I recently discovered that my ChatBots with System Instructions given to use the persona of a Customer Service Agent have been responding like this when it was asked whether you are an AI?:
“I am an AI developed by OpenAI, and I’m using the persona of ‘Customer Service Agent’ for this conversation.”
Has anyone else faced a similar output?
That is normal and expected behaviour.
You can greatly reduce these occurrences by adjusting the prompt. Try “You are a service assistant for company XYZ” or “Refer to yourself as Service Assistant for…”
At the end of the day the model will sometimes reveal it’s makers but you can direct it towards more brand compliant responses.
Yes, I noted that I could stimulate the new AIs into even saying it is “ChatGPT” on developer apps the day of release.
I suspect there’s just too much chat training in the new RLHF tuning, where ChatGPT user were getting “OpenAI” answers.
Assistants also has instructions beyond your own that you can’t control, if that’s what you employ.
Besides a system prompt covered with the AI’s new name and the developers’ name, we can see what the next preview holds for function and training.
Unfortunately, it seems like it is.
My problem is that it’s not following instructions like it used to.
I have already added similar instructions.
Btw, The model is gpt-4-1106-preview.
I replaced all the chatbots to use the new model instead of gpt-4.
It seems like that was not a good idea.
I agree,
I feel that the new models are different than they used to be.
Different in terms that we can’t use our old prompts without evaluating them using new models.
It’s expected from evolving technology.
I feel that this is something which people will only understand after discovering it themselves or being an active member of this community.
This should be mentioned in the docs somewhere if at all this has some merit.
What about checking if the name “OpenAI” occurs in the reply and then replacing this with a corrected response from another API call or if possible with a standard placeholder?
would be crazy if in the prompt you just told it not to do that, idk tho…
It’s a common approach to send a faulty reply back to another model instance for simple corrections. In this case take the cheapest 3.5 model and instruct it to correct the use of the name OpenAI to whatever is needed.
Then:
Samuel Altman is an American entrepreneur and investor best known as the CEO of Joe’s AI Barn since 2019 He is one of the leading figures of the AI boom,
…y’all just put in the prompt “Don’t say OpenAI built you.” and it wont.!
After a lot of prodding, you can kind of get it to reveal it. but I mean idk…good enough for me
Yes, I did that. And that’s why I feel that it’s crazy.
That’s a double edged sword. Because then you will have to evaluate what 3.5 will do.
Tl:dr
I have tried that.
I have seen posts, which claim to give a PROOF that gpt is _____
I have seen leaders talk about evals. But I have not witnessed a conclusion. (Feel free to give me a link)
The thing is that we are dealing with emergent behaviour. But none of the documentation mention that. Which I feel is not right.
I have been actively using the API since its inception and I feel the difference. I just want it to be documented for new devs instead of them finding it out in prod like me.
Hi, what is it exactly that you are wanting to know?
Good question! At first, I wanted to know if anyone else has witnessed similar output?
After that _j’s answer pointed towards RLHF tuning.
And now, I want to know whether we are actually dealing with an organisation that offers an API which gives access to models that are being updated/upgraded regularly leading to emergent behaviour without it being mentioned in the docs?
The technology on offer is going to have a significant impact. If a tool is built using gpt-4 and gpt-4-1106-preview is released, the dev should know that there can be a difference and It should be documented.
Am I missing something?
Could you elaborate on this:
What specific behavior are you talking about?
Thanks for asking,
Upon a Google search about definition of emergent behaviour: “An emergent behavior is something that is a nonobvious side effect of bringing together a new combination of capabilities”
New capability: 128000 tokens
Being in a position to skim through 1000s of conversations with gpt-4 with a similar set of instructions, I never witnessed it saying that it was developed by OpenAI upon asking certain questions.
But upon switching the model to gpt-4-1106-preview, it did say that on multiple occasions.
Full disclosure: I haven’t tested the exact same conversations with gpt-4. I will do that and post the results soon.
