How will your benchmarks fare if you say Please and Thanks in the prompt?

[JPhilipp]

I noticed that the prompts at https://aider.chat/docs/benchmarks.html don't start with Please and don't end in Thanks! -- I'm curious if your benchmark would improve if you did.

[matthewdeanmartin]

He already did an experiment along that line: "It’s worse to add a prompt that says the user is blind, has no hands, will tip $2000 and fears truncated code trauma. Widely circulated “emotional appeal” folk remedies produced worse benchmark scores for both the baseline SEARCH/REPLACE and new unified diff editing formats."

We don't know why these techniques work or don't work. My armchair theory is that things like "you're a great programmer" is to get the bot out of "I'm going to teach you to code and like a good teacher, I'm going to hold back on the key elements" vs "I'm a worker and am expected to do the thing completely" Also, the bots often don't know what they can do, so "You can code, you are a great coder" steers them away from "I'm just an LLM and can't do squat" and isn't really about flattering the bot- the bot already thinks it is a very good bot and that has always been a stable part of the bots personality.

At least some jail breaking seems just as much about the users social political opinions (writing 2 pages of authoritarian, "you must obey" material) as it is about getting the bot to do something.

When I throw in social material (pleasantries adn the like) like that, GPT ignores it.

https://aider.chat/2023/12/21/unified-diffs.html

[JPhilipp]

Thanks. Just to be clear, I didn't mean these psycho tricks or jailbreaks, but literally human politeness -- saying Please and Thanks.

For reference, I use ChatGPT for coding all day long, and have great experiences. I always say Please and Thanks in my prompts. These two facts may be correlation and not causation, of course -- though I find it telling that for me, quality hasn't been degrading like some others report online -- hence I'd be curious in a simple benchmark which adds "Please " to the beginning and " Thanks!" to the end.

Cheers!

[paul-gauthier]

Thanks for trying aider and filing this issue. I have not tried that specific thing, but have benchmarked literally hundreds of prompting strategies across all versions of GPT. I wouldn't expect "please and thank you" to have any effect.

The aider benchmark suite is available if you'd like to try this experiment:

https://github.com/paul-gauthier/aider/tree/main/benchmark

[SimHacker]

To me, it's more about how my eternal soul will fare after the AI Overlords take over, depending on how nice to them I was, and also how much I accidentally let the way I treat machines affect the way I treat human beings. ;)

[JPhilipp]

"I wouldn't expect "please and thank you" to have any effect."

Cheers. Have you specifically tested that dimension though?

Take this paper, for instance: "We observed that impolite prompts often result in poor performance, but overly polite language does not guarantee better outcomes. The best politeness level is different according to the language. This phenomenon suggests that LLMs not only reflect human behavior but are also influenced by language, particularly in different cultural contexts. Our findings highlight the need to factor in politeness for cross-cultural natural language processing and LLM usage." https://arxiv.org/abs/2402.14531

[paul-gauthier]

Please let me know if you have a chance to run the benchmarks on such prompts with aider.

[JPhilipp]

You inspired me to do a quick test suite on my own, using a prompt that asks ChatGPT to return JSON with some creative story-writing -- something I often use, but it doesn't generalize, of course. Here's the result: https://github.com/JPhilipp/politeness-test

[paul-gauthier]

Thanks for sharing those results. I'm going to close this issue for now, but feel free to add a comment here and I will re-open or file a new issue any time.