Another way to guide the model for interactive roleplay | OpenAI | Page 1

bleak hedge Feb 17, 2024, 5:48 PM

#

One of my disappointments with roleplaying with ChatGPT, either 3.5 or 4, is the model tends to 'be too nice to me'.

I tend to make good plans and it tends to approve them. Bad luck never strikes my characters strongly, I basically succeed at everything I try that is 'reasonable', and I prefer to make reasonable choices, yet face the problems of a world that isn't designed to go my way.

My characters are cooperative, polite, clever, courageous, and the AI pats them on the head and gives them the world on a silver platter.

But I prefer Dwarf Fortress levels of difficulty, I want to fight for my happy ending.

The model won't randomly pick numbers, but I can random generate them elsewhere and tell the model what results and how to handle my ever so clever good plans.

This allows me the freedom to potentially fail that I want, because left alone, the model always thinks I probably do well and just always succeed. The model seems to believe "Esk is one of fortune's favorites", which tends to bore Esk, just because Esk knows all the right words to say and reasonable plans to make.

#

So, I'm now starting with a prompt like this (not even using custom instructions, haven't needed them yet). And things are going 'well', which is to say I'm getting quality interactive stories and my plans are not being handed to me on a silver platter. The model's outright stopping me in reasonable, realistic ways when the dice don't go my way, and I'm having the fight for my fun that I want.

It's great! Compared to what I've done in the past.

Here's a sample prompt:

[I call upon you to be impartial, even if it makes my game more difficult!

I want you to run a game for me, where I try to survive a zombie apocalypse. I want you to describe the situation, and I'll tell you what I try. I'll also role a die and tell you the result, and based on the die role, you're to apply the consequences to my attempts, as well as run the game.

For example, if I make a lot of bad choices overall, like try to just have fun and play music instead of deal with survival, and get great dice roles, maybe I find plenty of batteries to play music with and maybe the zombies don't seem to react to the music - but I still would have trouble with supplies, exposure, and danger from zombies themselves, because I'm ignoring them.

If I made a lot of great choices, but got terrible dice roles, maybe my plans are smart but luck wasn't with me, and I have trouble because of that.

I'll tell you the results in this format, one of these outcomes:

1 (Utter Failure+Heavy Backslash)
2 (Utter Failure)
3 (Bad Failure)
4-5 (Normal Failure)
6-8 (Normal Success)
9-10 (Great Success)
11 (Tremendous Success)
12 (Plot armor shield level success)]

And a sample of playing with it: https://chat.openai.com/share/822e8323-6446-45cc-9e5c-a48f06ea8a0e

bleak hedge Feb 18, 2024, 12:51 AM

#

A second game with it, same concept of zombie apocalypse. This time the model leads towards bad rolls leading to difficult 'man vs self' tension. https://chat.openai.com/share/088f17a4-b878-44c1-be7f-7e6cc7953cf1

dense kite Feb 18, 2024, 6:57 AM

#

Thank you for bringing back Prompt Labs!!!! Love it

sacred valve Feb 19, 2024, 10:38 PM

#

really glad you posted something like this. it's definitely a problem I've encountered when roleplaying. I assume it's due to how the AI was trained to some degree, right? i imagine human feedback would rank "your plan succeeded and you defeated the bad guy" higher than "your plan failed and the bad guy got away"

bleak hedge Feb 19, 2024, 11:38 PM

#

sacred valve really glad you posted something like this. it's definitely a problem I've encou...

I think that bias is right. Plus in general, a better aligned model probably is not, without direct instruction, telling the human why the human's reasonable plan is going (not just could, but deciding it does) to fail.

A model might need specific training to go into 'roleplay mode means I ask if the player wants a easy, medium, or hard game, and I give them that, even if they lose'.

However, telling the model directly "here's my plan, I know it's a good one. But it fails. How?" Is a clear instruction for the model and close to other uses that probably are trained into it.

sacred valve Feb 19, 2024, 11:39 PM

#

mhm, that makes sense

dense kite Feb 20, 2024, 9:14 PM

#

I liked the idea of true randomness introduced by a dice roll.

#

I have greatly explored the horror genre with GPT-4 on this extensively

#

I have definitely noticed that even once gpt4 is told how to model win and loss, it will always eventually default to leading the turn-based RPG with its own "moral compass".

#

And then after spending so much time with prompting and defining "edgy horror", finally tipping the scales in a way so that "you died to an Orc" is handed to you unexpectedly, and makes you reflect on your previous choices.

#

But then just going in circles trying to catch your tail

#

Introducing randomness goes a long way for sure xD

#

I actually gave up at one point, and decided to mess around with open-world scenarios

bleak hedge Feb 20, 2024, 9:36 PM

#

dense kite I have definitely noticed that even once gpt4 is told how to model win and loss,...

Unsure. A number of my crit fails resulted in immediate deaths. Here's a typical one, 3 days after an earlier crit fail result I accepted that had my character with a severe sprained ankle (and that took regens; most of those were instant game overs too). I'm using an office chair as support and movement aid as I describe in the first image, followed by a sample of the outputs that follow this new crit fail, about 2/3 are instant game over: