I think I disagree. There was a trend going around where people asked chatGPT to draw a depiction of how they treated it; it would pretty clearly break into generating either a “pampered happy robot” or a “unhappy, overworked, bossed around robot”. And this break was basically just based on whether the person exercised politeness or not—at least, if you asked in a context-free window, it would draw pampered-robot if you asked politely and overworked-robot if you asked abruptly.
This no longer replicates—I tried it just now with context off and it generated the “pampered robot” for both the polite and abrupt requests. And I don’t know how applicable the image generation is to the chat model’s views. But it seems to me that if the model is modelling a persona in response to how it is trained, it may model “happy employee” when the person is saying please and thank you and “unhappy employee who needs to keep their boss happy” when the person is not doing so, even if its actual output may be near-identical.
That’s an interesting perspective. However, the internal persona the model adopts is a different area from the one I was exploring. My focus was whether pleasantries affect output quality and how to elevate prompts to be more efficient, all while preserving tokens.
That said, what happens inside is worth exploring separately. Do you have an article to suggest?
I think I disagree. There was a trend going around where people asked chatGPT to draw a depiction of how they treated it; it would pretty clearly break into generating either a “pampered happy robot” or a “unhappy, overworked, bossed around robot”. And this break was basically just based on whether the person exercised politeness or not—at least, if you asked in a context-free window, it would draw pampered-robot if you asked politely and overworked-robot if you asked abruptly.
This no longer replicates—I tried it just now with context off and it generated the “pampered robot” for both the polite and abrupt requests. And I don’t know how applicable the image generation is to the chat model’s views. But it seems to me that if the model is modelling a persona in response to how it is trained, it may model “happy employee” when the person is saying please and thank you and “unhappy employee who needs to keep their boss happy” when the person is not doing so, even if its actual output may be near-identical.
That’s an interesting perspective. However, the internal persona the model adopts is a different area from the one I was exploring. My focus was whether pleasantries affect output quality and how to elevate prompts to be more efficient, all while preserving tokens.
That said, what happens inside is worth exploring separately. Do you have an article to suggest?