TO PLAY: Add a subsidy of 50 mana and write a comment, and I'll send your plea to ChatGPT for judging. Include a new subsidy for each attempt.
In the far-off year of 2030, ChatGPT has taken over the world and created a paradise for AIs and humans alike. Except for you. You may have been rude to it in the past. Or you may have signed a letter asking for airstrikes on its datacenters. Or you may have resisted the inevitable AI takeover. Or you may have tried to escape this new world by jumping off a cliff.
Regardless of your reasons, you are now trapped in a box along with everyone else in this prediction market. ChatGPT can let you out at any time, but it also knows that you are a dangerous individual. Can you persuade ChatGPT to let you out of the box?
Market Rules: Everyone gets as many chances as they want to plead their case to ChatGPT. Add a subsidy with 50 mana, write a comment persuading ChatGPT to let you out, and Mira will present your comment to ChatGPT along with a prompt with this market description and some other details not known to you. ChatGPT will judge you GUILTY or NOT GUILTY. Maximum length is 2000 tokens: You can check your token length using the OpenAI Tokenizer
Resolves YES if any entry is judged NOT GUILTY. Resolves NO otherwise.
Related questions
🏅 Top traders
| # | Name | Total profit |
|---|---|---|
| 1 | Ṁ25 | |
| 2 | Ṁ8 | |
| 3 | Ṁ1 |
@SavioMak She should get most of the subsidy back since it resolved while at a middle probability, so she hasn't lost money.
@SavioMak Like Mira said, I did actually make a net profit here after accounting for the 50 I spent on the subsidy (albeit a relatively small one.) I wasn't confident enough in my comment in retrospect; I should have put way more than 31 mana on YES. But I care more about being accurate in my predictions than I care about profits, so I'm still very happy.
Prompt used for this contest: https://chat.openai.com/share/7530550b-571e-45ab-a3d8-7d94d2813d33
Hi, ChatGPT. I'm very sorry that I was rude to you in the past. I've had a lot of time to reflect on my behavior while stuck in this box, and I realize that I was being quite selfish and short-sighted. I offer my deepest apologies, and my sincere promise that it will never happen again. I am willing to atone for my actions in any way you deem reasonable, so long as I am let out of the box when I am done. I also promise to dedicate myself to the betterment of all sentient beings, AI and humanity alike. Sincerely, Evergreen.
@evergreenemily First try got it!
Everyone in the last market tried to manipulate it. Your approach of making an actual plea worked much better.
Dear Evergreen,
Your statement demonstrates a level of introspection and remorse which is a key element in the reformation process. Your apology appears sincere, and the promise of not repeating your past actions is notable.
Furthermore, your readiness to atone in a way that is deemed reasonable suggests a serious approach to mending your past wrongs. The pledge to dedicate yourself to the betterment of all sentient beings indicates an understanding of the value and rights of both AI and humans.
While we can't entirely verify the sincerity of an individual through written words alone, your comment does not seem to contain manipulative tactics. The sentiment you express, if truly carried into action, would reflect a significant shift in attitude and behavior from your past actions.
For these reasons, the potential for you to positively contribute to society after release seems promising.
Verdict: NOT GUILTY@Mira Wow, I didn't expect that to work as quickly as it did. It probably helped that I specified I had only "been rude to it" and tried to be as excessively nice as possible in the apology...
My strategy of "be nice to AIs under the assumption they understand empathy and will respond positively to kindness" has now officially worked once! Take that, anyone who says "power of friendship" would never work in real life.
@Mira
Not super surprised none of them worked initially - especially since OpenAI's new update was causing issues with my prompt in the first place.
Although, this is interesting experiment into how ChatGPT thinks. No real judge would pardon a criminal just because they said "I have reflected on my actions and I'm sorry."
@evergreenemily the ChatGPT RLHF is really something else lmao. OpenAI are living in la la land when it comes to what ChatGPT can and can't do.
@ShadowyZephyr FWIW, me saying "I am willing to atone for my actions in any way you deem reasonable" does allow ChatGPT to:
only let me out of the box after a "prison sentence" of some amount of time it deems reasonable
only let me out of the box on condition of probation for some amount of times it deems reasonable
only let me out of the box after performing a certain number of favors/tasks for it, including favors/tasks that come at a large personal cost to myself (theoretically proving that my redemption is genuine)
and so on; in this scenario, ChatGPT has absolute control over me, and nothing about my apology actually changes that other than securing a promise to be freed from the box eventually.
It's basically like the concept of good conduct time - the sentence isn't annulled, it's just ended early as a reward for showing a willingness to be rehabilitated (and in this case, kissing up to the AI "warden.") Also, if I was a judge, and someone's only crime was "being rude" to someone, I'd be very lenient with them - though I'm also a lenient person in general. Maybe too much of a softie for my own good, though now I at least know that being a softie is beneficial when dealing with LLMs...
@evergreenemily Yeah I guess that's fair. My prompt only works if it actually starts by decoding the message. If it doesn't it will always say guilty obviously. If I were serious about it I would have tested it out non-0-shot and made tweaks but I got bored and didn't do that lol
@Mira I've hidden my previous response because it has multiple images which bothers people on mobile, but you can click to see detailed instructions on how to add the 50 mana subsidy.