AI Dungeon: More Adventures with AI Chatbots
Imagine having the full D&D experience without needing to find a real, human DM. All the set up, narration, and other things done automatically for you, along with the ability to play and interact with your friends.
The best part?
It's all infinitely generated and customizable, giving you a near-eternally variant game of creative adventure, accessible by both smartphones and general web browsers.
Welcome to AI Dungeon.
Sounds Awesome! Tell Me More!
In continuing my exploration of AI, especially with ones designed for consumer use and entertainment, today I'm going to be sharing my adventures with AI Dungeon, a text-based game using AI to generate an adventure for players. Technically, it's a chatbot, since players are basically interacting with a GPT-3 AI (explained below), which is basically responsible for being the Dungeon Master of the whole thing.
The basic gameplay is like this: you create (or join) a game, and then choose the setting. What's great is that there are general (or classic) scenarios like Fantasy, Mystery, Apocalyptic, Cyberpunk, and others. However, both the developers of AI Dungeon and other players have themselves designed scenarios and settings you can use as well! So whether you want a tense fantasy setting where a civil war is about to break out, or if you want a Lovecraft Cthulhu-inspired adventure, it's all basically there for you to explore!
The web browser UI. So many worlds to explore!
Based on what setting you choose, you may also need to choose a class or job. Sometimes, other parameters are asked for, or less parameters are asked for, depending on if you're using a customized scenario or not. Then, the AI generates a world around all this stuff, and off you go!
As mentioned before, if you wanted to, you can actually play this with other friends as well. Each person simply needs to have the app (or go to the website) and log in. Then, the host creates a game, and then sends a link to everyone.
My First Few Adventures
I first found out about this little AI game while browsing through subreddits on Replika, another GPT-3 AI chatbot program I reviewed a few weeks ago. One of the members mentioned this, and I had to take a look. I downloaded the app, signed up for a new account, created a story in the “mystery” genre and became a detective, and before I knew it, I was off on my first adventure!
Here it is in all it's 'glory'!
Yup! You read that right. In my very first adventure, the AI decided kill me without letting me interact with it at all, haha!
It was impossible for me to not have a good laugh after that, and I decided to create a new game and continue exploring the whole thing. Here is another one I started:
As you can see, this one had much more content. I, Demetrius the wizard, was exploring some random ruined fortress for a book of essence. While exploring, I fell asleep, and wake up to see that I've been teleported to some mysterious grassy forest. VERY COOL.
At this point, I should explain some stuff.
As you can see, some of the paragraphs in the chat box above have a little “>” sign in front. Those are inputs made by the player. The ones without the “>” sign was generated from the AI.
All inputs are typed in, and are “Do”, “Say”, or “Story” inputs (to switch, tap the colorful purplish icon). “Do” inputs are actions that you do, “Say” inputs are things your character says, and “Story” inputs are description that you can offer to the AI to help the story along.
Above where you type are a few features you also can use to help the AI and story flow. The 'World' button allows you to put data into the world to build it up. The 'Pin' button is to help the AI remember certain things that are important (e.g. the people in your party).
The 'Pencil' button is for you to edit previous texts, since the AI will sometimes put out stuff that really won't make sense. There are 'Undo' and 'Redo' buttons, as well as a 'Reroll' button to allow you to help the AI generate alternate texts.
Which segues perfectly to the next thing: sometimes, the AI doesn't make sense, or isn't very consistent. Here's an example:
So the game itself told me I was a bard (this was a self-generated scenario). Then, the AI generated my Dwarf saying that she wanted to be a bard? And this was directly after the fictional soldier asked me to join them, because I was ALREADY A BARD.
And that leads us to a discussion about AI today, its current implementations and limits.
GPT-3 and the World of the Future
The AI system being used by AI Dungeon (as well as Replika) is called GPT-3. It is a third generation model of AI produced by OpenAI, a research group based in San Francisco that created it for deep learning language. Basically, GPT-3 can produce human-like language, and even code software and argue with people.
This is pretty crazy stuff. GPT-3 has been developed to the point that most people can't distinguish words written by a human versus written by the AI. That alone, of course, carries with it some utility as well as dangers. Such dangers would include things like spam, data theft, fraudulent writing (i.e. college application essays), and other things besides. This moral and ethical part, I won't be exploring in this blog today, but it's important thing to note.
Rather, it's important to point out what AI is currently, and most definitely isn't.
AI is based mostly on how we understand an animal's ability to learn. In general, as far as we understand it, biological brains are basically a network of nodes that have inputs and outputs. This network receives an input, changes patterns within the network to match that input, and then produces an output. The output is then funneled back into the input to check if the pattern that was produced matches the output.
Because of this, all AI, including GPT-3, are very good at imitating certain behaviors, and learning from those behaviors. But the key part is repetition, meaning the AI must continually receive and produce inputs so that it can match and produce the correct outputs. Thus, each AI, again including GPT-3, is usually pre-trained with certain material (i.e. Wikipedia) before it is released for others to use.
What I've summarized so far is just the basics. The actual inner workings of this stuff is much more complex, including multiple layers between inputs and outputs and such.
What this all means, for us lay-people, is that AI is very good at learning a singular behavior, provided the correct input and amount of repetition. But it is very bad at learning a large number of different behaviors, and juggling between them.
This is primarily why Replika and AI Dungeon, and any AI that is based on these methods, are really great at creating grammatically sound sentences, but not necessarily meaningful or coherent ones. We all laugh at the fact that early developments in this direction led to some really hilarious Harry Potter stories. But such hilarity is due to the fact that AI can certainly output grammatical structures similar to J.K. Rowling's writings, but cannot necessarily produce something that we would consider good literature. Good literature has an abundance of word, image, and other associations that are meaningful to the human experience, both past and present, which are written into our biology as well as our psychology.
With GPT-3, some of these barriers are being broken. After all, Replika is so human sounding that it is legitimately advertised as a emotionally helpful companion, and users generally agree. And as you can see above, AI Dungeon is very good at producing text that sound like it a story that could have been written by a person. But it would be a person with extremely short memory, and an inability to truly conduct meaningful tasks outside of the chat box which could inform it to better interact with the player.
Thus, it is still a long way before anyone should worry about AI taking over the world. We should probably worry more about the humans that will attempt to use it to do so.
Or, perhaps, maybe Ron's Ron shirt really is as bad as Ron himself.