AI getting into dungeon mastering, achieve kid level of dream logic

http://www.aidungeon.io/

https://www.polygon.com/2019/12/6/2...text-adventure-openai-how-to-play-nick-walton
You can do nearly anything you want in this incredible AI-powered game
The closest thing there is to an infinite game

https://twitter.com/nickwalton00/status/1202706532095258625

It has some limitations obviously, tend to mix up characters, have limited context and memory, crash and get stuck in loop, but it's also quite resilient to input.
And it's basically just a vanilla extension of the GPT ai system ...so it can be further improved, especially if we combine different system to help it (probably like a neural turing machine and data encoding that aren't just plain text)

 

Honestly that's pretty amazing. Some big steps in the right direction!

 

And it's just based on raw data input, that is, it has to infer the logic and semantic itself, even with the shortcoming that's incredible.

So it can probably further be improved, text adventure have more coded semantic than trying to decode raw text on the fly, we can take advantage of that surely. Or we can probably have a multi stage system, each fed on certain data semantic, to breed larger coherence.

If the system can track context and character better, we can probably add image generation on top. And even with no neural network, feed all of that to generic pcg.

The make game button isn't so far off

 

I haven't tested yet, but do I understand, user types response after > character and system produces most reasonable outcome?

I just wonder, how difficult system is, to implement lets say inventory, or collectable items?

 

Having played around with it for a bit I think it's less "responding" and more "contributing to the story". With a traditional text game you would enter an imperative sentence and the game would attempt to act upon it. With this you can enter lines like "an old man is standing nearby" and the code would attempt to weave him into the story as you continue.

 

@Ryiah This kind of reminds me like playing with a chatbots
This sounds however interesting.
Yet, when I had chance to give a try, all download mirrors are down :/
I will try it tomorrow.
I tried however AI Dungeon Classic, which is based on option 1-4 selection. At least what I experienced.

 

It's down because it was too popular, and host was saturated, it literally blew up overnight after one tweet

 

Am I the only one who isn't ok with that guy building a business, based on scraping copyrighted intellectual property from a website that various authors seem to have contributed to for almost 2 decades? I actually think this is worse than cloning a game with your own code and graphics.

The fact that no one seems to care about that aspect bothers me.

 

It bothers me, but so does most of the way china operates in terms of IP and games, and thats not going away. Ive learnt to accept that the way people treat copyright in the 21st century is vastly different to how it should be treated

 

@Martin_H weather we like or not, it is how world operates.

However, AI Dungeon 2 is on github with MIT licence, so anyone can contribute and benefit.
I don't see any difference to be honest, in comparison to Linux, Blender and many others.
You can use product for free. At least allow benefit creator in some form.

We making games and aps, in huge part relaying on free accessible source online.
We contribute, we use, so other can enjoy our products.

 

Imagine you have written books for over a decade, and I make some machine learn thingy that I then train on your books to write books in the style that you wrote in, and even though all my generated books are objectively worse than yours, people think they're so bad that they are hilarious again, and my S***ty AI generated books go viral and I start a patreon to fund further stuff, and I release the thing that now is basically the "essence of your books", under MIT license, so that every stranger can do with it as they please. Even if it took me hundreds of hours to get my side of the thing going, those hours are nothing compared to 10 years of you writing your books. And without your books as a training dataset no one would give a S*** about the thing I made. Without the proof of concept, it's pretty much worthless. I'm taking your work, monetize it for my own benefit, and even allow others to do the same. How is any of that making it less bad? If anything I think it's worse to distribute IP that isn't yours to distribute in the first place under a license such as MIT.

The Blender comparison shows you're totally missing my point imho.

 

@Martin_H you could argue even took you 50 years to write a book.
Then some AI did same thing in few hours. Unfortunately, you late.

Yes I feel your pain. But same I can argue, that someone did game faster than me, because they had funds and team, and I had only hobby part time dedication.

• Write Brothers, vs Samuel Pierpont Langley (who had huge funds)
• Nicola Tesla vs Thomas Eddison
• Graham Bell vs Antonio Meucci

Then

• CERN used machine learning, to generate scientific results in hours, which otherwise they been working for years.
• ROS foundation used commonly spread robotic related algorithms. Thx to that, we got robotics and AI advanced as we have now. That contributed to DARPA and famous DOG series of robots.
• Mobile devices use machine learning for super clear images, even in bad conditions.
• https://comma.ai/ uses driver user collected data, for createing self driving cars, which are much cheaper than any other current alternative.
• Search engines utilize our search results, to provide us best (well not always [ignoring all censoring etc.]) results, and translation. Is much more than just simple keyword. Is years of processing data from us. Mind, I am not fan of everyone data collection and spying. Nor targeted adds bombarding. But I understand implications and benefits.
• Academia uses range of sources available, to produce more sources available, as well as commercial products.
• For every commercial product, there are number of free alternatives. People will contribute, where commitment is.
• We use all that tech and social media to advertise our products. Smarter will know how to do it effectively and cheap. Other will pay big buks for campaigns.
• Google also contributed to range of open source products. Chromium, Blockly -> Scratch (for learning), etc.

To be fair, while I had time where that annoyed me, when MIT or relevant were used, for commercial products, but in the end, I use other contributions myself. I would be hypocritic otherwise.

Gosh most of us use Unity for Free

Worse than that in my opinion is, when someone rips your source code from paid product and makes its own. Now, you can be angry

Unfortunately world is not fair, but we can learn mitigate that, to our benefits.

 

Is it okay for a human to write books in someone else's style and make money?

 

This

Unfortunately we build on the shoulder of giants, and sometimes it doesn't work for us, we just lay foundation.

Here is the thing, here, the tools are available to everyone, that's not gated. It's not reusing old work but learning from it, the end result is very different from the source. Also it's every new way to do stuff, and figuring how to make it work was the actual true thing, not replicating old work.

I mean how else you can fight a dragon by flexing muscle at him, then invoke woodland creatures that will help you flex so hard the dragon will fly away in terror. Nobody ain't writing that game to make a dime, thrust me.

So it's not stealing as much enabling never seen before stuff.

We can embrace that, or be like people who tried to stop the car revolution by trying to be more efficient at horse delivery. Or people who tried to stop computer, etc... your skill still matter, but you have to contextualize it in that new inevitable shift.

I swear if I didn't have sweat because who will need a game designer if it only get just slighter better with different data.

 

Offline version
http://familiarcycle.net/aidungeon-on-windows.html

 

How it was made:
https://towardsdatascience.com/the-...2-shares-gpt-2-finetuning-advice-e5800df407c9

https://pcc.cs.byu.edu/2019/11/21/a...dventures-with-deep-learning-language-models/

the reddit
https://www.reddit.com/r/AIDungeon/
people are already modding it

 

Actually it seems that it was "simple" to do (conceptually):

1. feed input text and store it (context)
2. take player input and store it in the current context
3. add this context to a context list
4. Use the context list as input to the NN, up to 8 "context" in this version, to generate the next output.

Basically the memory is in the context, after 8 paragraph it lose track of what's not in the context list.

It use the open AI GPT2 trained model, then fine tune with 10mb of data scrapped from this site http://chooseyourstory.com/.

 

a horde of demons surrounded my house. i type "a mysterious screaming child bursts out onto the battlefield, the demon's heads explode when they hear the child's voice" all the demons died but the game recognised that was kinda bs so it killed one of my party members. next scene these guys where exposition dumping and wouldn't let me leave so i used my hidden technique "instant transmission" to teleport to where the rest of my party was. i used it again to find a secret castle i was so fed up looking for and the game called bs again and made it leave 2 party members behind. later i was grabbed by skeletons and my op child mage cast chain lightning and it actually damaged my character so much he needed heal to walk. ten outta ten

 

While it initially looked amazing it simply seems to regurgitate the most suitable text it can find without much understanding of context.
AI seems to be reaching a plateau and needs ever bigger amounts of data fed to it to function (& copy :/ )!

 

Well given it's an experiment that wasn't supposed to be this good, and the code has zero context for anything at all (ONLY the 8 last paragraphes). You can expect fully engineered version to encode context in the future, they will build from that.