AI Dungeon Features
Less moments where the ai completely forgets where the main character is. I know it won't be 100% fixed, but anything helps.
The AI is not nearly complex as any of you think it is. It is really just an advanced text predictor. It reads the last 20 prompts and the context, as well as using the training data, to try to predict what will come after your input... but it doesn't know that anything has happened or that certain things have to happen. It is basically looking at individual words and deciding "what will come after this word," it doesn't track events, people, things, or anything. It is up to you to guide the AI.
I told the AI I was Bob using "/remember" and 1 prompt later, the AI is calling me John. PC Version v1.1.8.
The game forgetting *where* you are is a pretty complicated problem to fix (although /remember is *supposed* to help.)Forgetting the player's name should definitely not happen, I'll put my character's name in the prompt and the /remember box and the game still forgets randomly.
This would probably be my number one ask at this point. Even if it was just a set of longer-term nodes that weighted words that should be considered from previous inputs would at least be something. This might, for instance, let the game engine track themes, characters, or other elements that appear often in user text or generated text that it can incorporate into the shorter term memory it's already generating. Oh, and definitely make sure "remember" is given a higher weight.
I hope you can fix that.
Sadly the nature of the AI is basically a very advanced text predictor. Keep that in mind while playing. Sometimes the best effect is achieved by formulating sentences that will make it easier for the AI to notice correlations between words. New models of the AI are being introduced over time and their performance is slowly improving, so you could say it's being actively worked on. I suggest pressing retry button to make the AI generate the output again or edit it if it's near perfect.
"Sometimes the AI’s reply doesn’t make sense and it doesn’t even apply to the story." (suggested by Behlah kanorwala on 2020-06-07), including upvotes (1) and comments (1), was merged into this suggestion.
There are other AI/ML methods for doing text summation. Summation already figures out a lot of important points. If the game implemented a summation engine for previous content that could work as an advanced /remember.
It would be pretty cool as well if the AI remembered your name and whatnot.
The IA keeps forgetting my race, my name, if someone is my wife or child, my title (king, warchief, lord, guard), my allies (all the time makes my allies attack me)
"best memory for ai" (suggested by Marcos Alexandre Alves Cazuza on 2020-08-02), including upvotes (2) and comments (0), was merged into this suggestion.
Perhaps a good approach would be to condition another instance of GTP-3 to write its own context paragraph, consisting of a summary of both the most recent generated text as well as the previous summary. This way it should retain the most important events indefinitely while filtering out any fluff.Ideally you would want this to happen per character, and maybe per location, but that simple "dungeon master's notebook" could add a lot.
I agree this should save along with the storyline!
I cannot use Dragon yet, but I am making an assumption that the games' base memory limits apply to it as well. I read somewhere that, as transformers become more powerful, their memory depth increases exponentially or something. Maybe remove any memory limits (or increase them to 8192?) I could be completely wrong about this, though.
Its actually the complete opposite. Doubling the memory limit quadruples the computation time as well as the model size. Without changing the model to a different model (ex. reformer, transformerXL) The memory size will always be limited.
What if you try connecting multiple smaller models that perform as well as a larger model and connect and train them together, with each of them attending to a different part of the text?
Of course, it might be difficult teaching them to stay optimized and not just throw everything away that they know to fit a new format, but I think it could work.
Multiple DistilGPT-2 models, linked together by a neural network that is trained to use their outputs as input.
Or even just at the 'edges' of the layers of each network, they connect, but act as separate models with different areas of the text that they can look at.
And they have a final layer that outputs their text.
If you try to distill GPT-2 down more, you may even be able to get away with even smaller individual models.
And with the AI models able to see, individually, separate parts of the text and communicate with each other, with each having their own computational 'bandwidth' (the whole model not having to attend to every part of the remembered text), they may be able to outperform a singular model.
And trained as a whole on the task of predicting text, while somehow attending to each individual model to keep them from forgetting their previous training and dropping their benefits, then these models could be coherent enough to make non=premium less of a constant clicking of retry.
Hmm, one thing that could help with this is training a small neural network to assign an "Importance" score to each prompt, and using a weighted combination of distance and importance instead of only distance when determining what prompts to keep in memory. You might also automatically /remember ones with exceptionally high importance score, especially if the player doesn't seem to manually use /remember (just check if it's been used manually in the current adventure so far).
A convincing, evolving story needs the AI to model and store a significant amount of state data. It should know roughly where you are (a woodland glade shouldn't suddenly morph into a train), details about the player char (name, age, sex), the genre (no cars or intercoms in mediaeval fantasy), and info about NPCs (physical description, age, disposition towards player). You *don't* need to process all of this at once, e.g. store NPC info when a character "exits" and reload it when they reappear.
That would require access to the AI's source code and a way to 'hack in' internal persistent AI memory, which actually could be done. But GPT-3 is closed-source, unlike GPT-2.
At the top of my new World /remember file it has listed that I am a human female, but characters are constantly trying to interact with my penis.
@Developers: You might look into distributed computing for this. You could give users the option to donate a portion of currently unused memory of his device to the AI. Then distribute that as additional game memory to all users who are contributing to the AI memory.
@Jos: You should be able to fix this by creating this world entry: key: you entry: you do not have a penis. you are female. you are not male
@FrederikE. That's not how the memory of the AI works. The memory length of the AI is absolutely fixed.
When I was on a city, called Gorgok (another character told me). After maybe 10-20 commands I typed, I made my character say "What is the name of this city again?"
The AI answered "I don't know" after that, I kept on altering it, some of the replies were the same and some don't. One reply said that "I think it starts with J".... Bruh
This is a limitation of both GPT-2 and GPT-3, it can only handle so much data so much before it starts discarding some of it. The only way to work around this is to add World Info based on things you need the AI to remember the meaning/purpose of, and add /remember stuff so it knows that whatever's there is of constant importance.Currently the token limit is about 1000 (1000 words), and that can't be easily changed, so the best that can be done is info management/compression, which is tough.
this is a must, when I was playing on my world, the AI forgotten the name of the city(City of Gorgok) that I was standing on. One character said "I think it starts with J".... Bruh
Keep this in mind, your 5 word input, that's 6 tokens (Because 'you' is added), and the last 20 actions, well that's 1200 tokens. That's too many for the AI, so 206 tokens are cut from the oldest part of the total text, and /remember and World Info count against this total as well, so that's why you suddenly went from being in Gorgok to being in Jicol, because any context relating to that was stripped ~16 turns ago. Auto World-Info needs Dragon to not suck, so just be diligent with /remember+WI.
"AI remembering most of the things" (suggested by The0IV3 on 2020-09-24), including upvotes (1) and comments (2), was merged into this suggestion.
Perhaps have it so new characters and locations are added as keys in the world info and have the information for the keys update as stories progress? Not sure how feasible this is but figured I'd suggest it anyway
The AI even forgets WHO the main character is; even with a "your name is ___" in the memory window, it randomly changes the main characters name constantly.
Having the AI more regularly and reliably reference the memory and World information would be helpful
"Make the AI better at understanding short and long text" (suggested by peepee on 2020-10-05), including upvotes (1) and comments (0), was merged into this suggestion.
Allow an option to generate https://www.ericswallace.com/triggers for a given prompt to enforce certain concepts:[WORD][WORD][WORD][WORD][WORD][WORD][WORD].You look up. You are in Jicol.You look up. You are in Jicol.You look up. You are in [PREDICTION]Flip each [WORD] token until the prediction changes to Gorgok despite the previous text.Could also be used to auto-generate training data, as long as the trigger is omitted from the data that is saved.
I understand that it's difficult for the AI to remember and understand complex things, but it forgets my name constantly, despite being in /remember. It seems to default to a certain name, depending on the genre, if no reference to your name was made in the previous few lines, completely ignoring the data in /remember. However, when I remove anything relating to the original template genre, it seems to actually remember the player's name.
I love this system but I feel it will only ever become what it is meant to be if it can stop being so forgetful. The AI, despite being pretty creative, constantly forgets important details. I don't even know why there's always a prompt asking for your character's name since the AI will just forget it two lines later. This is because the AI doesn't seem to track anything, it just generates text based on previous input. I guess it is kinda tracking stats now, but imI feel that being able to properly track characters, locations and other basic story info is a lot more important.
0 / 1,000
* Your name will be publicly visible
* Your email will be visible only to moderators
When submitting your email with a suggestion, here's our policy on what we do with the email.
We use your email to
Your email is never displayed to other users. AI Dungeon staff will be able to see your email address.
We’ll email you when the status of the suggestion changes. You can unsubscribe at any time with a single click.
* Comments and upvotes from this suggestion will be moved to another suggestion.
You cannot undo changes once they're done.
Edit the selected suggestion and create a new one.
Comments and upvotes will also be moved.