Create an improvised text adventure for AI Dungeon 2

Machine-generated text as a way to respond to text input provided by conversational humans has been a computing application, at least since Eliza, AI dungeon 2 It will lead it to a new limit.

It basically pretends to be one of those old Infocom-style adventure games, but there is no limit to what players can enter. Whatever the input, it improvises the event through prose to react to them.

It has not yet reached the level of classic or modern text adventure, and sometimes it can lose sight of the situation over multiple turns, AI dungeon 2 You may be amazed at the types of responses you can generate and the unexpected directions that can lead to your adventure.

The game was created by Nick Walton of Brigham Young University’s Institute of Perception, Control, and Cognition, and was so well received that it needed to be quickly reworked to reduce bandwidth costs.Walton answered some of the questions we had AI dungeon 2 Where will it be made and then go?

Who are you and what AI dungeon 2, And who helped make it happen?

I have been working on deep learning technology for the past few years. I have interned at several self-driving car companies and have studied at BYU’s Deep Learning Institute, BYU Perception, Control, and Cognitive Institute.

I have been working AI dungeon I came up with an idea at the hackathon in March. I continued to work on this idea with advice from my professor, Dr. David Wingate.Eventually it finally came together AI dungeon 2 At the beginning of December.

AI dungeon Is the first of its kind in a game where the response to the story and action is entirely generated by deep learning. The great thing about this is that the game is built on a huge machine learning model created by OpenAI called GPT-2. It adapts and responds to almost any action you can imagine, giving players a huge amount of freedom never seen before. Previous game.

AI dungeon 2The popularity of was soared almost overnight. What happened?

I was getting some attention AI dungeon 1.. When it was released in May, thousands of people were playing, so when I started teasing snippets AI dungeon 2 People have begun to pay attention.But when it was actually released AI dungeon 2 It exploded much faster than I expected.

It was released as a game that researchers could play on Google Colab, a way to run code on Google machines for free, but every time someone played it, they downloaded the 5GB model.

Within a few days, so many people downloading the model generated bandwidth charges in excess of $ 20,000. It had to be temporarily shut down until some great community contributors announced a peer-to-peer download solution.

now AI dungeon 2 Is playable as a web and mobile app, and teams are working together to improve the game. Great app developer Braydon Batungbacal builds a mobile app in a week, brother Alan Walton helps build an infrastructure to host a large model, and Thorsten Kreutz helps the long-term effort of this idea. I did. can go.We see AI dungeon 2 It’s just the beginning of an exciting change in entertainment.

Format AI dungeon 2The output of appears to emulate the output of the classic Infocom text adventure. AI dungeon However, you will get a text corpus from Choose Your Story. Was there a format of data massage needed to generate the available inputs and / or outputs?

Yes, there is a fair amount of model and data massage done on the game side to make sure the model works. Cut off the last sentence of the model output and the end of the text token and transform the player’s input to make sure it is in second person format. It also required a significant amount of work to control the iterations. This is a major problem with GPT-2, especially in low temperature (randomness) settings.

Creating your own text generator is a fun enthusiast programmer project. GPT-2 is much more sophisticated and uses much more datasets, but seems to be doing the same. Can you give us an overview of how it works?

GPT-2 is a deep learning language model. The language model basically means trying to predict the next word, which is most likely given the previous word set. The really nice thing about the language model is that you can train with any text (no labeling required) that can be learned from the huge amount of text data there. The largest model was trained with only 40GB of text data.

By learning to predict the most probable next word, not only will you learn how English works, but you will also be able to predict the most probable next event in the world. The model also learns (for example, if you are stung, you can die) and it allows you to use it to do some really interesting things.

The GPT-2 system can be downloaded for anyone who wants to try it out. How do you encourage interested people to get started?

There are some really nice Python packages that you can tweak GPT-2 and try for yourself. GPT-2-Simple by Max Woolf is probably the easiest way to get started and can be done with a Google Colab notebook like the one here.

Site AI dungeon 2 Please note that we are starting to sell mobile app interfaces to fund the servers and bandwidth needed for games. how are you? Are you still worried that you may have to close the project due to financial concerns?

We turned the game into a startup because there really isn’t a great experience without the financial sustainability of our business. We plan to make the base version of the game playable for free, but we will add a premium version of the game that will cover the cost of the server.

The premium version has a lot of interesting new game modes and features, so I’m confident that it will make the game sustainable.

is AI dungeon 2Is it perfect at the moment? Are there any plans for improvements or sequels? Maybe it allows more starting points or provides greater continuity between turns?

There are many things we are working on. In the short term, we’re working on adding multiplayer and text to voice support, but we’re looking for even more innovative improvements in the future.

Such a text generator seems to have a huge and hilarious application as well as simulating a text adventure game. Google strives to get people accustomed to thinking of AI in terms other than Skynet with free courses. Researcher Janel Shane has introduced his own AI to tackle themes such as kitten naming. Do you plan to continue text generation or move to another application?

Text generation continues to be a big part of what we are trying to make AI dungeon And its sequel, we will also start adding other forms of AI content generation using music, images, and audio wherever possible.

Most of our readers are game developers, and some of them use custom algorithms to randomly generate the terrain to explore, and so on. GPT-2 is built to generate text based on other text on the Internet. Do you think the underlying system will one day help create more traditional types of entertainment software?

definitely.Let’s go AI dungeon This is just the beginning of a major evolution of the game and we want to help make it possible. There are many exciting possibilities that this technology opens up and we look forward to seeing what the future holds.

Back to top button