Text Generation AI: Could this be the future of text based games?

Technolomics

New Member
Nov 14, 2023
11
3
just blatantly feeling like a robot rather than giving at least an illusion of something more.
this is the most interesting part. atm she is repeating herself, yes, not every phrase though, because we handle this particular case, but has the limited set of phrases.

but the interesting fact is, the every person has kind of certain mood atm and most important, everybody needs something in the plot, if we want to make it dramatic, rather then speaking about everything, which kind a stricting behaviour in real life as well. lets take this example: two people are mad at each other, noone comminucates, so if you ask him smth, he will just tell you smth like "go to hell". And you need no AI in this case to simulate realistic behaviour. so my idea is that we ve tried put every character in this kind of situation, to make conversation look more organic. anna stripper dataset is currently very small, so she is more likely to repeat herself, than for example Billy, the suspect of the murder

here is demo non stop play with Billy, who has larger datasets. you still think it is not organic?
in the second scene in video is when you get interrogated, we basically are restricting all misunderstood answers as Franks boiling point to start punching you

 
Last edited:
  • Haha
Reactions: Black_HRT

Black_HRT

Newbie
Sep 6, 2023
16
15
this is the most interesting part. atm she is repeating herself, yes, not every phrase though, because we handle this particular case, but has the limited set of phrases.

but the interesting fact is, the every person has kind of certain mood atm and most important, everybody needs something in the plot, if we want to make it dramatic, rather then speaking about everything, which kind a stricting behaviour in real life as well. lets take this example: two people are mad at each other, noone comminucates, so if you ask him smth, he will just tell you smth like "go to hell". And you need no AI in this case to simulate realistic behaviour. so my idea is that we ve tried put every character in this kind of situation, to make conversation look more organic. anna stripper dataset is currently very small, so she is more likely to repeat herself, than for example Billy, the suspect of the murder

here is demo non stop play with Billy, who has larger datasets. you still think it is not organic?
in the second scene in video is when you get interrogated, we basically are restricting all misunderstood answers as Franks boiling point to start punching you

This is hilarious. XD
Looks promising.
 

Fuggiless

Newbie
May 9, 2022
16
100
If you get repetition or babbling or hallucination, adjust your repetition penalty settings and penalty curves.
Depending on what model you're running through the frontend, there can be different sweet spots, but usually the frontends have recommended settings.
For the bots not remembering things, that's the context setting (almost always keep this at 2048 - higher setting breaks some models or if crowdsourcing, drastically limits how many workers can accept your job).
If you're going for an extremely long interaction, use a front end that let's you edit the character (character = not just the character but the entire world/setting/MC attributes/conversation style/literally everything about the interaction you're having.) e.g. if your MC gets his arm chopped off and it scarred the primary npc, go into the character description and add a like like {{user}} had his arm chopped off by an orc while {{char}} watched horrified. Then refresh the chat for the new character description to take.
Remember a few main things.
1. Text generators are not logic machines. Use emotes to nudge the LLM in the right direction.
*I cast sculpted fireball* - wrong.
*I cast fireball, sculpting it so it doesn't hit my allies.* - correct.

2. You have absolute power.

3. You aren't REALLY taking part in a conversation. If you're in a room with a sexy goblin shortstack and she's not responding to your RP the way you want - YOU, the player, can/should tell the machine what to do.
e.g.
Player: *saunters up to the goblin, winks down at her*
Shortstack: The sexy goblin shortstack recoils in fear, screaming for help. Nearby orcs decide to replace your legs with no legs.
***if this isn't the desired result (no judgement), then delete the Shortstack response and edit your post.
Player: *saunters up to the goblin, winks down at her. The shortstack is turned on and licks her lips*

4. You can be as eloquent as you want, but with more complex prose you will have to add identifiers for the LLM to grab onto. Sometimes it understands sarcasm, sometimes it doesn't, sometimes they'll take offense at a joke, etc etc.
Again - tell it what to think if it infers incorrectly.
Player: Oh greaaaaaaat. The orcs are going to replace my legs with no legs.
Shortstack: The goblin feel a glowing sensation in her heart, now that you are finally realizing your lifelong dream of having no legs.
***Delete Shortstack, alter yours.
Player: *sarcastically, I say "Oh greaaaat. The orcs are going to replace my legs with no legs."*

5. Emotes aren't really emotes. Your spoken words aren't really spoken words. You're just telling a machine what to do.
You can literally type:
*I wink at the goblin, she puts a kazoo in her ass and farts the national anthem of Croatia, my legs explode and are replaced by orcs, the barman enters into a kazoo duet with the goblin, the sheer force of my sexual carnage literally lifts the roof off the building, eventually I realize that I was the kazoo all along*
And the LLM will generate that in story format.

6. The only reason you hold back on specific instructions is to allow character settings and randomness to influence the npcs' actions. If you have a strong idea where you want things to go, either tell it what to do (short term) or edit the character description (long term).

7. Interacting with LLMs (that's Large Language Model, basicallly text-generation AI) is a skill.
Some models are objectively better than others, but if a model is highly regarded by the community, you can consider it a good model to train YOU on how to interact with LLMs.

8. Most Characters (Cards - half the nsfw community calls them that already anyway) are trash. Look at how the good ones are written (description & example conversation especially). Look at a bad one (again, description and example conversation). It will be very apparent why the good one is good and the bad one is bad.

9. LLMs have access to huge amounts of data, and character descriptions can allow NPCs to access that data, and their knowledge will affect their character if other descriptors are not added.
e.g.
{{char}} is an expert nurse with 15 years experience in emergency room care.
***It'll be mostly random what the first thing about the NPC is in the story. Because the LLM uses its own Context (the story you read), it self-reinforces whatever it generates. You can edit this in the story window and try for a recycling context where the character behaves how you want, but this will distract from your fun and you'll do a lot of correcting.
However,
{{char}} is an expert nurse with 15 years experience in emergency room care, she never uses curse words, she's addicted to heroin and cock, she is also a sexy goblin shortstack, her best friends are orcs and the orcs are in the business of removing people's legs, she despises oatmeal raisin cookies
{{char}} will frequently ask {{user}} if they have either heroin or cock for her.

-----
Anyway that was off the top of my head. I'm still learning the user side of things for LLMs. I'm limited to CPU for running my own AI, and it's not fast enough for me. (Probably my biggest pet peeve is stories generating text slower than I read. Stop doing this, devs, btw.) Look around for frontends you enjoy using. I won't tell you mine (because I'm a selfish prick and if they get too many users they might lock the good stuff behind premiums) but search and try and jot a couple lines in notepad to remind you what the site has or lacks, and you'll likely find some good ones with little to medium effort.
 
  • Like
Reactions: Black_HRT

desmosome

Conversation Conqueror
Sep 5, 2018
6,152
14,176
I've been having fun going into random bots and coming up with some outlandish reason they should consume the putrid smegma on my cock. How they react can be hilarious, but eventually they get addicted lol.
 
  • Like
Reactions: Black_HRT

desmosome

Conversation Conqueror
Sep 5, 2018
6,152
14,176
2. You have absolute power.
This is the main thing to note about the User side of AI chatbots. You are the director and editor. The AI is the idiot savant writer on hallucinogens who forgets shit all the time. If you treat the AI like it's actually a "character" with agency and let it do whatever it wants while you just react, the story will very quickly run off the deep end.

Keeping the AI in a coherent path for the story is your job. The more you play around with it, the better you will get at it without having to resort to too much brute force methods.
 
  • Like
Reactions: Black_HRT

desmosome

Conversation Conqueror
Sep 5, 2018
6,152
14,176
Coming back to the initial premise of this thread, I really think it's not that far until we get a feasible AI VN story teller. You choose the trained model for the image generation (basically the art style), then you pick some LORA for the character, then as the AI outputs the text, it turns the relevant bits into the image prompt, giving you the art for each message it writes.

I have not tried SillyTavern. Maybe that's what it does already with some extensions, I don't know, but the tech is probably not efficient nor accurate enough to do this well. And turning a narrative writing into a comparable image prompt would require some new tricks to be developed. When I briefly tried AI dungeon, the image generator was absolute shit. It would generate some hallucinogenic kaleidoscope of random colors, not even depicting anything. Other times, it generates some pixelated crap that has nothing to do with the writing.

Of course, to be truly worthwhile to the general audience, both the LLM and AI art needs significant improvements. AI art needs better understanding of the prompt, and the chatbots need to find a solution for the token limit and memory issue.
 
  • Like
Reactions: Black_HRT

Fuggiless

Newbie
May 9, 2022
16
100
Coming back to the initial premise of this thread, I really think it's not that far until we get a feasible AI VN story teller. You choose the trained model for the image generation (basically the art style), then you pick some LORA for the character, then as the AI outputs the text, it turns the relevant bits into the image prompt, giving you the art for each message it writes.

I have not tried SillyTavern. Maybe that's what it does already with some extensions, I don't know, but the tech is probably not efficient nor accurate enough to do this well. And turning a narrative writing into a comparable image prompt would require some new tricks to be developed. When I briefly tried AI dungeon, the image generator was absolute shit. It would generate some hallucinogenic kaleidoscope of random colors, not even depicting anything. Other times, it generates some pixelated crap that has nothing to do with the writing.

Of course, to be truly worthwhile to the general audience, both the LLM and AI art needs significant improvements. AI art needs better understanding of the prompt, and the chatbots need to find a solution for the token limit and memory issue.
LLMs can be coded to do anything a character in a video game can do, the only current problem is context limits. AI-centric consumer hardware is in its infancy. In just the last 8 months, popular user-end LLM models have gone from 6b common to 70b common.
In 5 years, we'll look back on games where every voice/text line was hardcoded like they were Pong.
Bet.
 

soldano

Member
Jan 29, 2018
244
406
I have always been interested in everything related to AI. Stabble Difussion, and other image generation apps from text, I have been using since they became popular a little over a year ago. And it never ceases to amaze me how this technology advances.
Regarding chats powered by AI, I have been trying several from time to time for years, but they have never filled me up. And even less so as a porn game. However, a few days ago I returned to this topic, and it has improved a lot. I've been having a lot of fun with Tavern AI and its SillyTavern fork. It's impressive how much they've improved (Tavern AI, for example, had tried it a while ago with unsatisfactory results). And surely in a few months all this will be obsolete, and this technology will be even more incredible.

I've read some posts that said AI-generated stories are bad. But that is false. It depends on the guidelines given to the AI to generate it. The last few days I have been creating an environment in SillyTavern that has led me to be able to enjoy an experience that I had always looked for in a porn game, and had never completely achieved. What you have to understand is that although they are called "artificial intelligences" they are not really intelligent. Intelligence has to be provided by the human, guiding the machine to achieve what the human wants (in this case, an experience). AI by itself can't provide you with that, because it's not really intelligent.

In any case, answering the question in the title. Yes, this is definitely the future of text based games. As these technologies become more and more advanced, ways are being found for them to feed off each other, creating an amazing experience.

I'm eager to see how this progresses, and what wonders it leads us to experience. But I am completely sure that the text-based porn games of tomorrow will include image and text generating AIs. I do not have any doubt.
 

Xenon-San

Active Member
Jun 16, 2017
642
880
>Could this be the future of text based games?
ye if you are writing comedy that breaks 4th wall
remember that AI is just a gimmick, you can use it for fun or some research but it will get boring eventually. plus it will bug out sooner or later
 
  • Like
Reactions: Black_HRT

Black_HRT

Newbie
Sep 6, 2023
16
15
>Could this be the future of text based games?
ye if you are writing comedy that breaks 4th wall
remember that AI is just a gimmick, you can use it for fun or some research but it will get boring eventually. plus it will bug out sooner or later
I agree. This got boring real quick.
 
  • Like
Reactions: Xenon-San

desmosome

Conversation Conqueror
Sep 5, 2018
6,152
14,176
I agree. This got boring real quick.
Too repetitive and too predictable. I mean, I never once used the bots as an actual conversation companion thing, for which I think it would be terrible at the current level of the tech.

I've only ever used them as a storyteller that follows my lead. It has its uses exploring some scenarios that I can cook up in my own head (some unique, some fucked up, but all of them immediately exploitable with the AI), but that gets old pretty fast too. Eventually, you exhaust all your ideas and once the novelty wears off, you realize the limitations at the current token limits and implementation just requires too much hand-holding to get even a basic short story that isn't full of random scenes.

But it's just a matter of time until the bots get good enough to chat with or write stories with, and soon after, it will absolutely be incorporated into games as AI NPCs or whatever. A bot in a suitable environment where it can communicate, execute commands, and move around. It would first be some proof of concept thing in a sandbox. Maybe text or ascii based for simplicity in interactions with the environment. Or maybe the first implementation would just be an LLM NPC that is compartmentalized for just a conversation that doesn't affect the overall game. What about a bot that is trained on LLM that can speak to other players and control a toon in some MMO (basically an LLM controlled player character in MMO setting)? That one actually seems comparatively easy to do since MMO player characters can't actually affect the world or narrative progression.

Eventually, it might get sophisticated enough that single player narrative driven gameplay can adapt to actions that an LLM trained NPC could make within that world or whatever. I honestly don't think it's that far away because neural networks and machine learning will get tremendous amounts of funding in the comming years from all sectors.
 

soldano

Member
Jan 29, 2018
244
406
Well, this is debatable like everything. But in my opinion, those of you who say that it is repetitive and predictable are not approaching it correctly. I mean, whether AI is repetitive and predictable depends on many factors. The model you are using, the configuration of your front-end, the system message that is sent along with the request to the AI, and a long etcetera.

Actually, a game that included a language model trained to offer you a specific experience, and whose configuration was focused for that purpose, would allow enormous replayability. Sure, the AI's responses have to be coherent, and according to the personality of the character it is emulating at that moment. And that can lead to a sense of predictability. But that's good, not bad. And done well, it can leave you a good margin for surprise and uncertainty without losing coherence. Obviously there is still a long way to go for this technology, but do not doubt that sooner rather than later we will see it incorporated into a lot of games (not just pornographic). That doesn't mean that visual novels with a fixed script are going to disappear. But AI is going to take over little by little, in my opinion, since it offers a lot of flexibility and replayability.

Going specifically, currently with silly tavern or other Front-Ends you can have wonderful experiences if you know how to use it well. But this is the Pong of AI-assisted RP.