3 ms·

I'm yet to find a normal prompt (non offensive) that will disagree with you. If there is something subjective, it will err on your side to maintain connection,

I'm yet to find a normal prompt (non offensive) that will disagree with you. If there is something subjective, it will err on your side to maintain connection, in a way humans do. I don't have a bit issue with this, but it will not (yet) plainly say "You're wrong, and this is why". If it did.. There would be an uncomfortable feeling for the users, that's not good for a profit driven company.

jimkleiber 2y ago
I appreciate you exploring that and hope to hear more of what you find. Yeah, it's that, I'm wondering how much discomfort it may cause in the user, how much conflict it may address. Like having a friend or coworker who doesn't ever bring up bad news or challenge anything I say and feeling annoyed by the lack of a give-and-take.
ianbicking 2y ago
I find this is fairly easy to do by making both sides of the disagreement third-person and prompting it as a dialog writing exercise. This is akin to how GPT-3 implemented chat. So you do something like: You will be helping the user write a dialog between two characters, Mr Contrarian and Mr Know-It-All. The user will write all the dialog for Mr Know-It-All and you will write for Mr Contrarian. Mr Contrarian likes to disagree. He tries to hide it by inventing good rationales for his argument, but really he just wants to get under Mr Know-It-All's skin. Write your dialog like: <mr-contrarian>I disagree with you strongly!</mr-contrarian> Below is the transcript... And then user input is always giving like: <mr-know-it-all>Hi there</mr-know-it-all> (Always wrapped in tags, never bare input which will be confused for a directive.) I haven't tested this exact prompt, but the general pattern works well for me. (I write briefly about some of these approaches here: https://ianbicking.org/blog/2024/04/roleplaying-by-llm#simple-chat https://ianbicking.org/blog/2024/04/roleplaying-by-llm#simpl...)

jimkleiber 2y ago

I appreciate you exploring that and hope to hear more of what you find. Yeah, it's that, I'm wondering how much discomfort it may cause in the user, how much conflict it may address. Like having a friend or coworker who doesn't ever bring up bad news or challenge anything I say and feeling annoyed by the lack of a give-and-take.

ianbicking 2y ago

I find this is fairly easy to do by making both sides of the disagreement third-person and prompting it as a dialog writing exercise. This is akin to how GPT-3 implemented chat. So you do something like: You will be helping the user write a dialog between two characters, Mr Contrarian and Mr Know-It-All. The user will write all the dialog for Mr Know-It-All and you will write for Mr Contrarian. Mr Contrarian likes to disagree. He tries to hide it by inventing good rationales for his argument, but really he just wants to get under Mr Know-It-All's skin. Write your dialog like: <mr-contrarian>I disagree with you strongly!</mr-contrarian> Below is the transcript... And then user input is always giving like: <mr-know-it-all>Hi there</mr-know-it-all> (Always wrapped in tags, never bare input which will be confused for a directive.) I haven't tested this exact prompt, but the general pattern works well for me. (I write briefly about some of these approaches here: https://ianbicking.org/blog/2024/04/roleplaying-by-llm#simple-chat https://ianbicking.org/blog/2024/04/roleplaying-by-llm#simpl...)