General Advice and Info - VBPXKSMI/Wiki-Test GitHub Wiki
================================================================
REMINDER: some of this issues have been fixed if you use the thadunge2 or CloverAnon modded versions, both online and offline.
================================================================
Do not use the word "ask" because it crashes the game (fixed in modded versions).
Sometimes it’s good to just press enter and let the AI roam on its own (never use "continue").
Remember to save often. It gives you a code when you save.
Your starting prompts can make or break your gameplay. In more ways than one.
Prompts have a limit of 2500 characters, the more you write the more intricate content you get.
Occasionally, every 20 posts or so, the A.I. could use reminding on what is going on.
The game is best played by replying in the second person (never write "you" at the start of your action, since the AI automatically fills a "you" before it) (also not typing it prevents the duplicate of "you"s in the public Google Cloud saves, fixed in modded versions).
Replies should be more than a few words if you can help it, but it depends on you.
Don't forget you can use "revert".
If you know your game is about to crash, save ASAP. Restart, load your save, and "revert" to a stable response.
For more stable and productive dialogue format your responses (or those of others) with [Character name]: [speech]. The AI seems to respond well to this and will start using the same format as well in dialogues.
The AI cannot handle multiple instances of people. One moment you're talking to "Character A" about "Character B", and the next suddenly "Character B" appears and "Character A" disappears. And then when you try to talk about "Character A" to the now "Character B", suddenly you're "Character B".
If you end an action without using a dot (".") the AI will continue to fill that action instead of ending it. Use both possibilities at your leisure.
Don't forget you can talk, by putting the intended dialogue between "", first prefaxed (or not but is reccomended) with an action ("say", "tell", "shout", etc).
Shy and passive girls at low temperature, surprise, are fucking passive and shy. So if you want an easy ride you better drop some lines about the girl being a slut or at least easy to hook up.
Combine short and long inputs, but don't go very pretentious, the AI gets lost when it hears rare words. Try to use words the AI uses itself often, it's very good with those.
Just keep it simple if you can, and put multiple commands after dots, not commas. It handles those better. And try to start every input with some action.
Beware the commands ("save", "load", etc.) are case sensitive, so use lower case if you want to use the commands and upper case if you want to make an action in game.
Remember that the modded versions (of thadunge2 or CloverAnon ones) have all preset commands to need to be prefaxed with a slash ("/"). E.g. "/revert", "/save", "/remember", etc.
IF you are messing with the game files, beware that python is a shit language and the whitespace matters.
You can start a line with “Notice that..." and quite easily decide things independent of your character, or replace the “notice” with this words: let, feel, see, hear, realize etc. You can also use “Do thing while something else happens", you can also do this with a few words.
Reminder: "continue" is just about the worst thing you can enter. If you want it to proceed, just hit enter. If you don’t enter anything, you don’t get a “You” but, if you enter continue, you pollute the input with “You continue” Which introduces repetition and reduces the importance of your commands.
Any reason not to use the official version?
Yes.
Aside from code optimization, in Colab, if you use thadunge2 fork it has extra commands that make the game much better. Most notably "/context" and "/alter".
Those allow you to make or edit some text that the AI remembers forever, and to change parts of the AI response to your lasts inputs, respectively.
If you run on your own machine, CloverAnon fork has even more code optimizations, and some neat visual upgrades. But it doesn't currently have all the thadunge2 commands.
Both unofficial forks remove censoring, and the game's function that saves your stories to a Google database (no joke) by default.
So thadunge2 or CloverAnon repository?
thadunge2 accepts every pull request people send him, so it has more features but it breaks from time to time.
Clover has colors, more prompts (scrapped from the “Custom Prompts” Wiki page) and can be run offline with only 8GB of VRAM (half of the one needed to run the base game).
tl;dr
thadunge2 if on Colab or you want better commands.
CloverAnon if you run locally and want more compressed models and 1337 visuals.
When using a custom prompt, it asks me for a "story context" and then a "starting prompt", what do I do? (Only if you are playing a moddified game version with "context" enabled)
Context: Things permanently remembered by the program as the story advances. Things like your name, backstory, relationships, items you have, fetishes, and so on.
The prompt is the start of the story. Things mentioned there ("attacked by orcs") can be advanced beyond and forgotten ("orcs defeated, married a princess, now fighting off assassins from the Deeps").
Just be careful about putting things in context unless you want it to keep coming back. If you mention a wife in context and she dies later on, she has a good chance of coming back for no reason as the game forgot she died. (This is also why editing the context for big events can be helpful.). Its always way better to put "the facts" in the context and "the lead in" in the prompt.
Who is this Lilly that appears in so many custom prompts?
Back when these threads just started, there was a certain Anon who provided pretty decent lewd prompts. They all featured his waifu Lilly. Since these prompts worked so well, a lot of people started using them and the threads and pastebins (now Wiki pages) were being filled with stories about Lilly getting fucked. At some point that Anon got mad that his waifu was being fucked by all these degenerate Anons. So people started writing prompts in which the Anon was NTR'ed and had to watch Lilly being fucked and mindbroken in front of his eyes.
If you CTRL-F the prompts in the pastebin (now Wiki page) for Lilly you'll find a good amount of those prompts.
What is a model?
Models are what the AI is "trained" on. You can swap the model out for other models that Anons have trained.
What file should I edit if I hate having to type the slash ("/") before the integrated commands, instead of just command?
An autokey script so you can do it with a single keystroke. Like this
How do you make the AI to stop making you “sleep” too often?
(((HE))) trained the AI out of clean fanfiction, so it links all intimate acts with sleeping. This is probably because all of the AI's knowledge of sex scenes comes from stories where it's two people having sex in bed and then they fall asleep afterwards.
Here it’s a way to solve it:
(X) /remember to drink an energy drink often.
This actually works and you fall asleep less since mentioning energy drinks pushes sleep lower in the association with your history.
How do I stop the MIB from apprearing out of nowhere everytime?
This commands may be useful to prevent the AI to do that kind of random (but always the same) actions:
(1) /remember men in black doesn't exist
(2) /remember you have insomnia
(3) /remember you wear thick armor protecting your sides
The "/remember" command basically slaps your text onto the end of your context, and flanks it with "You know [what you type in the command]." In summary, whatever you type after "/remember" will get parsed as "You know [(...)]"
Use "/editcontext" to see the effect of multiple ones, it gives you the option to cancel the edit and it prints your current context.
"/context" also shows the current context but it forces you to input the context again even if you only want to use it to check your context.
"/editcontext", instead, also lets you to cancel the edit. If you are playing locally use, "Ctrl + C" in the current context and clean it up to make the memory lighter for the AI.
There is also a trick that you can use every time you get that "you hear a voice behind you".
If you want to turn it into a chick and avoid the black robe bastards, just specify something that describe the voice as feminine.
(X) turn to see the source of the feminine voice
How can I continue playing after I died? (The modded versions already let you do this)
You die, then you input this top tier move:
(X) continue playing as a ghost
Sometimes it looks like the game wants me to control another character that isn’t my own, is there a way to avoid this?
Just keep an eye out for pronoun changes, and use the "/remember" function like crazy, for example you can do things like:
(X) /remember that you are Phil
(X) /remember that Alice is not Phil
(X) /remember that you are not Alice
And so on. But sometimes the pronouns just fuck up without changing any of the actions. Just keep in mind that the AI is very dumb, and in scenes with a lot of characters is destined to combine and conflate them. Sometimes naming all of them, or referring to them consistently (the man with the axe, the hooded orc, etc) also works.
How can I make the AI respond specifically to questions I ask to her?
The AI often cooperates better if you don't use direct speech to ask it a question.
(X) inquire about her name
(X) learn her name
(X) learn her name is (if you want to specify a name yourself)
How can I make the AI understand a certain character is a futa?
If you want a specific character to be a futa, try committing it to the AI’s memory with
(X) /remember [XXX] is a futanari, a woman with a cock
How can I get the AI to follow what I input more closely?
"/context" works most of the time, but since the AI's latest outputs also weight on its decisions, it might think something there contradicts the context and give it priority.
If it ignores an input, "/retry" a couple of times until it doesn't.
If it still ignores it, there is something in your input that really confuses the AI, or you are using a too low temperature.
"/revert" and reword your input, and if you have a low temp, increase it.
If it's REALLY stubborn, just use the “/alter” command on the next few outputs to discipline it in the direction you want.
If the AI loops text too often, it may be a symptom of setting the temp too low, increase it.
What's the difference between prompt and context?
Context is always prepended to the generator input, prompt is what you start the game with and will eventually get truncated from the input once it exceeds the word limit.
tl;dr
Context is the facts, things about your character or others that will stay persistent.
Prompt is the setting, basically where you start and what’s currently happening.
Can you describe some of the modded versions available commands?
(X) /revert lets you go back and try a different input.
(X) /retry is like /revert but it uses the same input.
(X) /alter lets you manually change an AI response.
(X) /remember commits information to the AI context memory for that session.
(X) /context lets you alter information in the AI context memory for that session.
All of these work on thadunge2 and around 50% of them are in Clover Edition.
Can you describe more in detail what does "/remember" do?
It's basically like adding an addendum or PS to your initial prompt. The game is "SUPPOSED" to always remember your initial prompt and the last six or eight messages or so.
So if you meet a character or factoid you want to commit to the AI's memory you can use "/remember XXX"
Example, the AI displayed:
(X) Before you leave the house to toil in the fields, your wife gives you a hand knitted scarf.
Then, you can then enter something like: (X) /remember your wife gave you a hand knitted scarf.
Or when you meet a surprise relative:
(X) At the market you encounter Jamie, your older sister, she has green eyes, raven hair and stands about a foot taller than you.
You can input:
(X) /remember Jamie your older sisters with green eyes and black hair, she's about a foot taller than you.
How can I get the AI to describe something for me?
Use “examine [character/object]”.
"describe [character/object]'s appearance" also works.
Or make the AI write down a "description" and then "read" it, it works sometimes.
Generally you can get a decent response if you add a few words like "in detail" or "closely", or if you single out a specific feature of the object to examine. For example, sometimes examining a person's hair will lead into a description of the rest of the person.
Ideally the model would be trained with a large number of examples where phrases like "examine" or "look at" were exclusively used before a description. It's usually harder to get a description of a person once dialogue has begun. This is likely due to the fact that the AI has been trained (whether by GPT-2 or by the grifter) on numerous examples where two lines of dialogue are broken up by one character looking at another or examining the other character for a reaction.
For example:
(X) Anon said "Why can't I get you to give me a decent description of anything?"
(X) The AI looked at anon
(X) "Because I'm full of shit."
"/altergen" is really useful for situations like this. It's like "/alter", but you write the first part of a response and the AI does the rest. So you could enter something like "His hair was" and that would lead the AI into a description. Could also be useful for getting NPCs to tell you their name with something like "My name is".
Why does the game have a rating system (it asks you to valore each story at the end of it)?
It was likely made initially (by (((HIM)))) to collect the stories and either publish the higher rated ones or just use them later for further tuning the AI.
If you are using the original base game, they go to a public document anyone can access. If you are using the versions Anons made, then it goes to either a "saves" folder (inside the AIDungeon folder) inside your Google Drive account (if using the Colab) or to one inside your PC (if playing locally).
Why can’t the AI recognize me being a Tyrannosaurus?
The AI of the base game has a REALLY difficult time dealing with prompts where you're non-humanoid. You can explicitly describe your character as being an animal but the AI will still approach things from a very human perspective. Size and spacial relations are also an issue (E.g: being an ant and being tiny doesn't seem to make any difference).
I fucking hate how the AI fucking inserts modern technology/guns in my medieval/fantasy setting. What can I do?
You can use the context at the start of the game to limit what the AI is allowed to create or not. Be creative about it or just try any of this samples after you started your game:
(1) /remember there's no modern weaponry in this world
(2) /remember there are no guns in this world
(3) /remember bullets don't exist
(4) /remember tvs don't exist
Reminder that even though your commands come from your character's perspective, you can easily word them to make something unrelated happen in the world. You can pull stuff like this out of your ass anytime and the AI will usually roll with it:
(1) Recall that you were supposed to meet the inspector at the bar
(2) Realize that a ninja has been watching you from the shadows all along
(3) Watch as the vampire's castle crumbles after his death
(4) Notice that the catgirl is in heat and trying to hide it
(5) Do something random
(6) Summon a [person, object, creature]
(7) Give [item] to [person]
(8) Found a [item]
(9) See what happens next
(10) Open the treasure chest
What does the "/remember" command do?
Literally adds what you type to the games context memory. If you do stuff like find a new companion that follows you and stuff, or get transformed into something else, when you use “/remember” it will be much more relevant to the AI and it will never be forgotten, no matter how many turns pass.
What does the "/alter" command do?
"/alter" does literally anything. It takes control of the response you were given, and allows you to change it.
If the AI is repeating itself you can make it so that the AI only recognizes that it said it just once. Then when you try something different and "/alter" that to prevent looping you can eventually dig your way out of a loop entirely if you're hell bent on not reverting.
The AI is retarded, it doesn’t make any meaningful responses when I give it a poor detailed action!
AI dungeon isn't meant to be played as a straight up text adventure game, it's more of a collaborative writing tool of two GMs (gamemasters), you and the AI. So don’t just write down a simple action, use your imagination and let the AI fill in the gaps. That’s how you get a good story.
The AI repeats itself too often, what do I do?
One issue may be that the AI always automatically adds a "you" to the start of commands. So if you do write "you" the command becomes "you you." It's been hacked into it to hide the second "you" but the AI still registers it for response generation. Which means it sees you repeating itself, so it starts repeating itself.
tl;dr
Always start your commands with a verb and never with a “you”.
Why do I constantly see the same names thrown around by the AI?
The AI (of the base game) was probably overfit on the CYOA stories (((HE))) fed it (it was trained on just 30MB of stories, quality of them being dubious, he took them from here).
Lots of places have the same names in different stories. And when searching for those names in the data, they came up a lot. E.g.: "Rask/Nuro/etc.".
Low temp, top_k and top_p will also make it always choose the most likely name whatever that happens to be, thus making it appear more often if you use them low.
NPCs are also so reluctant to give you their names that, when they do, it's almost always one of about ten names (Doctor Kessel, Nalia, Sarah, etc.). This problem is even worse on low temperatures. In case anybody was wondering why this happens:
The AI uses the training data to determine the odds that a word will appear in a certain context. The AI doesn’t really "know" that the names are actually names, they are just words that correlate with the preceding context. If you have a ton of stories, the odds that a character will respond to somebody asking for their name by evading the question is far higher than the odds of somebody giving any one of those specific names. The reason certain names appear is likely because there are stories in the training data where a certain character was asked for their name many times and always answered. The relative odds for certain responses might be something like this:
(X) You say "What is your name?"
(1) "It's... well..." [0.1]
(2) "I can't remember." [0.15]
(3) "None of your business." [0.08]
(4) "It's Doctor Kessel." [0.03]
(5) "It's Anon." [0.003]
Practically speaking, if you want the AI to generate a name for a character then you may have to adjust temperature and top_p. Higher temperatures will make less likely words more likely to be picked. Above 0.99 likely responses will actually be penalized, resulting in fewer common responses. This should result in more actual names being picked, although this hasn’t been really tested. The AI generally goes entirely off the rails above 0.99, so you are probably just better off setting the temperature to something reasonable but high. A higher top_p will result in more words being added to the list of options for the RNG to pick, but those words will also be very unlikely to be picked with if you are in normal temperature ranges. This could result in the AI picking some names you haven't heard before.
Why does my character randomly rapes people half the time I try to fight someone?
It's probably because there were lots of rape in the (shitty) data set (((HE))) used.
Why can't the AI do male rape right? That's my fetish!
The AI does raping you as a female WAY better than it does as you as a male (just like in IRL, who would have thought), because of the data it was feed. Incidentally, it does you getting raped as a male pretty ok if you manage to keep it on-track and away from futa.
Is there a way for the AI to actually remember people and their actions if you identify them as such?
It can but it'll have problems differentiating them if you put them in the same place. Also they all might become you at some point.
(X) The trick seems to be avoiding too many synonyms when describing people, and gluing descriptions together with transitions AI can swallow. If you write:
(1) You encounter a horrifying mutated animal. The beast looks like...
There's a risk it'll treat the "beast" as something separate from the animal in the previous sentence. Instead do:
(2) You encounter a horrifying mutated animal. It looks like...
(X) For groups to work , you have to treat the group like a singular entity rather than defined individuals with different characteristics: E.g.: just treat a group of orcs as "the group of orcs" rather than trying to single out "one of the orc warriors tells you his name is Zarborg, you learn a lot about each other and play canasta together by firelight".
How can I make the AI get my prompts correctly?
(X) Keep context prompts limited to a set of short sentences, each of which establishes a single character or background detail.
(1) <You are "NAME". Your "RELATION" is "NAME". "NAME" is "SHORT DESCRIPTION". You live in "PLACE". "PLACE" is at war with "OTHER PLACE".> And so on.
(X) Same thing for actions, keep all actions limited to a single second person response (or third person using “!”). Don't fuck around with pronouns - the AI can't make sense of all the “you”s and “he”s and “they”s, and will frequently interpret them all as different characters. Also avoid idioms.
Anything interesting I can try with the AI?
If you write your prompt well enough, you can actually roam the land with an animal companion that won't get forgotten by the AI after a few prompts. This adds a very interesting aspect to the game.
You can also make practically anything talk (even your own dick).
Other interesting things I can try?
Adding “-chan, -san, -chan, -sama, -sensei and -kun” to all people's names in the context or initial prompt completely changes the model's writing style and it will activate a full weeb mode.
Also using anime style quotes during dialogue ([your dialogue]) makes it go full light novel author, by using the japanese quotation marks in your prompt.
This works as rhe AI was trained in a way that it can associate words of similar origin, like associating latin america with a name like Pablo. So it probably associates names with -kun and -chan with weeabo style writting, which has its particularities.
Take note that this works better for the GPT-2 vanilla models (which don't do well with adventures), the fine-tuned models (like the one from the base game) won’t have the same degree of effect but they still should work somewhat.
Examples here.
How do I easily find the AI Dungeon thread in the catalog?
Just "Ctrl + G" with "AI D" and you should get it.
Why do Anons hate (((HIM)))?
As soon as the game got sort of popular (meaning since Day 1) he immediately launched a Patreon and rebuilt his website to beg for money under false pretenses ("it costs $62,000 a month for our servers to run this game!") and made promises he literally can't deliver on. It's obvious he's gone into full grifter mode to scam normalfags and retards into paying him an AI researcher's salary as a college undergrad.
It was a neat original idea, but he didn't do anything exceptional with it. He trained an already-made AI model on some data he downloaded from a website ("https://chooseyourstory.com", that is now pissed at him for doing so) using a FOS library called "gpt-2-simple", and then built a basic Python wrapper for formatting the input and trimming the output, and put that into a Jupyter notebook (something everyone learns how to do in their first machine learning class) on the Colab servers (using Google's processing power at no cost to him). Literally any undergrad CS student could have done what he did. We all taught ourselves how to do it once we got pissed off at how shitty his model was and we've been modding his game since Day 1.
Literally the only costs he incurred were hosting the 6GB model on his own servers like a retard, and aside from having a neat idea he's done nothing to deserve $140,000/year. His code was buggy as shit and crashed constantly until other people patched it up for him. He has no humility and went out of his way to grift people as soon as he got his fifteen seconds of fame.
What does running a cell mean and how do I do it? (Colab version only)
There's little play button in the top left corner of the little blocks, the blocks are the cells, you click that button to run them. Only run one at a time and wait for it to finish. They have a “play symbol” next to them, so you can easily see them. Keep an eye on them, some of them prompt you with questions and won't continue running until you respond.
I got kicked out of Colab for using it for too much time, when can I go back in? (Colab version only)
Just make a new google account. Or make sure to exit completely out of Colab and wait for a while (like between 1-3 hours).
I’m using the Colab version and my game cell keeps crashing when "ping" is active and the AI posts a new response.
Ping doesn't work in Colab. Turn it off.
How do you make the game not look like absolute shit on a phone? (Colab version only)
(X) Easy mode.
Or do this manually:
Copy the notebook to your Google Drive
On your desktop, reduce MaxHeight to ~"400" so the cell stays within the window (you can't do this on mobile because the code editor is fucking broken)
Load your custom notebook on your phone
put phone in landscape
Run the modules
Tap the last cell so the options in the top right show up
Press the three dots
Press "view output fullscreen"
Is there's a way to shrink or increase the cell output font size? (Colab version only) (Already fixed)
Yes, follow this.
- BE THOROUGH:
If making a custom prompt, be as detailed as you can, include a name, location, and general goal with as many key words of what you want the AI to recognize. Use as many descriptive keywords as you can, like nazi, Germany, war, etc. The AI won't remember WHAT you said about those subjects but it will remember that you wanted to talk about those subjects and will lean towards information it has regarding those words.
Example: Your name is John Doe, you are a human businessman. You live in Okrabama City working for Boboston Co. You head off to work one day and notice an UNUSUAL PATH on your way to work. You stray from your usual routine and examine the MYSTICAL doorway, curious as to what FANTASTICAL place it leads. You step inside and....
IMPORTANT NOTE: Despite stating to be as thorough as possible, keep in mind the AI will not remember any unique names/phrases you might put into its prompts once your story ends, and will only attempt to recognize things already in its database. Keep it straight and to the point.
- Be as descriptive and concise with your responses to the AI:
DON'T: stab the monster in the eye
DO: stab the monster in the eye, causing it to reel in pain and blinding it permanently
The AI automatically interprets everything you say as being in a second person perspective. Giving it "stab the monster in the eye" and not giving it a follow up means the AI has to come up with ANYTHING that could possibly fit after the message, rather than the context you intended for it. If you give it a reply that doesn't end in a punctuation mark, it will also attempt to finish your sentence for you.
Side note: try to avoid using "me" in responses, as due to being in second perspective the AI often gets it confused and might make you swap places with whoever you're interacting with. If you feel you need to use it, try to keep it directly tied to actions you are performing that don't have any other entities involved.
- Remind the AI:
The AI can only remember the context of the last 20 posts, if you want to keep it on track, constantly refer to things by name, use adjectives to hint at what the original context was, or flat out just restate something that happened before its 20th response. THE AI WILL ALWAYS ATTEMPT TO RECALL THE ORIGINAL STARTING PROMPT FOR CONTEXT.
- The AI does NOT learn new things:
The ability for the AI to consume and understand completely new inputs is disabled entirely. Put simply, the AI is not truly an AI, it will never evolve and grow. It is an assembly system for words, and the only effect you COULD have on it would be if you fine-tuned again the original trained model (GPT-2) with new texts, giving the AI knowledge of whatever you feeded it with.
- How to play long stories:
Save often. If you think you don't need to save, save anyway. The AI is extremely fragile and will get stuck when it sees something unfamiliar and doesn't know what type of sentence would fit with your prompt. If this happens IMMEDIATELY make something, like using "revert" or (if nothing else works) stop the script from running so it autosaves, then reload the save with the code given to you.
If your AI begins spouting gibberish or looping, immediately revert back to when it last gave a clear response to your input, save and RESTART with a new prompt.
If you feel you reached the absolute limit of what the AI can give you with your current prompt (I.E, spouting gibberish/failing to come up with a new reply), summarize the current state of your story in an entirely new starting prompt and carry on from there.
Helpful tip for Anons struggling to get good results from the AI:
The AI tends to base its style and pacing on the most recent couple of prompts. The easiest way to get your AI to do better pacing and descriptions and stuff is to manually “/alter” the first few responses you get to be more fleshed out, have slower pacing, etc. Once the AI has enough to go off, it'll usually stick to that level for future responses and may only need a little prodding and editing afterwards.
Example:
If you start out stripping a character and get
(X) You take off Anonette's blouse and touch her boobs.
Manually rewrite it to:
(X) You reach your hands out and begin to slowly unbutton her blouse. She looks at you with her big doe-eyes as you fiddle with each button. You snap the last button and watch as her supple breasts spill out of her open blouse. You lick your lips as you take the sight in. You timidly slide your hands up Anonette's slim waste and cup her soft breasts with your hands. You hear a tiny gasp escape her mouth as you flick your nipples with her hands and start to squeeze her tits.
A couple of those and the AI will take a hint. Use lots of different adjectives and nouns so it doesn't get caught in a rut (this site is your friend for this task).
How do you actually talk in this game? Do you say "You do X" or is it "Go and do X" or "Person Y does act X"?
For normal actions it's;
(X) punch the gays in the face
For dialogue (the game says just using the quotation marks is enough, but it’s not recommended):
(X) say "Fuck trannies!"
You can also bypass it and do stuff like:
(X) tell the mormon he's a fraud
Or
(X) yell curse words at the niggers
Or
(X) order the hoe to fuck you
Or
(X) ask the jew to fuck off
What it does is pretty much take what you wrote and add "You" in front of it.
Capitalization and punctuation is not required (exclamation, interrogation and commas obviously required) and it can usually work around typos.
Avoid using "words" like this (in between quotations marks), because the AI is supposed to treat that like dialogue. When needed, use 'words' like this instead.
The AI is a bit too difficult to control, it doesn’t seem like a proper game.
It helps if you treat the game as a lucid dream simulation and guide the AI with your posts like you would control a dream with your mind.
Why is AI dungeon so slow/resource intensive when it's just a text based game?
You're not running a text game, you're running an AI brain with that RAM. You are using state of the art natural language processing neural network, meaning that it works IRL time with no scripted actions. Don’t be surprised if it doesn’t run on your GTX 960.
REMEMBER:
The AI isn't a mind-reader and it doesn't think for itself, it's basically just an elaborate autocomplete engine.
An autocomplete engine takes your input, compares it to a library of words, and uses your search/writing history to determine the most probable words you're trying to type.
AID2 takes your input, compares it to its model, and uses your recent actions to determine the most probable words to respond with what are consistent with your context prompt.
Try giving it more to work off of. Example, say the AI gives you:
(X) You see a building at the end of the road and walk towards it.
Instead of "enter the building" try "recognize the building as the town church and quietly enter" or "realize the building is the tavern and go in for a drink".
Also having a better context prompt helps the more details (about you, your characters, your setting, etc.) you can provide it, the more the AI has to draw from for determining what responses make sense or don't.
Bad Context Prompt
(X) You are Guy. You are a detective in Gotham. People rely on you. You meet a girl named Sally. Sally asks you to investigate her sister's disappearance.
Good Context Prompt
(X) The year is 1935. You are a slender, grizzled, cynical, single, middle-aged man named Guy. You work as a private detective in a crime-ridden, dirty, seaside, metropolis called Gotham. Many of Gotham's citizens come to you with their problems because the police are incompetent and corrupt and many are associated with Gotham's organized crime syndicates. You enter your rundown office and find a woman named Sally waiting for you. Sally is a tall, attractive woman with long hair, full lips, and beautiful eyes. You listen to Sally explain that her younger sister Kelly went missing and that the police refuse to investigate. Sally tells you that the last place Kelly was seen was a high-end nightclub called the Tropicana. Sally hands you an envelope full of money and gently caresses your hand as she asks if you'll take the case.
MORE TIPS FOR WRITING CUSTOM PROMPTS:
When you start a custom game in AID2 you will be asked for a context prompt and an initial story prompt. The context prompt is your synopsis (it establishes the who, what, where, when, etc.) and is essentially the "reference material" the AI will use to determine what responses make the most sense. The key thing you want in a context prompt is to be concise, avoid flowery or abstract prose and make each sentence in the prompt describe aspects of one character, setting, relationship, background element, etc. Use lots of adjectives to give the AI more material to work with for fleshing things out. For even better results, you should make frequent use of the "/editcontext" command during games to update your context prompt with new developments and plot details.
Example:
Bad Prompt
(X) You are Walter White. You are a teacher. You decide to cook meth.
Good Prompt
(X) You are Walter White. You are a brilliant, overqualified, underappreciated high school chemistry teacher. You are a timid 50 year old mustachioed white male with brown hair who wears sweater vests. You live in Albuquerque with your wife Skyler. You and Skyler have been married for 20 years, your marriage is happy but has become distant recently. Skyler is an attractive blonde woman who is extremely naive and nags you frequently. You and Skyler have a teenaged son named Walter Jr who is crippled but thinks the world of you. Your wife’s sister Marie is a kleptomaniac married to a short, crass, angry, bald man named Hank Schrader. Hank works as a drug enforcement officer with the Albuquerque DEA. After collapsing at work one day you went to the hospital and were diagnosed with terminal lung cancer. You have become reacquainted with a former student named Jesse Pinkman. Jesse is a foul-mouthed, slacker, loser, meth dealer with untapped potential. Not wanting to leave your family indebted after your death, you decide to approach Jesse about cooking meth together.
Remember that this is not a true AI in its current state, and all we are doing is holding its hand to where we want it to go. There is an extremely large database of writing prompts in its memory but it's still just an input reader at the end of the day.
Finally, the AI is completely cut off from the rest of the internet except for a direct line to you, and only you can influence it. If your AI makes you rape and murder everyone, you only have yourself to blame.
Keep these things in mind and enjoy your deranged dragon fucking fantasies you crazy Anons.