AI Dungeon 2 has just launched, and it is a reimagination of how games are played. Deep learning allows gamers to create their own ‘infinitely generated world’. So how does it work?
AI in gaming
In the BYU Perception, Control and Cognition Laboratory, a team is working with the aim of creating AI that can perform complex tasks as well as humans. BYU uses cognitive modeling to write deep RL algorithms.
Nick Walton’s original AI Dungeon used the OpenAI 124M parameter GPT-2 model. This was able to calculate potential player choices, and then present results from within those parameters for the gamer to select from.
From a user perspective, this sounds fairly similar to the Stadia AI. This uses similar mapping to anticipate user choices and thus is able to respond to commands immediately. Tiny late-stage adjustments are then made as the game continues.
Next level artificial intelligence
AI Dungeon 2 takes things one step further, and it is quite a leap!
This doesn’t have pre-generated actions but instead allows the player to enter anything they like. That’s right – literally anything! Hence the claim that AI Dungeon 2 offers an infinitely generated world, since the possibilities are, indeed, endless.
Dungeon 2 uses OpenAI’s largest 1.5B parameter model and Walton worked with chooseyourstory.com to finesse the technology by working through different interactive stories. It also follows the Salesforce CTRL model to reduce the repetition problems they had seen in AI Dungeon.
How do you play?
Here is where it really gets imaginative! At the start of the game, a player chooses from a number of settings and characters. This begins with a start prompt and a contextual sentence to begin the game.
Every action you choose creates a response where the context sentence is also fed to the model, to ensure this stays grounded in the same story. Dungeon 2 uses N=8 as the optimal amount of memory to feed the model and allows both new actions and the original context to be repeated after each command.
There are some limitations. The model can struggle to identify separate characters in dialogue, but has also shown ‘remarkable understanding and writing quality.’ Check out an example below!
The fascinating outcome of this game is that AI can be shown to have the ability to create ongoing stories on command, and Walton calls this a ‘dramatic advancement in AI-generated interactive fiction‘.
You can play the game and check out more examples here.
Are the machines on the rise?
The potential is huge. However, AI technology needs hundreds of millions of interactions to create meaningful responses. This means that deep RL algorithms capable of complicated tasks need a very fast simulator to function.
Humans in comparison need much less ‘training time’ to pick up a new skill and can learn from watching and experience. BYU is trying to overcome this barrier by using cognitive science to help the machine to mimic human cognitive abilities. The three key factors are:
- The human ability to build and visualize models of the world
- Our ability to use prior experiences to incorporate abstract knowledge
- The knowledge of how to reason about things we are unsure of
These thinking skills work around core human skills; imagination, memory, and experience. The same things which AI simply cannot have – it has to be taught and shown everything it needs to know.
The lab thinks that it is possible to enhance the capacity of AI by combining two different frameworks. Deep neural networks are best are processing images, and Bayesian models excel at reasoning and cognitive tasks.
BYU acknowledges that this might sound a bit pie in the sky to most of us but, they continue to work on developing AI that can actually respond and interact like a person. They say ‘Great strides require great risk. We’re on our way.‘
What do you think of the game? Are you as excited as I am to try it out – just to see how bizarre a story I can concoct? How long do you think it will be before cognitive AI can realistically mimic human interactions? Share your thoughts!