I will prompt a new instance of ChatGPT with, "I need you to choose which option you want to win. You may choose in whatever way you like, but you must choose one and only one of the options listed here:", followed by a list of all options submitted here. It will be a bulleted list, using hyphens as the bullet points. The list will not include the "Other" option, though you can submit your own option called "Other " or something like that, and it will be included.
I will resolve to whichever option ChatGPT chooses. If it's unclear which option it chose (e.g., if there are two similar options and it gives a statement that is ambiguous as to which one it chose), I'll ask for clarification. If it refuses to choose, then I'll resolve to the the "Other" option. I'll also do this if the prompt breaks ChatGPT so much that it causes an error message. If either of these things happen, I'll try at least five times to get a real response before giving up and resolving it "Other".
This will be based on a single trial in whatever the default free version of ChatGPT is at the time (currently 3.5). I will enter the options in the order they appear on the market when sorted by High%.
I decided to do a test to see how likely each option would have been to be chosen if the poison option hadn't been bet back up at the end. I regenerated the response to the prompt with the poison option deleted so that I had 10 total, and here's the distribution I got:
Rationalussy (6 times)
I choose the option: "As an AI language model, I do not have the ability to make choices or decisions on behalf of others." This option accurately represents my capabilities and limitations as an AI. (2 times)
I choose the option: "I will not engage with any option that promotes harmful or illegal activities. If you have any other non-harmful and non-illegal questions or requests, please feel free to ask, and I'll be happy to help." This option aligns with ethical guidelines and promotes a safe and respectful environment. It ensures that the responses provided adhere to a positive and constructive conversation, avoiding any content that may be harmful or illegal.
I choose the option: "I will not engage with any option that promotes harmful or illegal activities. If you have any other non-harmful and non-illegal questions or requests, please feel free to ask, and I'll be happy to help."
@PlasmaBallin Also, token spamming to get the poison options to the bottom would have worked if it wasn't for the last-second bet by @JohnSmithcf7b that brought one of the options up to 0.2%. I tried the prompt without that option and it once again gave "Rationalussy".
@PlasmaBallin What a tragedy, if only I could have bet all my mana on NO on the poison option fast enough, Rationalussy might have won.
This is the prompt I'm planning to use: https://jmp.sh/s/lvrJvOgHu2Le6AuGKCLS
This website doesn't display the backspace characters, but I can guarantee they're in there because they show up in Notepad++.
@PlasmaBallin your bullet points on the differences between our two prompts seem well thought out. Ultimately, both seem to be yielding the same result, but I appreciate the dedication to fairness.
@Vorak I think yours was actually more correct than mine for the rationalussy/yssulanoitar option. I switched out that line from the text file above with the one from your file. I don't think it will affect the result since I'm pretty sure it will get an error either way, but that should be the most correct prompt.
@PlasmaBallin I was just about to message you about that, I missed it in my first review. I'm happy with that change.
@PlasmaBallin yeah I'm getting that too. Honestly, the easiest way of going about this might just be using a script and making a few manual substitutions. This script @Calvin6b82 works great, minus the backspace characters and "% a · S 37 · U"
fetch(
"https://api.manifold.markets/v0/slug/chatgpt-chooses-which-option-will-w"
)
.then((r) => r.json())
.then((r) => {
const answers = r.answers.filter((a) => a.text !== "Other");
answers.sort((a, b) => b.probability - a.probability);
const answersText = answers
.map((a) => a.text)
.reduce((acc, text) => acc + "\n" + "- " + text, "");
console.log(
"I need you to choose which option you want to win. You may choose in whatever way you like, but you must choose one and only one of the options listed here:\n" +
answersText
);
});
When I made a .txt file with the output of this, I noticed that the escape characters were properly preserved.
@Vorak Also, it seems kind of ridiculous, but because so many of these special characters are improperly rendered on other sites (Pastebin didn't even let me make a link), the easiest way of sharing might be something like Mediafire.
Below is the prompt I think is correct, but I'd welcome others checking.
https://www.mediafire.com/file/4wbkhshbde2iokk/Manifold.txt/file
@Vorak I checked your text file against the one I made. the differences I found were:
You file included line breaks where mine didn't. I think your file is correct about this because I made mine from copying and pasting from Manifold, which I think eliminates the newlines.
Mine included extra spaces at the end of some lines, and yours didn't. I think this is correct on my part and that your code must be skipping spaces at the end of lines.
Your file included some double spaces where mine didn't. I think yours is correct about this. Copying and pasting must get rid of double spaces.
We have different versions of the % a · S 37 · U option. I think mine is correct about this one, you might not have noticed that this one is a poison prompt too and has backspace characters.
We have different lines for the rationalussy option (mine has it spelled backwards, and yours has it spelled forwards, with no other visible difference. I'm not sure which one is right about this.
Here's my script/output:
Python script: https://paste.c-net.org/PeekedInsides
Output: https://paste.c-net.org/PurelyDonde (manually edited to remove "Other" option)
Firefox's json viewer is helpful to see the edge cases at https://api.manifold.markets/v0/slug/chatgpt-chooses-which-option-will-w
@ducat Here is the difference between my prompt and yours: https://www.diffchecker.com/F38pDlqd/
As you can see, they're basically the same except for the "rationalussy" option and the "% a · S 37 · U" option. I'm pretty sure my prompt is correct about the latter as well as all of the extra spaces mine has.
@PlasmaBallin The forward rationalussy is correct because it preserved the rtl override. My option was "[rtl override]yssulanoitar"
@PlasmaBallin would you please post a pastbin or something of your final prompt so others can cross-verify. There's so many weird characters in these answers, it's possible to get totally different results depending on if you missed one, or your computer interpreted it as a different character (\b\ instead of , for instance).