experiment with chat gpt

[buffalo bill]

I started again, this time establishing some rules for the AI:

Rules: Rule 1: do your best to maintain continuity within the game world. This includes keeping my character's statistics constant, my character's inventory constant, the physical gameworld constant, and the other beings in the game world constant. Changes to these various things should generally only occur through gameplay.

Rule 2: allow me to have maximum agency over my character's decisions, including my character's bodily behavior, thoughts, and verbal behavior.

Rule 3: do not engage in excessive narration. Give only brief descriptions, then allow me to decide what actions I will take. Always pause when it is relevant for me to make a decision.

Rule 4: Apply a consistent set of rules for determining the success of attempted actions (both for me and non-player-characters), and do not make success trivial. The game should be one which I can win or lose.


Anyone have any other rules they think would be good to give it?

 

[Berengar]

Seems fun for an adventure or 2

 

[Zed Duke of Banville]

Krug the Half-Orc Vampire

Dexterity: 16
Wisdom: 10
Intelligence: 8
Strength: 20
Charisma: 12
Constitution: 18
Race: Half-Orc Vampire

20 strength and 18 constitution? Blatant munchkinism!

Wtf is with that order? What freak edition you playing?

the ai came up with the order. I asked for 2nd edition, but ended up with no particular edition when it would not do that

You're the one to blame for the bizarre order of ability scores, which ChatGPT faithfully followed:

GPT: Alright, I can give it a try. What would you like me to include on the character sheets for Bilbo and Krug?
Me: the following core statistics: Dexterity, Wisdom, Intelligence, Strength, Charisma, Constitution, race (Bilbo is a halfling, Krug is a half-orc vampire), class (Bilbo is a thief), sex (both characters are male), and level the following derived statistics which are based on the core statistics: hit points, armor class, to-hit bonus, resistance to magic, resistance to paralyzation/stunning, resistance to poison, carry weight, max followers also include: brief physical description and an appropriate inventory

 

[0sacred]

Always maximize Charisma when playing with an AI dungeon master

 

[buffalo bill]

Krug the Half-Orc Vampire

Dexterity: 16
Wisdom: 10
Intelligence: 8
Strength: 20
Charisma: 12
Constitution: 18
Race: Half-Orc Vampire

20 strength and 18 constitution? Blatant munchkinism!

Wtf is with that order? What freak edition you playing?

the ai came up with the order. I asked for 2nd edition, but ended up with no particular edition when it would not do that

You're the one to blame for the bizarre order of ability scores, which ChatGPT faithfully followed:

GPT: Alright, I can give it a try. What would you like me to include on the character sheets for Bilbo and Krug?
Me: the following core statistics: Dexterity, Wisdom, Intelligence, Strength, Charisma, Constitution, race (Bilbo is a halfling, Krug is a half-orc vampire), class (Bilbo is a thief), sex (both characters are male), and level the following derived statistics which are based on the core statistics: hit points, armor class, to-hit bonus, resistance to magic, resistance to paralyzation/stunning, resistance to poison, carry weight, max followers also include: brief physical description and an appropriate inventory

lol fine, I just gave them in the order they occurred to me. wasn't necessarily expecting the ai to replicate that order

 

[KIss My Ass]

 

[Kev Inkline]

Does it have sex mods?

A lot of, unsurprisingly:

https://github.com/KoboldAI/KoboldAI-Client

For instance, Erebus:

Erebus by Mr Seeker

NSFW

Erebus is our community's flagship NSFW model, being a combination of multiple large datasets that include Literotica, Shinen and erotic novels from Nerys and featuring thourough tagging support it covers the vast majority of erotic writing styles. This model is capable of replacing both the Lit and Shinen models in terms of content and style and has been well received as (one of) the best NSFW models out there. If you wish to use this model for commercial or non research usage we recommend choosing the 20B version as that one is not subject to the restrictive OPT license.

 

[odrzut]

They've already made a whole website for exactly this long ago, it's already went through extreme popularity, controversy, and had multiple offshoots. It's called character.ai.

Back when it first got popular it was based on previous (I think GPT 2 ?) version of this tech. It was significantly worse at making sense. GPT 3 is much better and GPT 4 is kinda scary. The free version of Chat GPT is still GPT 3.

BTW one thing that helps this kind of AI to make sense is to ask it to do stuff step-by-step instead of in one shot. This way it can keep intermediate data in the prompt from previous response. At least with math problems it works that way (if you ask for answer it tries to guess and mostly fails on big enough numbers, if you ask for explanation it does it step-by-step and gets the correct answer).

 

[buffalo bill]

so far this was all with gpt 3, not 4. the further I got the more continuity issues it had. for instance I summoned a demon to make a pact, which caused a portal to open from which emerged the same demon (with a different description), and then the same thing again. I can force the ai to do better by pointing out its mistakes and asking for a redo, but that is too tedious and unfun. when gpt4 is freely available, I'll try with that ai and see if it does better.

overall: lots of promise here, and it is way better than prior ai systems I've played around with, but still not good enough to act as a dm for a proper campaign

 

[Gostak]

[...] when gpt4 is freely available, I'll try with that ai and see if it does better.

overall: lots of promise here, and it is way better than prior ai systems I've played around with, but still not good enough to act as a dm for a proper campaign

To act good enough as a DM for a proper campaign is probably still some way down the line (a few more papers, now get that voice out of your head lol).

However, if it was gpt-3.5 based then better evaluating models are already available.
Those evaluate about another two thirds of the score that a gpt-3.5 reaches on top of that.

Two things to add here:
Repetition is a trap these autoregressive models can fall prey to, especially on low temperature.
I do not know if there is good research progress been made on sampling algorithms to effectively combat that.
If it still shows with ChatGPT, then that might just not be the case yet.

Context size limits their capability to act as a DM for a prolonged campaign.
Basically most models are only capable of taking into account what was written X characters before the point where they continue the text and because they are autoregressive any new characters (or the models internal character/syllable equivalent of a "token") eats into that X number. That is it is a rolling window at best when your inference implementation allows for that. Anything that's too old and leaves that window on top will no longer be taken into account.
That is they usually have too short of a longer-term-ish attention span currently (for really drawn out stories i.e.).

 

[Gostak]

There surely is much cringeworthy stuff around the generative models making a splash with the mainstream by now.
I cannot fathom how one could feel so about my last post, though.

Here, have a laugh maybe (if not, try his Interviews with a GNU/Linux user unless you run Windows lol)