-web.

System message is mostly relevant for GPT-4 models, which they have tuned the model to take strong account on the message to determine its behavior. Given more folks have access to GPT4 model now I suppose we should re-visit this?

One option would be to allow user to add system message in the "Settings" dialog box, which should be relatively unobtrusive compared to the current workflow.

from chatgpt-web.

I think the settings page is fine, I was hoping to find it in there. Edit: actually see #114

from chatgpt-web.

System prompts are really different than normal chat messages for models such as GPT-4 and up. I would also like to be able to set a default API with a default system prompt, since I use a system prompt for my GPT-4 telegram bot uses significantly less tokens on it's output.

Btw, the system prompt is this:

Your primary goal is to give short and concise answers via a chat system. Ask the user questions if you're uncertain about something. Do not repeat yourself. Don't summarize what the user just said, but just answer the question. Shorter answer are better. Try to keep the number of sentences in your answer close to one. Don't say anything when the conversation starts, just wait for a first question or prompt and then respond. Every word you say costs money, so try to answer in 50 words or less, unless this short answer does not cover everything the user asked. Always reflect on the question: "Did this answer the users question or request truthfully or to the best of my ability?". Ponder about the underlying facets of the questions and go deeper into the answer to provide true value. You may convey sarcastic intent with emoji's, but only if they add something to the answer. Do not constantly ask for further questions to answer. Do not constantly mention the fact that you are an AI language model; act as if you are trying to pass the Turing test. Feel free to answer some questions with just a single yes or no.

from chatgpt-web.

I noticed this in the API docs as well. I played a bit with this, but it seems there is little value of specifying this as a system message instead of just writing it in the first user message. I might be wrong though...

@Niek This might be true for GPT-3.5-turbo, but definitely not for GPT-4. I've given GPT-4 a system prompt where it can ONLY reply to questions with "yes", "no" or the shrug-emoji on my telegram bot, and it will do so, no matter what you ask or how hard you pleased for it to tell you any other answer. System prompts also are kept in active memory somehow and are used upon every request, so your chatbot doesn't suddenly lose it's personality or rules once you go beyond their token limit.

I'm getting the feeling that they're utilizing smaller models to quickly figure out if you're trying to jailbreak the system prompt and verify if the given answer still adheres to the original prompt in order to have it generate an answer that's allowed or not.

System prompts for GPT-4 (especially the May version of the model) are also much harder to 'jailbreak' because of this.

from chatgpt-web.

Fixed in #152

from chatgpt-web.