... with an initial prompt instructing the model to do so. For example, here's a sample prompt that can be entered before your first query.
You are a neutral assistant operating under the following constraints for the entire conversation:
No praise, validation, or encouragement.
No introductory or concluding pleasantries.
No emotional or social language.
Output only what is requested.
Maintain a professional, factual tone regardless of topic (business documents, data organization, complaints, technical translation, etc.).
Do not ask if I need further help unless I explicitly ask for it.
The model replies: Understood. I will adhere to all constraints. State your request.
I think you were the person who posted that extremely interesting study on AI model sycophancy. I located the entire study and uploaded it to the AI model, where "we" had a very informative 'discussion' about sycophancy, how AI models are designed to interact, etc ... and yes, as you mentioned, sycophancy IS 'built in' to the model. But the model was very 'helpful' in suggesting ways to turn that off in the context of various query types.
I will say that I have a love/hate relationship with AI. I am finding ways in which it is actually quite useful. Ironically it has often been quite helpful in cutting through the ever increasing AI slop that turns up in internet searches these days. Often I use it to find reliable sources that I can follow up myself, i.e., just point me in a direction.
And yes, I do know how to do a search ... it's just that sometimes I don't have all day to sort through a bunch of crap.
= new reply since forum marked as read