October 6, 2022

Next Steps on AI Dungeon

We’ve been talking to many of our players about their experience on AI Dungeon and have received lots of great feedback. We’re grateful to all the users who have shared their suggestions with us.

We identified some common themes from the feedback, and they’ll become the focus of our work for the next few months.

1. Stability

Many of our players have been understandably disappointed by bugs, app crashes, and downtime they’ve encountered while playing. No one likes when they’re excited to play a game or continue building their story only to encounter an annoying bug or lose some progress.

We’ve already completed several key rewrites to address code that isn’t stable when we’re patching. But despite these improvements, users still experience far more issues than desired, and we realize we haven’t done enough yet.

Here are some of the areas we already improved or have plans to change to improve stability:

App System Details Status
New Payment System Returned mobile payment, more consistent
Code Stabilization TypeScript throughout AI Dungeon app + API (this helps catch bugs earlier in development)
Updated Help Center Check it out at help.aidungeon.io
Updated Login/Authentication We’re migrated to using Firebase (a Google product) for our login system. See #3 below.
Mobile App Stability Upgraded mobile packages, retired support for old app versions (Android 8.0+, iOS 15.0+)
Updated Front-end Design + Component Library See #5 below
More Consistent Ad Experience See #2 below
Updated Model Datasets Improved coherence, reduced model errors, see #4 below
Refactored Navigation Better handling of navigation within the app using the latest versions of React Native Navigation
Improved Beta Shifting Beta version from feature flags to different code versions
Refactored Core Game Loop Performance Improved editing experience with various performance enhancements
Refactored Model Infrastructure Behind the scenes improvements to our code for handling model calls and text manipulation
Improved Error Management Fewer errors happening with more informative error messages for players when they do happen

Last Updated 11/1/2022

Please continue to share feedback about bugs at https://bugs.aidungeon.io, which we review daily. We also respond to issues and talk with players in the AI Dungeon Discord server.

2. Ads

We’ve heard the feedback that ads have been frustrating for many players. Ads are the latest solution we’ve implemented to support a free version of AI Dungeon (energy before). We’ve learned a lot about how ads work and are exploring the next steps to improve how we support free play.

We’ve been running several experiments to better understand the impact of ads on the free experience. We’ve tested fewer ads, removing ads, increasing the number of actions awarded, and different types of ads. Our goal is to have a clear understanding of what an optimal free experience looks like and how ads should play a part in that.

We’re still early in the process, so we don’t have a specific solution ready to share today. In the meantime, we’re working to address two consistent issues players have: 1. Ads don’t load, and 2. Ads don’t consistently reward actions. We will share results as we continue to experiment and appreciate continued feedback.

3. Authentication (new login)

Some players have shared continued issues with our login system. Functionality has been inconsistent across the platforms we support, especially on the Android app. We recently launched a new system based on Firebase Authentication (backed by Google), which has already been a much more stable experience for players. In this new system, players will now need to authenticate using email (we can’t support username login anymore). We have also ended support of Steam login.

If you need help figuring out your email because you normally logged in with your username please send us an email at [email protected] and include your username and anything else that can help us verify your account.

Steam purchases can still be linked to your AI Dungeon account to verify the purchase, and your Steam benefits can still be enjoyed on all platforms you use.

4. AI Quality

Better model quality continues to be one of the most requested improvements by players. We’ve recently put renewed focus on improving the models that power AI Dungeon and the data used to train them, which will help solve issues and improve the general experience for every player. Here’s what we’ve been working on and have planned:

  • (done) Wyvern as a new model
  • (done) Hydra mode for Griffin + Wyvern
  • (done) Improved AI Feedback
  • (beta) Doubled context length for Wyvern (2048 tokens)
  • (beta) New 3rd person finetune dataset
  • (alpha) Updated 2nd person finetune dataset
  • (planned) Story Styles

Story Styles will tune the AI to a specific genre (fantasy, romance, and historical fiction), fictional universe (such as those featuring superheroes, space adventures, or even vampire and werewolf love stories), or writing style (like Jane Austen or Mark Twain).

Not only will we be creating a set of Story Styles, but players will be able to create their own Story Styles. We can't wait to see what you come up with!

Wyvern-Hydra continues to perform as our leading model in both our Train the AI experiments and according to player preference. Please continue to submit feedback about the language model's performance using the AI Feedback button in AI Dungeon. Also, join our Discord server and let us know what your experience is like with the updates to Wyvern. We plan to keep making all of our models even better!

5. UI / UX Design

We are especially excited about updating the UI/UX of AI Dungeon. We will be using a component-based library called NativeBase which will provide a more stable app experience. And, given recent survey feedback, we will be making slight updates to the look and feel. Here are some of the early mockups for the designs we’ve been exploring (there will be changes before we launch, but this gives you an early idea of the direction):

We've begun work on this update and will launch it as a new beta instance for early testing when ready.

We look forward to any feedback and the next stage in our journey forward together.