/aidg/ Guides and Resources

Read this page before asking questions. Chances are, they have already been answered many times.

/aidg/ FAQ

A collection of the questions that are commonly asked on /aidg/ threads. Do not hesitate to point out any wrong or outdated information. Last updated: 2021-17-03.

AI
Energy
AIDS
1. What are AI Dungeon Scales (AIDS) used for?
2. How do I get AIDS?
Worlds, World Creator, and NPC Generator
Input
1. General Questions about the Input
2. POV and Tense
  1. Out of 1st, 2nd and 3rd Person POV, which one is superior to others?
  2. But, all novels and books use 3rd Person… 2nd Person triggers bad training data, therefore, it should not be used.
Author’s Note
A/N script
World Info
Formatting
1. Hey, if I use JSON, caveman, Zalty, Onyx, Monky, EWIJSON formatting, will AI remember and reference things better?
2. Why should I not use any formatting and use prose?
Guides
Miscellaneous

AI

General Questions about the AI

Is it possible to run AID offline?

GPT-3 can’t be run on consumer hardware, it’s just too demanding. Old, GPT-2 AID forks (thadunge, clover) and GodAI are still available, but they are inferior even to Griffin in every way possible.

Does the AI learn in real time?

No, the AI is already trained and can’t learn new information.

What do you mean the AI’s already trained?

The AI that the game uses has completed its “training” in 2019, where it has consumed 410 billion byte-pair-encoded tokens (words) from a filtered version of Common Crawl—representing about sixty percent of its weighted pre-training dataset. Other sources include 19 billion tokens from WebText2 representing 22% of the weighted total, 12 billion tokens from Books1 representing 8%, 55 billion tokens from Books2 representing 8%, and 3 billion tokens from Wikipedia representing 3%. If you wish to learn about this further, the link to the research paper can be found here.

Does the AI know about media X?

If it existed before/during 2018–2019 and information about it could be found on the Internet, then it very well could, assuming that the media you’re looking for is sufficiently popular enough for the AI to have encountered it during its training.
If the media in question first came into existence after 2019, then no, it does not.

Can the AI simulate my waifu perfectly?

Any attempts at simulating a specific character will heavily rely on fanfiction content, as there are (theoretically) more of it in the training data in regards to character interactions than it does from any other sources. Test AI’s knowledge yourself, it literally takes ~2 minutes to do.

Is the AI sentient?

No, you are a schizo if you think that. The AI has no intelligence of its own, it’s just a really, really good text predictor. Do take your meds daily.

Is there a magic technique to get the AI to perform exactly as I want?

No, there isn’t. Different people use different methods to get the desired outputs. Ultimately, how good AI is depends mostly on (You)r writing.

Which model is used for the very first response of the story?

The same one you have chosen inside your “Settings”. Previously, GPT-2 was used for the first output of the story, that’s why you may have heard about using “filler inputs”. Since some undocumented update, this only applies to completely empty prompts. As long as you actually type anything inside that field, it won’t be an issue.

AID doesn’t work! It outputs garbage! I can’t get anything good out of it!

Many of the issues with the AI’s behavior can be caused by improper use of /remember & world info, high randomness, and clunky formatting/grammar. In other words, git gud.

AID Website

AI doesn’t output naughty words, HELP!

Coomers need to manually disable Safe Mode in Settings.

Is there a censorship applied to the AI even with Safe Mode turned off?

The word ‘rape’ (this spelling and capitalization specifically) is prevented from being generated in the outputs. Reportedly, some slurs are also filtered out.

ALL my stories and prompts are gone! Help!

As a first response, try resetting your search index inside “Account” tab in “Settings” to restore them.

If that doesn’t do anything, try taking a break and waiting to see if it magically gets fixed along the line. Remember, AI Dungeon is updated live and intermittently, so always back up stories you are fond of as you MAY LOSE currently active adventures. If all else fails, the only thing that you can do is to put up a support ticket and pray that the devs can restore them somehow.

How do I back up stories? Surely there must be some kind of story/prompt export feature…

There isn’t. You just “Select All” on your story/prompt, “Copy” and “Paste” it into a text file.

I have a problem with Android/iOS app.

Usage of the mobile app is discouraged as it is in a state of being constantly broken. Just use the mobile website instead.

Fine-tuning

I keep encountering Count Grey, Sir Kit, Jacob, Karth, Kyros, etc.

The model was repeatedly violated by brute-forcing it to output CYOA text. To avoid seeing them, use descriptive prose, and stop using /do and /say.

Can I get rid of Count Grey for good?

There is a “Banned Words List” inside “Game” tab in “Settings”. It will prevent anyone from typing “svelk” in your multiplayer games AND reduce the amount of this word’s appearances. Take note that it will not cull the words inside it completely from outputs. Count Grey will still lurk in the shadows… Read up more on this here: https://aidungeon.medium.com/controlling-gpt-3-with-logit-bias-55866d593292

Dragon

Is Dragon good?

Yes, Dragon helps and is better than Griffin.

How good is Dragon exactly? I think Griffin is fine and Dragon is a scam.

Dragon’s training data is more than 10x bigger than Griffin. It is the largest GPT-3 model currently with 175 billion parameters, known otherwise as DaVinci by OpenAI. Griffin is the second largest GPT-3 model at 13 billion parameters, known as Curie by OpenAI. This means that Dragon knows how to handle specific fetishes better, easier and in much more detail. See side-by-sides yourself: https://catbox.moe/c/806g5o.

Should I get $10 sub?

There is a free trial available for it. Try it out and see if you like it.

Is Platinum sub worth it? I don’t want to pay $30 a month.

No, $30 tier isn’t worth it unless you literally do nothing but use AID day in and day out.

Why is everything so expensive?! Why do I even have to pay for words on screen?

First, GPT-3 requires incredibly expensive hardware to run, you being able to use it for private smut generation is a blessing already. Second, OpenAI overcharges its customers by driving their prices through the roof due to monopoly. Third, Latitude has no clue how to run their business, bleeding money where it could be avoided.

I have a grandfathered $5 sub I got in July. Now, the “AI Model Type” setting has Dragon highlighted, but when I click on it, I get the notice about upgrading my plan.

It’s a visual bug, you are still on Griffin. All old $5 subs are now called “Silver” and give access only to “Unlimited” Griffin. To use Dragon, you have to upgrade your plan.

Energy

Why is the energy bar not showing up for me? Am I on the Unlimited Dragon already?

To make your energy bar visible, toggle it in the settings menu under the “Display” tab in “Settings”. It is hidden by default. Set its appearance to “numbers” too to have a more accurate representation of how much of it you have left.

How much energy is used for a single action?

1 energy is depleted per action, BUT the first 2 /retry ARE free. That does not count if you /undo and then hit enter, or if you edit any context information (i.e. your previous outputs and /remember, A/N, or World Info).

How fast does the energy bar recharge?

Energy refills 1 unit every 5 minutes until it reaches 100 (~8 hours), then 1 per 20 minutes until it caps out at 2000 (~26 days).

AIDS

What are AI Dungeon Scales (AIDS) used for?

AIDS can be used to recharge energy (the only thing you should use them for), buying Worlds (read next question), and avatars.

How do I get AIDS?

You can get a daily login bonus by clicking on a wrapped gift icon on the home page.
People give you AIDS if they like (You) or (You)r prompt.
If you paid for premium, each time it renews, you get a fixed amount of it, depending on your tier (100, 200, 500 respectively).

Worlds, World Creation, and NPC Generator

What are Worlds?

Worlds are basically a pre-packaged World Info dump, consisting of location info, races, factions, classes, and exclusive world events that are random pre-generated happenings within the scenario. For more information on World Infos, see this section

Wow, that sounds like a good deal! Should I buy Worlds, then?

No. Worlds are shit. Due to the AI’s limitations, it doesn’t handle massive settings well and is limited to about 1400 characters of context, which includes your /remember, World Info, and A/N. The devs don’t know how to use their own tech, so they somehow managed to create a playground where your starting /remember can reach up to more than 1000 of those characters, without even mentioning any of the World Info. What this means is that the World Info that you’re paying for is likely to NOT be put into context as it is cut off by the character limits. You are better off with writing your own settings.

Okay, but what if I made my own World using the World Creation feature?

The World Creation feature still has the previously discussed hard context limits, but a skilled writer can take it into consideration when trying to generate one. That being said, it is still recommended to just write a World on your own, as World Creation costs you AIDS to use. The general consensus is that the World Creation feature’s only recommended usage is to use it as an idea generator, as its randomised outputs may have interesting results.

What about the NPC Generator feature?

The NPC Generator also uses AIDS to generate its results if you’re not doing things manually, and its randomised outputs can be a hard swing and a miss. It can be used as a repository of characters you would like to see frequently (say, for example, a recurring antagonist or companion), but it is still seen as inferior to keeping things saved in a text file.

What do all these formats in the NPC Generator do?

See this section.

Input

General Questions about the Input

Why should I use /story? That is just like writing the story myself.

/do and /say garble your input, lowering the AI’s output quality drastically. Also, everything /do is capable of is easily achievable with /story, along with complete control over your input.

Hm, but I like random outputs AI gives me when I use /do.

Write “you attempt to” to make the outcome of an action uncertain and add RNG.

AI forgot about my characters…

Use /remember and World Info to maintain continuity and permanence in stories.

POV and Tense

Out of 1st, 2nd and 3rd Person POV, which one is superior to others?

It’s a personal preference.

But, all novels and books use 3rd Person… 2nd Person triggers bad training data, therefore, it should not be used.

If you write your scenes as an actual prose and not as a shopping list, it literally does not matter. If you are a decent writer, you can experience kino outputs no matter what Tense, POV, dialect or quotation style you are using. Read this question again and get better at writing.

Author's Note

What is Author’s Note? I see people referencing it, but I can’t see it anywhere?

Author’s Note (A/N for short) is currently a premium only feature. HOWEVER! You can use it for free if you read the following sections. You either click the pin or type /remember to view it. Here is how it looks:

I don’t understand, where is the text I put inside the A/N shown?

Text inside A/N field is injected 3 sentences above the last sentence of the output+input. It is not shown inside the main story text. But it is sent to the AI and it impacts its outputs drastically.

Wow! That sounds very interesting. But I am a free user…

Kind anon made a script superseding the default A/N completely with new features. All users can use it, regardless if they know anything about programming. Read up on it here — https://github.com/CoomersGuide/CoomersGuide.github.io/tree/main/Resources-And-Guides/Scripts/AuthorsNote.

A/N script

What does A/N script do? I don’t understand.

I suggest you carefully read through the documentation again. If you still don’t understand, read the next question.

How do I install it?

You download the scenario.zip file from the repo. DO NOT DOWNLOAD IT OFF SOMEONE ELSE! You can rename it to something like scenario-an.zip so that it would not be overwritten by other scripts.
Go to ‘My Stuff’ —> ‘Scenarios’ —> ‘Create Scenario’
Scroll down to ‘Scripts’ and click it.
Click the ‘Upload’ button .
Choose the file you have downloaded in the step 1.
(OPTIONAL) In ‘Shared Library’ tab (the already active one by default) there is a line:
```
state.authorsNote = "" // String for Author's Note.
```
Set something inside the quotes so that A/N will be active during the very first output (the one that happens once you click the ‘Play’ button . Example: state.authorsNote = "This a female sexual domination story." // String for Author's Note.

I think I installed it, but how do I use it again?

You type something like:

/an This a high fantasy story.
The protagonist wandered for days among the ruins —

This will set A/N to This a high fantasy story. AND send The protagonist wandered for days among the ruins — to the AI at the same time.

How do I turn off the A/N display above the text input box?

Type /anv.

I can’t set up Author’s Note, and `/anv` does not work.

Users who reported these issues used Android app and /do mode. Use /story and web version (yes, even on the mobile).

Still can’t get it to work.

If you are sure you followed every step correctly and it just refuses to work, ask in the thread about it; script’s creator is active in it.

World Info

What is World Info (WI)?

There is a basic guide available for it. Read it to understand how it works.

Are WI keys case sensitive?

No, they are not. Capitalization is used for logical understanding of information (it just makes sense to capitalize names, regardless whether there is a difference or not).

Should I separate my WI keys with spaces after/before commas or not?

Spaces should be used for short or common words: cat ,pig ,key , If you do not use spaces in the above examples, words categorical, pigment, keyboard will trigger a WI entry. REMINDER! The usage of such common keys is not recommended, since any mention of a cat in your story would load WI entry associated with it. Unless your story will only ever have ONE cat in it, this will confuse the AI.

Wait, but… I used wrong spacing for all my WI entries up until now… Should I go remake all my previous prompts?

If you styled your WI entries like this: Emma, sister, Big Sis the only case when WI entries would not load during usage would be a start of the paragraph.

Your sister¹ Emma¹ is very shy. Sister² — such a good thing to happen to you. Big Sis¹ looks after you on a daily basis.
Big Sis² loves you very much.
Emma² loves you unconditionally.
'Sister³ is such a good friend to me!' you think to yourself.

Here sister¹, Emma¹, Sister², Big Sis¹, Emma² trigger a relevant WI entry. Big Sis² and Sister³ do not trigger a WI entry.

If you think that warrants a massive respacing of all your WI entries across all of your previous prompts, go right ahead.

So spaces in WI are a no go after all?

The spaces before commas are fine for short/common words (read this question again). Since adding a space after commas is grammatically correct, it was assumed that should be done for WI entries as well. The only thing to be said about not enunciating it enough in the thread is “AID never offered much documentation”. If your WI worked all this time despite this, it means it’s not as much of an issue as you might think.

Formatting

Hey, if I use JSON, caveman, Zalty, Onyx, Monky, EWIJSON formatting, will AI remember and reference things better?

Don’t use any programming/game formats. Write everything in normal English.

Why should I not use any formatting and use prose?

Short answer: they don’t work. There is no evidence they help, they leak into output and once they are criticized, their proponents respond with “It feels better to me”. Thus, it is recommended to use prose: if nothing else, the only leaking you’d have to worry is from repeated sentences.
Long answer would require a whole separate debunking article. Once that will be written, the link to it will be here.

Guides

“/AIDG/ Writing Guide” and “A Coomer’s Guide to AI Dungeon” contradict each other. Which one is legit?

They were written by different people during different periods of time, with each author having his own approach to AID. Try both methods and see which one works for (You).

“/AIDG/ Writing Guide” is outdated. Why is it still up?

All information (besides GPT-2 for the first output) inside it still holds true to this very day. The fact that it predates Author’s Note does not negate its usefulness.

Why is HTML version just a converted PDF file?

HTML version is intended for mobile devices and browsers without the PDF viewer.

LaTeX is fucking garbage, you should have sticked with Markdown.

If you don’t like the guide’s visual look, port/change it however you want and submit a PR.

“Minimalist” version of Coomer’s Guide is not minimalist enough.

See this question.

Miscellaneous

Who is the Mormon?

He is the creator of AI Dungeon 2.

Why do you call him that? Does he have an actual name?

Mostly out of spite for his incompetent handling of the Colab version. Even though it’s old history now, it’s just a ton of fun to use that nickname. If you actually think it’s offensive and malicious, I recommend talking to actual people more, instead of lashing out at “4chan hackers”.

Wasn’t /aidg/ responsible for a DDOS attack on AID some time ago?

No, it was not. We are people interested in AID prosperity the most. Any claims saying 4chan threads carry ill intent towards AID and Latitude are propaganda.

What is the Lovecraft model and why is everybody gushing about it?

The Lovecraft model was an AI model that was fine-tuned to the works of H.P. Lovecraft. Due to it being tuned to actual literature, it manages to avoid the pitfalls of the usual fantasy CYOA works that the vanilla AID is trained upon, and is thus far more eloquent and suited for generating passages. However, it was taken down due to its cost and/or the lack of interest by the plebeians of the community. We are hoping to bring it back by voting on it in the suggestions page here: https://features.aidungeon.io/ideas/AID-I-1073. No, it will not turn Count Grey into niggerman.

This site is open source. Improve this page.