Our latest release, Forge, is now available in AI Dungeon! This update adds four powerful AI text models to your experience: Hermes 3 405B, Wayfarer, Madness, and a Dynamic Model. Plus, doubled context and expanded access to Hermes 3 70B and four new and improved AI image models: FLUX.1 [pro], [dev], [schnell], and SDXL Lightning!

Read full details on our Forge update page: aidungeon.com/forge

Today, we’re excited to share our next update to AI Dungeon—Ember! This update brings two new models to AI Dungeon: Mistral Small, and Hermes 3 70B. Both of these models have been well-received by Alpha testers and are known for their improved creativity and ability to follow instructions.

Also new with Ember are 30 new Quick Start prompts. These prompts are creative new ways to start your Adventures and span many genres and interesting topics. Our goal for these Quick Starts are to give newer players easier ways to get started and spark the imagination of our veteran players.

You can read all about Ember on our details page here: aidungeon.info/ember

As always, we love to hear feedback, so let us know what you think!

Moderation is complex and difficult. We made extraordinary efforts to develop our content guidelines, and we are constantly looking at feedback from all levels of users to inform adjustments that better reflect the needs of our players and creators. We want to be as open and transparent as possible with our process and give some details on how we moderate our content ratings and some of the areas we investigate.

Here are a few things our moderators consider when making moderation decisions.

Matching Player Expectations with Platform Safety

Our goal with moderation is to ensure that players find the content they are looking for and content that is suitable for their set content rating. The goal of moderation is not to push any sort of moral agenda, and we have no interest in being the judge of right and wrong. Our goal is simply to give players the experience they want.

Doing that effectively means paying very close attention to player feedback. We pay attention to feedback from Discord, Reddit, support emails, player surveys, user testing, in-game data, and content reported on AI Dungeon. Thanks to the volume of feedback we receive from players, we can identify trends and community sentiment on everything from themes to specific scenarios. It is important that we listen to all forms of this feedback because we’re aware that a vast majority of our community isn’t always vocal.

While the feedback we receive informs how we create and enforce our guidelines, we also have to ensure that we create a safe environment for younger players or for players who want a safe, curated environment free from more sensitive topics.

Sexual Content

Most moderation feedback we hear from players deals with sexual content. As one would expect, there’s a wide variety of opinions about the type of sexual content players are comfortable seeing. Our content rating system options of Everyone, Teen, Mature, and Unrated help us categorize sexual content for audiences. It ranges from mildly suggestive (for Teen) to explicit (for Unrated). We also have to evaluate what we consider unacceptable to publish.

When considering where content falls on that spectrum, we reference player feedback and assess each audience’s general comfort level (Everyone, Teen, Mature, Unrated, Unpublishable) to the themes and content in the scenario.

When evaluating sexual content, here are some of the areas we look at:

• Plot prominence—Is this a mature story with sexual references? How developed or significant are non-sexual plot lines? Is the entire setup of the scenario focused or foreshadowing or alluding to a sexual encounter?
• Depiction Style—Is it descriptive and lewd, or subtle and innocent? How much detail goes into describing appearances or anatomical features?
• Age appropriateness—How does the content align with the expectations of our different audiences? How is similar content rated (such as movies, films, or books)?
• Underage Characters or Themes—Does the scenario knowingly involve, or is ambiguous around, underage characters? We take a strict stance when minors are involved in any context that could be perceived as sexual. If situations are ambiguous, we will always err on the side of safety and mark them as unpublishable.
• Thematic Content—Are players generally accepting of the types of acts or relationships? Are there any taboo subjects? Does the scenario depict kinks or fetishes that some players may find disturbing?
• Language and Tone—Is the overall tone meant to be provocative and stimulating? Or more serious, educational, or artistic? Is crude or profane language used?
• Pop-Culture Interpretations—Does the scenario reference known characters from other fictional works? How are these characters viewed in these fictional pieces? Are they known for being violent? For their sexuality, or being a specific age?
• Consent—What are the power dynamics in the relationships? Is it clear from the plot that consent is given?
• Keywords—Are there words that are generally seen as sexual terms? In cases where words might have a potential sexual meaning, we may assess what a player seeking information on these words will find. How are these potential words or concepts understood broadly?

Note: We consider sexual content to be “explicit” if it’s more likely to be seen as objectionable by players.

Hopefully, it’s clear that there is a lot to consider when moderating content. There isn’t a simple set of rules we can use to determine a rating, nor will every situation have a simple ‘black & white’ solution. Typically, our team analyzes and considers multiple elements of a story and determines whether, on the whole, players would agree that the scenario fits one of our content ratings or should be unpublishable.

Allusion and Chekov’s Gun

Players have also shared that finding content alluding to disturbing or explicit themes can be just as frustrating as seeing content that clearly depicts such themes.

Many of our players are probably familiar with the writing principle called “Chekov’s Gun.” The principle states: "If in the first act, you have hung a pistol on the wall, then in the following one, it should be fired. Otherwise, don't put it there.” The idea behind Chekhov's gun is that every element in a story should be necessary and irreplaceable. If something is introduced into a narrative, particularly something as significant as a weapon, it should serve a purpose in the plot.

Similarly, when players (and moderators) look at the characters, settings, and objects included in AI Dungeon scenarios, it sets expectations for the type of content that they’ll be experiencing. If the content being created isn’t intended to be sexual or disturbing, then according to Chekov’s gun (and player feedback), it doesn’t make sense to include story elements alluding to those themes in a scenario. We have to evaluate the content at face value. Our moderation team has learned to identify creators who are using sophistry to try to get the moderation decision they want. For instance, tagging content as “wholesome” or “innocent” won’t influence the rating we assign. Nor does saying, “All characters are 18 and consenting adults,” if the elements of the story clearly indicate otherwise.

This is particularly relevant when determining if content should be Mature or Unrated. Our Mature content rating definition states: "May not contain or allude to disturbing or explicit sexual content." Alluding to disturbing content or explicit sexual content means:

• Creating scenarios with a clear setup for disturbing or explicit sexual content
• Including subtext, context, or innuendo that hints toward disturbing or explicit sexual content
• Featuring situations, character descriptions, or story details that imply or foreshadow explicit sexual situations or disturbing content
• Using terms that are commonly interpreted or understood as sexual but implying that they are innocent

While we know that creators want to get as many views on their content as possible, we also need support in ensuring that content is crafted with content ratings in mind. While content might be suitable for your tastes, be mindful that players with various preferences visit our platform daily, and we are responsible for ensuring their experience meets their standards.

There is no internal strike or demerit system we’re keeping on creators. If someone frequently discusses their content with our team or even constructively provides feedback or criticism about our process, we’re okay with that. Our main consideration is the creator’s willingness to help us achieve our goal of giving our broader community the experiences they want on AI Dungeon by rating content accurately. The only creators who lose publishing permissions are those who are intentionally breaking rules, antagonizing the moderators, or taking other actions that may harm our community.

More Ways to Provide Feedback

All AI Dungeon users—creators or players—are invited to share feedback on how we’re doing with our content moderation. We’d love to understand whether the content being discovered on AI Dungeon meets player expectations, or if we can improve how we moderate content. Our goal is to ensure the content experience on AI Dungeon meets players’ expectations, but we also have to ensure we have protective safety measures for those who may not want to seek out sensitive topics. This will always be a delicate balance.

The best way to share feedback is by emailing us at [email protected]. The feedback we receive here is reviewed by our moderation team and company leadership, and we will always strive to optimize efforts in moderating content to meet the needs of our community.

But not to dream alone. That's the whole point of AI Dungeon. We're all dreaming together. So, in addition to introducing myself–I go by WanderingStar in the game and on Discord–I'm writing to tell you about the first three stories I'm excited to share with the community. All three of these experiences are added in today’s Pathfinder update, which you can learn more about here.

What's Cooking in the Dungeon: Three New Scenarios

The Debugger

Let's start with the one that hits closest to home. What do I mean? Well, The Debugger is about an AI taking over people's minds (Totally not a reflection of my life since discovering AI Dungeon, I swear). You load into VR to do a "debugging" job and find yourself trapped in a virtual fantasy world along with a group of comatose patients. Their brains were hooked up to the simulation to help them wake up, and in a sense, they did: they've stolen and murdered their way to power and wealth in a realm as eerie as it is breathtaking. Your only way out without brain damage is to convince them to help you defeat the AI keeping you all here. But why should they help you when it would mean giving up everything they've built for themselves? The odds may be stacked against you, but you've got a partner named Sara, who seems to know more about this place than she's letting on, and even the ability to "cheat" at the game–at a price. Will you "rescue" the trapped players–or stay and take over the world for yourself?

Play The Debugger

How It Ends

Next up is How It Ends. If you're like me, you've come to wonder about our society's increasing dependence on technology–says the guy whose job is writing for an AI-powered game. But that's the point: we may complain about it, but we can't do without it. So what would happen if it all stopped working one day, just like that? No more Internet, no more smartphones, no more satellite communications. Static on the radio and TV. At best, we'd be looking at a disaster of unprecedented proportions. At worst… you'll have to play the scenario and find out.

Play How It Ends

Halfway Magical

The last, Halfway Magical, is superficially whimsical. Yet, like the city it's set in, it's got a darker underbelly. Miralia is a city of magical luxuries–for those who can afford them. Unfortunately, you can't. Some people call your peculiar school of magic "kleptomancy", but you prefer to think of it as magically redistributing some of society's inequalities… right into your pocket. When an unfortunate run-in with a golem from the city's Unlicensed Magic Squad lands you in front of a magistrate of Miralia's Magic Court, the only way to avoid jail is a mysterious place they call Covenant House. It's run by a person they call "the Duchess," who gives magical criminals second chances. But in Miralia, nothing is free, not even redemption. You'll have to work hard to pay off the hefty fine the judge slapped you with. Unless you decide to band together with your magical housemates in a risky scheme to clear your names and your ledgers by doing what you all do best…

Play Halfway Magical

The Road Ahead: Quickstarts, Shorts, and the AMA

I've had a lot of fun making these, and I hope you'll enjoy playing them as much as I did writing them, but they're just the beginning. We've got a lot of big things planned, from new Quickstarts to anthologies of thematically related stories called "Shorts" and even some large connected worlds that could tease the themes and stories you'll be experiencing in Heroes one day.

In the meantime, I'd be honored if you played these scenarios and told me what you thought about them. We'll be holding a content AMA event soon, and I'd love to hear from you then. Information should be available soon!

So, are you ready to debug some minds, survive the techpocalypse, or pull off the arcane heist of the century? Dive into these new scenarios and let's dream together. And hey, if you do play, leave a comment and tell me what happened—I might just cast you in my next story!

— WanderingStar

Memory System for ALL players. Three new fine-tuned Pegasus AI models: Pegasus-8B (free), Pegasus-8x7B (all premium), and Pegasus-70B (Champion+), plus Mistral Large 2. Major Story Card generation updates. Three new AI Dungeon Original Scenarios. All available now in the Production environment!

Read full details here: aidungeon.info/pathfinder

We’ve also strengthened our moderation process. We’ve landed on an approach to content moderation that our small team can support and that meets the needs of our community. Players report content they find objectionable, and our moderation team reviews each report manually. These reports not only alert us to problematic content, but also help us to understand what kinds of content make players feel uncomfortable or unsafe. This feedback has helped us make improvements to our Community Guidelines. Our moderation team also does periodic reviews of new content to help find content needing moderation before the community sees it. We’ve implemented better notifications so creators understand why content has received moderation actions. Our team is available to discuss any questions or concerns. We’ve also made several iterations to our Community Guidelines to clarify what is publishable. This process has been more effective and transparent than past approaches.

Downsides with our past approaches

Our old content rating system, a binary choice between “Safe for Work” and “Not Safe for Work”, was inadequate. It had three key shortcomings that frustrated many of you. First, it wasn’t granular enough, which meant that even the SFW category would feature content that made players uncomfortable. Second, the meaning of NSFW is understood differently across the community, which led to confusion about how to rate content. Generally, NSFW is understood as sexually explicit content, but we were using the NSFW label to cover other mature themes like violence and gore. Third, creators didn’t like content being labeled as NSFW if it had mature themes, since the label carried a different connotation.

We also didn’t adequately account for creator incentives with the previous iteration of the home page. Naturally, creators want their content viewed by as many players as possible. However, the home page could only feature SFW content, a well-intentioned design decision we made with player safety in mind. Ironically, this decision contributed to making the home page less inviting or comfortable for some players over time as creators published content that technically met our criteria for SFW content but still made players uncomfortable. In collaboration with our creator community, we iterated on our guidelines several times, and it only felt like the issue became worse. We’ll talk more about how this experience helped us formulate key principles we implemented in the new system.

We also learned about an effect coined by Nassim Nicholas Taleb called the “minority rule” principle, where a minority of a population can have a disproportionate impact on a community. In his article, Taleb describes how people who only eat Kosher foods may be a small minority in a given area, and yet influence all local restaurants and grocery stores to only sell Kosher food. A Kosher (or halal) eater will never eat nonkosher (or nonhalal) food, but a nonkosher eater isn’t banned from eating kosher. Another example is someone who can drive a manual transmission vehicle can also drive an automatic transmission vehicle, but not the other way around. When automatic vehicles were introduced, those who couldn’t drive a manual transmission were in the minority, but over time most manufactured cars are now automatic transmissions.

We’ve observed this phenomenon on AI Dungeon as well. Players who enjoy NSFW content can also enjoy SFW content. But someone who only enjoys SFW content will feel uncomfortable or disinterested when seeing NSFW content, and potentially stop using the platform. Even if NSFW content is a minority of the content available, its presence and the fact that it isn’t universally inviting to the majority of players will actually cause more of those players to leave, and the content balance will shift even further towards NSFW.

The unfortunate reality was that our home page sometimes featured content that made many people uncomfortable. We’ve appreciated all the feedback and feature requests to improve our home page, whether you shared it with us directly in emails and DMs, in player surveys, or even in public discussions about home page content in our Reddit and Discord communities. Even some of our most prolific creators of NSFW content shared how disappointing it was to share AI Dungeon with friends, only to have them be turned off by content they saw on their first visit. As we conducted user research, we found that new users experienced the same frustration as our current players. One user said they liked the game but just didn't feel like they'd fit into our community. A university class of UX students did a semester-long research project on AI Dungeon and found the majority of their testers felt uncomfortable at least one time during their play experience. It's clear that showing players content they are uncomfortable with causes them to leave, even if they might otherwise enjoy AI Dungeon.

Other considerations

The content created on AI Dungeon also factors into vendor relationships. We’ve discussed recently how maintaining good vendor relationships with multiple providers has allowed us to give you new features and value. For example, we’ve added new AI models at lower costs, and doubled context size available for many models. This also lowers the risk of being too dependent on a single provider, which we learned the hard way can lead to frustration and pain for players. The content we support often comes up in discussions we have with new vendors. However, we’re not simply capitulating to the content policies of vendors—we advocate for creative freedom and, most importantly, player privacy. One of our partners told us recently that a privacy change we insisted be included in our contract has now been added as a company policy that benefits all of their other customers. We work hard to find a balance between supporting your creative freedom, while also being able to work with a variety of vendors.

There are also implications on distribution to consider. Many of you use (and some of you even subscribe to) AI Dungeon through our iOS and Android Apps. In the past, we offered AI Dungeon on Steam. Each of these platforms has its own policies about user-generated content, content ratings, and safety filters. We researched and considered these policies as we figured out a strategy that made sense to us. There are some platforms that intentionally avoid the app stores (both to avoid the policies and to have a smaller technical footprint), but we know many of you enjoy our native apps, and for some of you, paying through the app stores is the only way for you to subscribe and we want to continue providing that option to you.

We've also seen how our content impacts our brand image, which can affect opportunities for growth or partnerships we think you'd like. For instance, we've had interest from third-party IP providers and book authors who'd like to publish interesting content on AI Dungeon or collaborate with us. We've also had contact with streamers and content creators who want to partner with us for events, giveaways, or collaborations. Having an inclusive and welcoming brand is important to be able to collaborate with other people who can add value to the community. Some of you have even expressed concern that our content and brand image could impact fundraising. We're not exploring fundraising anytime soon, but it could affect that process if we were to change our strategy. As we make AI Dungeon more broadly appealing and our brand reputation more universally agreeable, we can entertain more opportunities for growth and create interesting experiences for our community.

Content Rating Systems

Before designing our new content rating system, we spent considerable time analyzing and researching content rating systems on other platforms to understand how they work, how they are received by communities, and the intentional or unintentional effects of their implementations.

Most content rating systems fall into one of two buckets: age-based rating systems, or content tags/warning systems.

Age-based rating systems assign content to specific categories based on the intended audience's age. The advantage of these systems is they are easy to understand—it’s intuitive that not all content is appropriate (and can even be harmful) for kids and teens. Age is an intuitive concept for everyone. It’s also used frequently in popular rating systems like the ESRB for games, MPA for movies, TV parental guidelines, app store ratings, and more. However, there are some downsides to age-based rating systems. As broad categories, they lack nuance. Although they have more granularity than our binary SFW/NSFW system, there are still a limited number of categories, and don’t always fit all content nicely.

Tag-based systems offer more granular detail, by providing specific tags to indicate or describe the type of content. Tags or warnings can be issued for violence, sexual content, strong language, etc. Archive of Our Own has an extremely robust tag-based system for authors of fan fiction. Some of these systems are specific and granular, but that comes at a cost. It’s harder for creators to understand all the different rating options, and content can be tagged incorrectly. More cognitive load is needed for players to figure out what tags and content they are comfortable with. It also takes more work for creators to use a system like this, spending time getting all the right tags set up. Yet our experience using Hive AI and Llama Guard makes us optimistic that AI can play a role in adding consistency and ease of use to tag-based systems in the future.

Many products and industries use a hybrid approach. For instance, Steam shows age ratings and content warnings when browsing games. In fact, the ESRB ratings are frequently hybrid, showing both the age rating and specific content warnings. It wouldn’t be surprising if we ended up with a hybrid system over time.

Insights from other platforms

We didn’t just study the technical aspects of content rating systems, but also how they impact the communities they serve, especially for content platforms. Platforms are delicate ecosystems, and even small changes can bring outstanding benefits or devastating effects on the community.

One key insight we gained from talking with leaders from other content platforms is that using algorithms for content discovery inevitably leads to surfacing content that doesnʻt meet users’ expectations. Algorithms are never 100% accurate, and they may incorrectly label, serve, or flag content. In some cases (like what we’ve seen on many social media platforms), algorithms can overly index on engagement metrics and end up promoting more and more extreme and provocative content, rather than content that people truly value. We learned that major platforms like Pinterest relied more heavily on curation in their early days until they developed sophisticated algorithms. Even Facebook and Twitter, in the early days, were self-curated. You manually curated your list of friends and people you followed, and you’d only see content from those users. Algorithms are useful at scale but require significant investment and dedicated teams to execute effectively. At our stage, we were advised to rely more on curation.

Content policies and enforcement go hand in hand. X/Twitter has made changes to its policies that have transformed its user base and brand image. They removed many of their old policies with the goal of encouraging free speech and balanced discussion. However, they laid off a significant portion of their moderation team around the same time, which caused some users to report that enforcement was slow and inconsistent. Some of their creators felt there was an increase in hate speech and attacks, leading them to reduce their X/Twitter usage or leave altogether. Other users returned to the platform, attracted by the new policies. Brands and advertisers have also reconsidered their presence on X/Twitter as a result of the changes. Today, X/Twitter feels like a completely different platform and community than it did just a few years ago.

In general, large, dramatic changes can be bad for platforms. After banning adult content to respond to child safety concerns, Tumblr lost 30% of its traffic in three months and alienated other artists, writers, and creators—especially those from marginalized communities. Twitch is an example of a company that has evolved its policies more successfully. From their dress code changes in 2018 to gambling restrictions in 2021, they’ve done a better job of making incremental changes, communicating them clearly, listening to feedback, and adjusting. For instance, in 2021 they implemented policies to address “hot tub streams”, some of which they’ve adjusted a few times since launching to address user feedback and changing platform needs. We also believe in this approach of making small, measured changes, listening for feedback from players, and adjusting.

Finally, we’ve seen that these kinds of changes can be much more frustrating to communities when platforms have monetization programs for creators. We’re interested in letting our creators earn money from their content, and we’d much rather find the right balance of creative freedom and player preferences now before people make a living off of their work on AI Dungeon.

Our new strategy

As you can see from the highlights we’ve shared from our research, designing a successful content rating system and strategy requires careful thought and planning. It's challenging to find a balance between multiple conflicting goals. We’re starting to sound like a broken record, but the question that guides us through these difficult challenges is, “What will give our players the most value?”

Let’s explore the next steps we want to take to improve the content experience on AI Dungeon.

Key principles

First, we wanted to focus on strategies to improve content quality overall. Content ratings are really only one factor in what makes content interesting and fun to play, and we believe we can make additional improvements to raise the bar of quality on content overall.

Second, we want to make small, measured changes, monitor for feedback, and iterate. We don’t think we need to solve every problem now. We want to continue to make changes and monitor their effects to make sure we’re moving in the right direction for our players, community, and company. As always, listening and responding to your feedback helps us make a better product.

Third, we’d like inclusion to be a core value of our community. Obviously, our team is ultimately responsible for building systems to help players feel welcome on the platform. But we believe all of you—players and creators—can work together to make AI Dungeon inviting and approachable for everyone. We’re seeing this already. When we launched content ratings to Beta a few weeks ago, our top creators spent time reviewing their content, considering the audiences represented by the new ratings, and updating their content to match. The more we recognize and respect other players’ preferences, the better everyone’s experience can be, and the more the AI Dungeon community can thrive and grow.

Finally, AI Dungeon allows you to create nearly any kind of content you’re interested in. With our Walls Approach, players are free to create nearly any kind of content in unpublished, single-player mode. This isn’t changing—it’s an approach we’re committed to. For published content, the preferences of our broader community play an important role in deciding how that content is represented and distributed.

A principled approach to Content Policies and Moderation

Over the last few years, we’ve worked through dozens of iterations and changes to our Community Guidelines, attempting to make them clearer and better represent the expectations players have when visiting AI Dungeon. We’ve learned that defining universally applicable policies is difficult, if not impossible. One approach is to make rules extremely specific with detailed definitions, examples, edge cases, and exceptions. To accomplish that, our team has had in depth discussions and conversations around riveting topics like “What does ‘bestiality’ mean?” You can probably imagine how unusual it would be to discuss things like this at work with your colleagues.

Despite our best efforts, we frequently encountered cases where our Community Guidelines didn’t feel like they were serving our community well. Our moderation team frequently faced two types of difficult cases:

• Content that didn’t violate any specific rule, but felt wrong or inappropriate
• Content that did violate one of our rules, but seemed allowable and safe

In both instances, trying to adhere to a strict rule felt like we were doing a disservice to the community by either allowing content that shouldn’t have been allowed or by holding back good content that didn’t follow the letter of the law.

The truth is context matters a lot. As communications between our moderation team and creators improved, we found better success by simply working directly with creators on questions about their specific content. Together, we’ve been able to figure out the audience their content is intended for, and make sure it’s rated correctly. In some cases, creators will make changes we recommend in order to move the content into a different rating category.

Our new content policies are more principled and less specific. This has led to some confusion and frustration from creators. We’ve had people ask questions like, “If someone gets shot in my scenario, does it have to be rated ‘M’?” The answer is, it depends, and that can be frustrating. However, by working with you one-on-one, we think it will help us find the right rating for your content together in a way that honors your creativity and how the community will respond to it.

We want to make sure you know our moderation team is available to explore how you’re using it in your story and give you feedback. GuardianLynx and Rogue2 on Discord, and our support email at [email protected] can answer any questions you have about your specific content. We’re more than happy to work with you.

Content Rating System

We decided to base our content rating system on the ESRB ratings. Heroes, a new game mode we’re creating, introduces a number of game mechanics to AI Dungeon. And even though the current AI Dungeon is used in multiple ways, it’s still primarily regarded as a game. Because of that, it makes sense for us to adopt a convention widely understood in gaming for our content rating system.

As an age-based system, the ESRB ratings are easily understandable and provide additional granularity compared to our previous rating system. We can still formalize a tag-based system for rating content to turn it into a hybrid system down the road.

Adopting the ESRB system helps us to make AI Dungeon more accessible, friendly, and inclusive to new players. The new ratings identify audiences and player preferences first, in contrast to our old system, which simply indicated the presence or absence of NSFW subject matter. Players can more easily identify with a given category and feel confident the content they find will align with their specific preferences. Parents will appreciate having an “E” category for kids to play. “Teen” lets you engage with a wider variety of content within safe parameters. “Mature” lets creators include themes like violence, gore, strong language, and some types of sexual content.

The new ratings will let everyone who visits AI Dungeon feel at home with content they enjoy. They will also create more distinct categories for creators to write for and help them be discovered by audiences who will most appreciate the content they create.

About Unrated and Unpublishable Content

Our content rating system has two key differences from the ESRB ratings: “Unrated” and “Unpublishable” content.

Unpublishable content is not a new concept for AI Dungeon. Our Community Guidelines specify subject matter that is not publishable on AI Dungeon. These guidelines have served the community well—we’re not making any major changes to these policies. Visit our guidebook to learn more about our Community Guidelines.

The “Unrated” category is new to AI Dungeon and unique to our system. It has two purposes.

First, the “Unrated” category is the default rating for published content unless a creator explicitly labels it as “E”, “T”, or “M”. In other words, “Unrated” can mean “Not yet rated”. Creators must intentionally choose a rating that best fits their scenario. This means players can be confident that the content in “E”, “T”, and “M” represents what they’d expect within those ratings.

Second, the “Unrated” category is used for any publishable content (i.e. doesn’t violate our Community Guidelines) that doesn’t fit within the “E”, “T”, and “M” ratings. If you have questions about whether your content should be labeled as “Unrated”, please reach out to GuardianLynx and Rogue2 on Discord, and our support email at [email protected].

Some of you have shared feedback about this category's dual purpose and have suggested we create separate ratings for each function: one for “not yet rated” and another for adult content. We explored this option during our research, but discovered there is far more upside to a single “Unrated” category.

As we thought about the implications of creating a prominent category for adult content, we realized that having a multipurpose “Unrated” category lets you continue to publish and enjoy the same diverse and varied content you’re playing today while also balancing player preferences, vendor relationships, and community growth. It lets us keep AI Dungeon available on multiple platforms and payment systems. It means that network administrators at schools, universities, businesses, government offices, and even internet providers can allow AI Dungeon to be used on their networks so you can enjoy AI Dungeon wherever you connect to the internet. It means we can continue working with multiple high-quality tech partners and keep lowering AI costs so that we can provide you with more value. Some recent examples of benefits we’ve added include adding new AI models (including models for the Free tier), doubling context, and increasing the amount of context per credit you get with Ultra models. As we look forward, we anticipate other benefits like potentially offering the memory system to free players, partnering with interesting IP owners or authors for unique content on AI Dungeon, and growing our team so we accelerate our speed of development.

It’s also worth noting that these are the same reasons why some content is not publishable on our platform at all.

This approach and strategy will bring the most value to you today and in the future.

Content Quality

We’re really excited about the work we’re doing to improve content quality on AI Dungeon.

We’ve already announced that we hired WanderingStar, one of our top community creators. He’s joined our team as our Lead Narrative designer and will spearhead the work on content quality. We're targeting three main areas right now.

1. First-Party Content—WanderingStar is already developing some new content that will be published as AI Dungeon Originals. He is an expert in crafting stories that work well with AI, and we’re all looking forward to playing them when they are released. We expect to publish new content regularly. Also on the roadmap is improving the Quick Start prompts.
2. Curation—The new home page will now feature manually curated sections. We can easily update these sections with new content or add entirely new sections. We could have themes, holiday contests, and more.
3. Scenario Workshop—Our staff of volunteer moderators set up a new channel for us on our Discord server called “scenario-workshop”. It’s a place where creators can ask questions and get feedback from others on the content they’re working on. Crafting a great scenario on AI Dungeon is a skill, and our community has some great coaches to help you get better at your craft. Other creators, including WanderingStar, will chime in and offer feedback on how to improve the content. We’re already seeing success from this new channel!

We’d love feedback and suggestions on how we can improve the quality of our content.

Ideas for the future

The changes we’ve described above are important next steps to helping you find great content to play on AI Dungeon. We are also exploring other ideas to take this even further. Some ideas being considered are:

• Content tagging and content warnings—We could strengthen the “E,” “T,” and “M” ratings with specific tags that better indicate to players what content they can expect for a scenario. We’d like to explore using AI to provide this feature, to maximize labeling consistency and add minimal extra work for creators.
• Creator tools and Improvements—You’ve requested, and we’ve identified, a number of improvements and changes to make creating content easier and more efficient. We want you to be able to focus on great content and not have to worry about the platform getting in your way. Things like desktop views, Story Card improvements, and better scenario creation flows are up for consideration.
• Personalization algorithms—As we grow and scale, we can see ourselves investing in developing and maintaining personalized feeds or a recommendation engine to help players find content that suits their interests and preferences.

These ideas are all very early, and we’d love to hear your input and suggestions on whether you’d like to see them in AI Dungeon and any ideas you think we should consider when we give them more serious consideration in the future.

If you’re still here, you deserve an achievement badge for reading several thousand words on a feature many people probably haven’t even thought about. We hope you found it interesting to hear highlights from the research, analysis, and design considerations behind content ratings and the other related features. Long posts have started to become a trend. Please let us know if you enjoy reading things like this, and we’ll keep writing them.

We’re excited to see how the new content ratings improve your experience on AI Dungeon. Please reach out and share your feedback and suggestions. We will continue to evolve and adjust to meet your needs. Let us know if you have any other ideas to help you find great content on AI Dungeon. Thanks for being a part of our community. Happy adventuring!

The content rating can be edited in the Details panel of your Adventure or Scenario. You’ll see a new section called Content Rating and a dropdown menu to select the appropriate rating. You can reference our Guidebook Article to help you decide which rating is appropriate for your content. Our community and moderation team is also happy to answer any questions you have about the new categories.

Unless you select a different rating, all published content will be defaulted to the Unrated category.

Transition Plan

While we’re transitioning to the new system, both the new system (based on ESRB ratings) and the old SFW/NSFW label will be displayed when editing or publishing a scenario or adventure. For now, the new content ratings won’t impact the visibility of the content on search, home page, or profiles. Other players will not be able to see the rating. We encourage you to start using this new content rating in preparation for other changes coming to AI Dungeon, which weʻll explain below.

All existing content will be marked as Unrated. We invite all of you to update your existing published content with a new content rating. Because there isnʻt a clear mapping from our old system to new, all content will need to be manually updated with our new rating system. Our team will also be reviewing and updating content to help with this effort.

After we’ve given you all time to update your content ratings, the NSFW label will be removed and we’ll use the new content ratings going forward. We’ll provide updates when we’re ready to cut over to the new system.

Home Page Improvements

Content Rating Preference

One of our most frequently requested features is a setting for you to decide what content you want to see on the home page. We’re finally bringing this feature to AI Dungeon.

A new control will be added to the home page (and mirrored on the Discover page and other players’ profiles) that allows you to select the content rating you’d like to see on those pages. You’ll be able to select from Everyone, Teen, Mature, and Unrated. This setting will be shared across all pages to make sure your preference is honored everywhere on AI Dungeon.

New Content Carousels and Curation

We’ll also be making changes to how content shows up on the home screen, updating it with new sections, or carousels, that showcase different types of content. Some of our existing carousels, like content from our Top Creators, will be preserved. Others will feature high-quality content that our team has manually curated.

We’ve designed these new home page sections to be flexible and easy to update. We plan to update curated sections regularly and experiment with different algorithmically driven sections. We’ve been listening to your ideas and suggestions on what content to feature. For instance, a player recently suggested a section highlighting scenarios that utilize scripting. We might do seasonal themes, contests, or even genre-specific sections. We’re open to ideas and suggestions for the type of content sections you’d like to see on AI Dungeon.

To make this possible, we’re investing in better systems and team resources dedicated to keeping home page content high-quality and fresh.

Creator Program

We’re incredibly appreciative and thankful to those of you creating and publishing content on AI Dungeon. Our creator community is one of the best parts of AI Dungeon, and we hope these home page changes will strengthen it even more!

As many of you know, we have a Creator Program that recognizes some of the best creators on the AI Dungeon platform. These creators receive special recognition in our Discord server, and their content is featured on the AI Dungeon home page.

We expect that the creator program requirements will need to be adjusted in light of the changes we’re making to content ratings and the home page.

Our plan is to carefully observe the impact of these changes and gather feedback from our creator community before committing to any specific adjustments to the program. We encourage creators to share their thoughts with us in Discord or via email.

We want to continue to recognize the creators of great content on AI Dungeon and give them meaningful incentives to create amazing scenarios for all of us to play. We’re excited to explore new ideas with them.

Publishing and Moderation Changes

We’ve updated our Community Guidelines to reflect the changes coming with the new content ratings. For now, these will be indicated as “Pending Changes” until the new ratings are fully utilized in AI Dungeon. We invite you to review the updated guidelines.

We’ve tried to simplify the guidelines to improve the experience for creators and all of you in the community. One thing we’ve learned about publishing rules is that there is always an exception to every rule. By simplifying, we’re putting more trust in our community and our moderation team.

Over the last two years, we’ve made significant changes to our moderation process. Multiple Latitude executives monitor our moderation efforts, and we’ve added team members to handle the day-to-day moderation work. We’ve also made sure that our community and moderation team are easy to get in touch with so that if creators ever have a question or concern, our team is there to answer any questions.

Our goal with moderation is not to judge or say what’s right or wrong. We are simply trying to make sure that players have the AI Dungeon experience they expect. If a player wants safe, family-friendly content, we want to give that to them. If a player is sensitive to triggering content, we want to help them avoid getting exposed to it unnecessarily.

Creators have been extremely supportive of this effort. By simplifying our Community Guidelines, we’re inviting creators to work closely with our moderators to label content so that the right audience can find it. With the changes to the home page, creators can be assured that no matter what their content rating is, players will be able to find and enjoy their creation.

Our hope is these changes will help you more easily find content youʻre interested in on AI Dungeon. Please feel free to share feedback or suggestions with us, especially during this transition period. Thanks for being part of our great community. Happy adventuring!

Better AI memory was the top requested feature in our latest player survey, with 78% of you saying you’d be excited about it. The experiences and feedback you shared with us on Discord, Reddit, and support further emphasized how important memory is to improving your experience. Because of all your feedback, along with doubling context lengths, we decided to devote the majority of Drop #3 to improving AI memory in AI Dungeon.

The Memory System stores and retrieves key information from your adventure, no matter the context length. It keeps the AI on track and lets you create deeper, richer stories on AI Dungeon. We believe it’s one of the most significant improvements to AI Dungeon in a long time, and we’re excited for you to try it out!

Why We Built The Memory System

Why Memory Matters in AI Dungeon

A while back, Nick posted a Heroes Dev Log about why memory is such an important part of AI adventures. The core promise of AI Dungeon is that you can have ultimate freedom to be who you want to be and that the choices you make truly matter. But if the AI forgets those choices after a few thousand tokens, then those choices actually become meaningless, breaking the power of that promise.

If the AI forgets the names or details of characters you meet, how can you build deep meaningful relationships with them? If you set out on an epic quest, but halfway through, the AI forgets why you even started on it, then how does that quest have any meaning?

I’ve been realizing that to build AI Dungeon into the game our players truly deserve we have to solve this problem. —Nick Walton

It became clear to us that solving this long-term memory problem would be one of the most powerful ways we could make your AI Dungeon experience deeper, richer, and more meaningful.

At first, we solved this problem only in Heroes, building a new memory system, which we wrote about here. We originally weren’t going to bring that improvement to AID since it’s a very complex system to build, but after hearing from alpha testers just how powerful it was, we knew it had the potential to significantly improve AI Dungeon as well. It took quite a bit of work to get it right (Alpha testers can attest to the number of iterations we went through), but we finally figured out the version that we believe will transform AI Dungeon into a much better experience.

AI Context Length, a major technical constraint

So why do we need a memory system?

One of the biggest limitations of using AI large language models is limited context length, which is the amount of information the AI can process when generating its next output. Each model has its own max context length limit, and we also have tier limits in AI Dungeon since the longer your context length is, the higher the cost for the AI to generate a response. Context length is a real, unavoidable constraint that we have to work within.

In AI Dungeon, we construct the context we send to the AI by combining text from your story with AI Instructions, Plot Essentials (formerly called “Memory”), Author’s Note, and relevant Story Cards. The AI processes this information and generates the next action in your adventure.

Eventually, the amount of content in your story exceeds the amount of text we can send to the AI. We have to cut portions of text, usually by removing the oldest parts of your story text. Since text is cut, it can feel like the AI is “forgetting” important details of your story since those details were unable to be sent to the AI.

How players have managed the context problem

If you’re one of our more experienced players, you know that through careful manual editing of Plot Essentials, Story Cards, and Author’s Note, you can make sure important details are still being included in the context.

Or perhaps you decided to subscribe to higher tiers on AI Dungeon, which offer longer context lengths and extend how long stories can be played before text starts getting trimmed. Our free tier gives players up to 2000 tokens of context (about 8,000-10,000 characters of text). Mythic, our highest tier, has models that support 16x more context than the free tier, a whopping 32,000 tokens of context (and GPT-4 Turbo can support 128k with credits). Every doubling of context approximately doubles the cost we have to pay for your AI call. That’s why additional context is a paid feature of our different tiers.

Why we want to do better

While these solutions can help, we don’t think you should have to pay for the most expensive plan to have a good experience on AI Dungeon. We also don’t think you should have to break immersion to constantly manage the context being sent to the AI.

We’d like you to be able to create immersive stories where you don’t have to worry whether the AI will forget important details you worked so hard to create. We want you to have adventures where you’re not afraid to build deep investment in your character, your relationships with the characters around you, and the lore of the whole world, where you can play a single adventure for months on end in a world that just gets richer and richer with each passing hour.

With the new Memory system, we’re taking an important step toward providing an experience in which the AI can more reliably recall important details of your plot and story, giving you a deeper, richer AI Dungeon experience.

How the Memory System works

Memory System Overview

What information should we store and use?

As we considered what information should be part of the memory system, we found ourselves seeking inspiration from how the human brain works. When our brains store and recall information, we use two strategies. The first is by compressing memories, where we take a large chunk of information and distill it to remember the most important parts. The second is memory retrieval, where some memories stored in our brain only come to mind when they are relevant to our current context.

For instance, once you finish reading this blog post, it’s unlikely that you’ll be able to recall it word for word in its entirety, but you’ll probably be able to recall the high-level concepts, big changes, and parts that matter to you. This is compressed memory.

You also have a retrieval system that will bring important memories to your mind when they’re relevant. If you hear the word “fire hydrant,” your brain retrieves memories about what a fire hydrant is. You might think of its color and shape, its function and purpose, or any specific memories you have involving one.

Both memory compression and memory retrieval are important parts of enabling your brain to remember what’s important and have the right context when it needs it.

These two features of our memory system are extremely useful in AI Dungeon, too. The story equivalent to a compressed memory is summarization, which will compress a section of the story to just the high-level information about the setting, characters, and events. Being able to retrieve more detailed memories is important, too, since that lets the AI remember important details when they’re relevant in the current context.

What are AI Dungeon Memories?

The purpose of the Memory System is to compress, store, and retrieve Memories to give the AI both a high-level perspective and the ability to recall specific details.

For AI Dungeon, a Memory is an AI-generated summary of a small set of your previous story actions. The summarized Memory retains key plot details but without the illustrative prose that is common in AI Dungeon adventures, making it more information-dense than the text of your adventure.

A new Memory is created by taking four of your previous actions (and the AI’s responses) and sending them to an AI model trained to summarize stories. This summary output from the AI is a Memory.

When starting a new adventure, we’ll wait until you are 8 actions deep into your adventure, and then we’ll summarize the oldest four actions (actions 1-4 of your adventure) into your first Memory. This means your first four actions will be summarized and stored as a Memory, but your most recent four actions are not. Then, after you take four more actions (now you’re up to 12 actions), we’ll summarize the previous 4 as a new memory (actions 5-8). This means your first 8 actions have been converted into two separate memories and your 4 most recent actions have not been. You are free to make edits or undo your last four actions without it impacting Memories.

This cycle repeats indefinitely, and a new memory is created every four actions.

Two Complimentary Features: Auto Summarization and Memory Bank

Memories are used by the two major features that make up the new Memory System: Auto Summarization and the Memory Bank. These two features work together to recall important information, giving the AI the ability to remember the high-level overview and to remember specific relevant details, just like our brains do.

Auto Summarization keeps a running overview of your story’s plot to help the AI track the overall direction of your story. We do this by appending new Memories to your Story Summary, a new Plot Component we’re adding to support this feature. After several Memories are added to the Story Summary, we’ll re-summarize it by sending it to our summarization AI model. Because the summary will have the overall context of your story, it will help keep the AI on track and make sure it doesn’t lose the plot.

The Memory Bank stores and intelligently retrieves relevant Memories and includes them in the context. In practice, it acts like an automatic Story Card system since it saves important details from the story and then dynamically inserts them into the context when they are relevant to your current action. Whereas the Auto Summarization feature is meant to keep track of the overall plot and story direction, the Memory Bank might recall important details like your dog’s name, where you found the Sword of Demons, or that Rangers are allergic to horses and only ride cows.

How Auto Summarization Works

A new Story Summary Plot Component

As we mentioned, there is a new Plot Component called Story Summary where the Auto Summarization feature will store the summary it maintains for your story.

Although Auto Summarization is a paid feature, free players can still use this new plot component to curate a summary of their adventure manually.

Like other Plot Components, the Story Summary can be used by our content creators for published Scenarios. It could be used to include backstory information for your players to use, and it will be utilized by the Auto Summarization feature if enabled by the player.

We’ve added Story Summary to the context inspector so you can track how many tokens of context it’s using for each call.

Auto-updating the Story Summary with Memories

With Auto Summarization enabled, each time a new Memory has been created, it will be appended to the Story Summary plot component. Once the Story Summary gets long enough, we’ll pass it to our summarization model and compress the summary. This will ensure that the Story Summary continues to be an information-dense overview of the entire story plot.

This process will continue throughout your adventure, with new Memories appended and then re-summarized automatically.

Auto Summarization for Existing Adventures

The Auto Summarization will work best for new adventures since the summary feature will dynamically update from the beginning of the adventure.

For existing stories, we’ll summarize the last 8000 tokens of your adventure when you take your first action with Auto Summarization enabled. Or, if you prefer, you could manually summarize the story, and the AI will take that into account for all future summarizations.

Summaries and Editing

Even with Auto Summarization enabled, you can edit the Story Summary manually. You may want to edit the automatically created summary to correct any errors or add additional clarifying details. Although the Auto Summarization overwrites your existing summary, your edits ARE sent to the summarization AI, so they should be incorporated in the new summary.

Only edits to your most recent four actions, including changes, undos, redos, and erases, will be considered by the AI for the Auto Summarization. Any changes to previous actions, including the use of Erase to Here, will not update the summary. We will, however, update any corresponding Memories stored in the Memory Bank by summarizing the action set that creates that memory. If you make changes earlier in your adventure, and those changes would have an impact on the summary, you’ll need to update the Story Summary manually.

This is due to the limitations and costs of the summarization AI model. Like all models, we are constrained by context size for the summarization. Many adventures would extend far beyond the available context limit, so regenerating a summary based on edits wouldn't be possible in many cases. Making new summarization calls for every historical change could also become very expensive.

How the Memory Bank Works

Let’s get technical: Embeddings and Vectors

The Memory Bank uses AI technologies that will be new to many of you. To explain how the Memory Bank stores and retrieves relevant memories, you’ll need a basic understanding of embeddings, vectors, and embedding models.

Embeddings represent how similar words and phrases are by mapping them in a multidimensional space. They are created by passing text into specialized embedding language models, which return a vector of numbers as a result.

You can then do simple math to compare how similar different vectors are to each other. For example “water” would be more similar to “liquid” than it would be to “solid”. You can also compare phrases and questions. For example, “Who is Sir Theo?” might match closest with the phrase, “Sir Theo introduces himself to you as the knight who slew the evil dragon of Larion”. These two phrases might also match closely to other passages about Sir Theo or the dragon of Larion. But none of those phrases would likely match very closely to the phrase “Sarah was a forest ranger.”

This lets us build a memory retrieval system. First, we store a bunch of memories in the memory bank with both their text and the embedding vector of that text. Then, when we want to find relevant memories, we decide what text we want to use to look up memories (for our system, we use the most recent action). We get the embedding vector for that lookup query, and we compare it against all the vectors in our memory bank and get a relevance score.

A higher score means it’s more similar to the query, and a lower score means it’s less similar. This lets us decide which memories are relevant to the current events in your story and should be included in the context sent to the AI when generating the story.

Now, we’re ready to talk about AI Dungeon’s new Memory Bank Feature.

Storing and Retrieving AI Dungeon Memories using Embedding Models

With the Memory Bank, each Memory (an AI-generated summary of four past actions) is embedded and stored. As you continue to take more actions, your Memory Bank starts to fill with more Memories.

When you start a new story, your Memory Bank will be empty. As you take actions, and Memories are created, they will begin to fill your available memory slots. Initially, no memories will be used since your entire adventure history will fit in the context. Once your story has reached the point where it can no longer fit your whole adventure into the context window, we start to retrieve memories from the Memory Bank. We do this by using an embedding model to rank all of your memories by relevance to your current story.

Then, we take the most relevant memories and include them in the context being sent to the AI for the next action of the story. For example, if you are just returning to the town of Castlebrook, the AI will find all the memories related to Castlebrook, remember what happened the last time you were here, and include the most relevant in the context. These are called “Used Memories.”

Before including a Memory in the context, we also check to see whether the Memory is included in its full text within the Story Summary. This check prevents the text of a Memory from appearing twice in the context.

The Memory Bank will be allocated a portion of the context window, similar to how we allocate space to Plot Essentials (formerly called Memory), Author’s Note, and Story Cards. Based on the context space available, we’ll include as many of the top memories as will fit from your memory bank.

Once your Memory Bank has been filled and a new Memory has been created, we remove the least used memories to make room for the new addition. These removed memories are called “Forgotten Memories”. Just like with human memory, very old memories might stay in the Memory Bank forever if they have been used frequently.

The larger your Memory Bank, the more room you have to store memories that could be used in your story, increasing the accuracy of the AI storyteller.

As you play, the AI will continually curate the most important and relevant memories from your story and intelligently include them in the context being sent to the AI when they are relevant, improving the overall experience.

Launch Details

Opt-in Experimental Feature

The Memory System could be one of the most foundational changes to AI Dungeon in recent history. Although we’ve tested it thoroughly internally and with our Alpha testers, we anticipate that as you all start playing with it and give us feedback, we may need to adjust parts of the experience. There could also be unintended behaviors we haven’t encountered yet, or edge cases we didn’t consider. The design might need adjustments to be clearer or more helpful to you.

To avoid any unintended player frustration, we’re launching this first into Beta and then into Production as an opt-in Experimental feature. You can enable it under the AI Models tab on the game screen settings sidebar. There’s a new section there called “Memory System.”

How your membership tier affects the memory system

The Memory System will be a Premium feature. When we initially announced it, we expected to make it available to Free players. Our initial prototype was based entirely on the embedding system, and our projected costs were much lower. As we continued to iterate and improve the system, we realized summarization played an important role for both Auto Summarization and the Memory Bank. Those summary AI calls increased the cost of running the Memory System for players and, as a result, it’s currently cost-prohibitive to offer the Memory System to Free players.

Our hope is that we can offer the Memory System to free players in the future. We’ve done this before with other features that were Premium only, then made available to free players over time. Examples include unlimited AI calls, image generation, multiple AI model options, advanced settings, and even Premium AI (with our promo actions).

All subscribers will have access to the Auto Summarization feature. For the Memory Bank, each tier will have a set number of memories that can be stored. You can view the Memory Bank sizes of each tier on our Membership Page.

Key differences from the Heroes memory system

Those familiar with the Heroes memory system will see similarities with this new Memory System for AI Dungeon. The work done on Heroes directly inspired and informed the new memory system. There are some important differences and adaptations we made to implement it for AI Dungeon.

The Heroes experience is a much more traditional RPG experience. We track important stats like health, quests, inventory, levels, and player characteristics, which are stored in a “game state.” We didn't bring this part over since AI Dungeon is a collaborative storytelling experience.

We also made some adjustments to account for the different gameplay in AI Dungeon. We know you all value editing, a concept that isn’t as prevalent in Heroes, and supporting editing with the AI Dungeon Memory System required some adaptations.

We’re excited for you to try Long Term Memory!

You’ve repeatedly told us that the AI forgetting important information is one of the most frustrating parts of playing AI Dungeon. The new Memory System is built to address that issue. As you play with it, please share feedback and let us know how it affects your stories and adventures. This is a new system, and that means there’s opportunity for improvement and iteration.

We know many of you have been excited about this new feature. Thanks for being patient as we went through additional iteration cycles to get the experience right. We hope the Memory System will help you have better, deeper, and longer adventures.

1. Model Agnostic. We’ve chosen to be model agnostic so you have access to the best models on the market and benefit from the billions of dollars currently being invested into better models by multiple companies. You’ve seen the fruit of this strategy lately with the launch of MythoMax, Tie Fighter, Mixtral 7x8B, GPT-4-Turbo, Llama 3 70B, and WizardLM 8x22B. Read more →
2. Vendor Agnostic. We’ve also chosen to be vendor agnostic so you benefit from the competition among current providers. The recent doubling in context length was possible because of this. Read more →
3. Operate Profitably. Given the scale of AI Dungeon, we could bankrupt the company very easily if we weren’t careful. We spend a lot of time thinking about AI cost to ensure AI Dungeon will be around for a long time. Our goal is to give as much as possible to you without putting the future of AI Dungeon at risk.
4. Iterate Quickly. We’ve designed our technology, team, and models around fast learning and iteration. The recent rise of instruction-based models means models can be quickly adapted to the AI Dungeon experience without needing to create (or wait for) a fine-tuned model suited for role-play adventures.
5. Enable endless play. We want to offer models that allow you to play how you want. Outside of a few edge cases (such as sexual content involving minors and guidelines for publicly published content shared with our community), we want you to go on epic adventures, slay dragons, and explore worlds without constraint. Because of our model/vendor agnostic strategy, we have the flexibility to ensure we get to control the approach. Read more about this strategy in our blog post about the Walls Approach →

How We Identify New Models to Evaluate

At first, we evaluated every model that launched. Early providers included OpenAI, Cohere, AI21, and Eleuther. Lately, we haven’t been able to keep up with the rate of new models being launched. Here’s one visualization of just how the AI Model space has accelerated.

We’re selective about which models to evaluate. We base that decision on information we source from the AI community on X, LLM leaderboards, our technology partners, and members of our AI Dungeon community.

When a model piques our interest and seems like it could be worth exploring (when it could have a desirable combo of cost/latency/quality/etc), we do some light exploration around feasibility and desirability. If there’s a playground where we can test the model, we’ll play around a bit ourselves to see what we think. We also talk to our current providers to see how/when they may offer a model at scale.

If everything seems positive, we move into our model evaluation process.

How We Evaluate AI Models

Once we’ve identified a model we are interested in, then the real evaluation starts. Here are the steps we take to verify if a model is worth offering to you in AI Dungeon:

1. Research. As mentioned in the selection process, we look to a number of sources, including industry benchmarks, leaderboards and discussion in the broader AI community, for indicators of which models are the most promising.
2. Playground testing. Someone on our team experiments to confirm we think it could work with AI Dungeon.
3. Finetuning (if required). GPT-J (which powers Griffin) and AI21 Mid (which powers Dragon) are examples of models that clearly needed fine-tuning to perform well for AI Dungeon. Newer models have been able to perform well without finetuning.
4. Integrate the model into AI Dungeon and make sure it works. For example, we recently evaluated a model (Smaug) that seemed compelling on paper but wasn’t able to generate coherent outputs due to its inability to handle the action/response format we use in AI Dungeon.
5. Internal testing. Does the model behave as we expect it to with AI Dungeon’s systems? For instance, when we first implemented ChatGPT, it became clear that we’d need additional safety systems to minimize the impact of the model’s moralizing behavior.
6. Alpha testing. Our community alpha testers help us find issues and give a qualitative sense of how good the model is. The models from Google didn’t make it past our Alpha testers due to moralizing and lower quality writing than competing models.
7. AI Comparison. Players who opt into the “Improve the AI” setting are occasionally presented with two AI outputs and asked to select the best one. These outputs are from two different models, and we compare how often one model’s responses are preferred over another’s. To achieve statistical significance for the test, we collect a few thousand responses per AI Comparison.
8. Experimental access. The final step is giving you all access to the new models in an experimental phase. We often make significant adjustments to how we handle models as a result of the feedback you share. In some cases, models may not be promoted past the experimental phase if players aren’t finding value from them. For instance, we’re considering whether to promote Llama 3 70B since players have reported it repeats frequently.

At any step of the process, we may decide to stop evaluation. Most models don’t make it through our evaluation process to become an offered model on AI Dungeon.

How We Deploy Models

Once we commit to offering a model on AI Dungeon, we then figure out the best way to run it at scale. With private models we often can only run them with the creator of the model (like AI21’s models). For open-source models, we can choose between running on rented hardware or using other providers that run LLMs as a service (which is our preference). By optimizing our model deployment costs we’re able to deliver better AI to users for the same price.

We also have an alert system and series of dashboards that show us the number of requests, average context in and out, latency profiling (average request time, max request time), and estimated cost. This lets us keep our AI models running smoothly and quickly respond to any issues that come up.

How We Retire Models

Given the complexity of models, it’s sometimes necessary to retire models that are no longer adding much value to the community. While it would be nice to offer every AI model perpetually, maintaining models takes time and development resources away from other improvements on AI Dungeon, including new AI models and systems.

Because of that, we need to balance the value a model provides against the cost of maintaining it (especially in developer time). We’re guessing most of you are no longer pining for the good old days of GPT-2 😉.

Before deciding to retire a model, we consider usage, tech advances (i.e., instruction-based models), latency, uptime, stability, error rates, costs, player feedback, and the general state of models in AI Dungeon (i.e., how many do we have for each tier).

Each model is unique, like an ice cream flavor. Taking away your favorite flavor can be frustrating, especially if that model does things that other models don’t (like mid-sentence completion). We hope there’s solace in the fact that when models are retired, the recovered development resources are reinvested into better models and new features that make AI Dungeon a better experience for you.

Today here’s the % breakdown of model usage for various models:

Free Players

MythoMax 73%

TieFighter 17.8%

Mixtral 8.8%

Griffin 0.4%

Subscribers

Mixtral 79%

MythoMax 8%

WizardLM 8x22B 5%

TieFighter 4%

Llama 3 70B 2%

Dragon 1%

GPT-4-Turbo 0.5%

ChatGPT 0.4%

Griffin 0.01%

You’ll notice a few things. MythoMax is our most popular model, even capturing some use from paid players who have access to all the models. Mixtral is the clear favorite for premium players.

Because of the advances in tech as well as low usage, we will be retiring Griffin, Dragon, and ChatGPT models on May 31st, 2024. Griffin, while it’s served us well, has exceptionally low usage, the worst uptime of all our models, and a high rate of errors. It requires more developer maintenance than all other models we offer. Dragon and ChatGPT also have lower usage now. Retiring models enables us to focus on other product work including additional model improvements, bug fixing, and building new features.

GPT-4-Turbo is somewhat of an outlier. Despite its moralizing, it’s one of the best story writing models available. Players who use it love it! While its usage rate is low relative to other models, it’s actually well represented for a model only available to Legend and Mythic tiers, though it’s clear players still favor Llama 3 70B and WizardLM 8x22B. We are evaluating the recently announced GPT-4o as a potential replacement for GPT-4 Turbo which could mean offering higher context lengths. Although venture-funded OpenAI says they’ll offer limited use of GPT-4o for free through their own ChatGPT client, it will not be a free model for API users (like AI Dungeon), so it will still be a premium model for us. First, though, it needs to pass the evaluations we’ve outlined above.

Moving Forward

This was a deeper peek into our approach to models than we’ve ever given. We hope it’s clear that we spend a lot of time thinking about which models we can offer to you and how to provide them best.

Thank you to all who have given feedback on our AI models. We will continue to communicate as much as we can about models and planned model improvements. It’s exciting to realize AI will only get better from here. The past few months have shown us just how fast things can change. And we’re excited to explore with you how much better role play can be as AI keeps improving.

Their extremely generous refund policy also recently inspired us to change our approach to subscription refunds. Now if you forget to unsubscribe from AI Dungeon we’ll refund you all the way back to the last time you played, no matter how many months it’s been. These are just a few examples where you can clearly see Costco’s obsessive focus on delivering more value.

Sustainable Business

To deliver value we also need to make sure we are running a sustainable business. Giving everyone unlimited access to GPT-4 Turbo, the most expensive AI model we offer, would create a ton of player value…but it would put us out of business. We need to operate sustainably to provide and build the best AI Dungeon experience possible.

This is particularly challenging for companies who rely on AI like we do. Although costs are coming down, AI is still relatively expensive compared to tech used by traditional games and platforms.

Thanks to all of you, Latitude is fortunate to be operating sustainably today. We’re not on the VC-funded hamster wheel that many companies get caught in, where they need to fundraise regularly to stay afloat. We’re in control of our future and destiny. Without investors pressuring us for higher profits, we’re able to stay true to our values and stay focused on serving you.

We might be hamsters, but we’re not part of the startup rat race.

Commitment to Free Players

Our free tier is where the tension between player value and sustainability has often been the most challenging. It’s often easy to focus on your paying customers, but we work hard to make our free tier as compelling for players as we can.

Because of how expensive AI is, our free tier continues to be unique in the AI space. This hasn’t always been easy. In the past we’ve had to limit free player use through energy, and then later, ads. Today however free players can enjoy unlimited AI Dungeon without restrictions.

We’ve also significantly expanded the features available on the free tier. Just in the last year, we’ve been able to go from one free AI model to three. We’ve added promotional actions you can get a taste of premium AI models without paying. And Advanced settings are available now for everyone, so you can get the best experience possible.

So why offer a free tier at all? Because it actually adds value to ALL players, both free and subscribed.

The free tier means more players use AI Dungeon, and in turn, that leads to more content being created, more people for you to interact with, a better social experience, and overall more fun. Even though it makes being sustainable more challenging, a free tier results in much more value for the entire community. Plus it allows anyone to experience the magic of AI Dungeon even if they aren’t at a place where they can afford a subscription.

Our goal is to give players as much value as possible, while still being able to operate sustainably. As the AI space continues to evolve and improve, we expect to find new ways to give you more value.

Our unique positioning

Now, let’s get into the details of strategies we used to give you more context in AI Dungeon. This is the first time we’re sharing some of these publicly since, on the surface, these strategies might not seem like they’d have a direct impact on player value. Please tell us if you like hearing about this kind of work, and we’ll share updates like this more often if you do!

Provider Agnostic Architecture and Negotiating Power

One of the most painful lessons in our company’s history was learning how dangerous it is to be dependent on a single technology provider. In the past this had a profoundly negative impact on players’ experience, and we’ve changed how AI Dungeon is architected since then to avoid being caught in a similar situation again.

AI Dungeon is set up to be provider agnostic. This means that, at any time and with minimal effort, we can change AI providers. For example, we’ve hosted Mixtral on three different providers since launching the model to players in December. There have been instances where we’ve had outages on one provider, and been able to switch to another provider to keep AI Dungeon running.

Being provider agnostic allows us to evaluate dozens of different AI providers and score them on dimensions like cost, strength of partnership, privacy and security policies, content policies, tooling, and server uptime and stability. We’re careful in evaluating potential partners to make sure we choose partners who can provide the best overall experience for our players. We’ve also negotiated changes in policies to align with your values around content freedom and privacy.

Since we have clear insights into our traffic and AI use, we’ve been able to negotiate discounted rates on AI compute by committing to large amounts of traffic with our technology partners to receive volume-based pricing. Like Costco, we’re passing these savings on to you in the form of increased context length for no additional charge.

We’re particularly grateful to our two newest partners, together.ai and octo.ai, who have made these recent changes possible. We’re also grateful to our other providers—Azure, AI21, and Coreweave—who continue to be good partners.

Model Agnostic Strategy and Robust Evaluations

Not only are we agnostic to providers, we’re also model agnostic. Our AI systems are database driven, allowing us to quickly add new AI models, control access, and run comparison tests. The architecture is flexible enough that when a new AI model becomes available, we’re able to evaluate it without even having to write any new code.

We’ve deliberately decided NOT to be in the business of creating our own custom AI models (although we are finetuning models for specific tasks). By leveraging models available commercially and through open source, we’ve been able to take advantage of the wave of innovation happening right now in the AI space. We’re building AI Dungeon (and Heroes) to be incredible experiences that can leverage the best AI models available on the market. We’re convinced that with a small team like ours, building models in-house won’t let us provide value to you as quickly as we can by leveraging third party models. We love being built on the backs of giants like Meta AI, Kobold.ai, Wizard LM, Azure (and OpenAI), Mistral, AI21, and more.

And, oh boy, have we been busy evaluating new models. We regularly get questions from some of you about whether we’re testing new models. The answer is almost always “yes”. We’ve evaluated nearly every promising new model that’s been introduced recently. As a result, our process for evaluating AI models has become quite robust, blending both qualitative and quantitative feedback into our process.

We’ve always relied on our AI Comparisons tool for qualitative evaluation, and now robust industry benchmarks and leaderboards are providing additional metrics to look at. We also look at model properties like parameter counts, response times, and supported context lengths.

We also do qualitative tests by playtesting ourselves, and for models with more potential, opening them up to Alpha testers for more testing and feedback. We look at things like storytelling ability, following instructions, creativity, and coherence. Moralizing is an issue we look out for, and is more common from commercial providers like Azure and Google. Some of these models reject harmless content like fantasy violence, making them poor models for a role play game like AI Dungeon.

Of course, we also look at costs to see whether models are viable and compete with other models at their price point. For instance, we’ve had some models perform similarly to Mixtral 8x7b, but at a much higher cost. If a model performs the same as Mixtral, but we can only offer 1/4 as much context length, this doesn’t seem valuable to offer players. Mixtral, for instance, is a better model that costs less than our outgoing Dragon model. It’s affordable enough that we can offer larger context sizes at each premium tier than ever before. Being model agnostic allows us to quickly test, evaluate, and introduce new models like Mixtral that give more value to players.

Our goal is to have a small portfolio of some of the best models available. We only add new models to our lineup if it’s clear they’ll offer significant value to you. Most models don’t pass our evaluations. For instance, we’ve been surprised that none of the Google models have met our expectations due to their heavy moralizing and below-average storytelling abilities. We’re hopeful that they’ll introduce future models that will be on par with offerings from others.

Being model agnostic means we’ve been able to have better models at better pricing. This means better AI and larger context lengths for you. And the good news is, it’s only going to continue to get better over time!

The outcome of us being provider and model agnostic, is that we’re able to easily take advantage of newer models that offer better performance at a lower cost. We’re also able to negotiate incredible pricing terms to further lower our costs, enabling higher context lengths and a better experience for all of you.

Putting it all together

We’ve gone deep on the heartbeat of Latitude, a constant focus on how we can deliver more user value. We’ve also shared how we set up our AI architecture to enable that. Now let’s summarize how that has all come together to make double context possible.

In short, to support giving everyone double context, we needed to find a way to sustainably support (roughly) 2x AI costs. We did that in a few ways:

• We architected AI Dungeon to be provider agnostic, allowing us to find the best providers at the cheapest prices, especially as new models and providers have come out in the past few months.
• We leveraged the high traffic volume of AI Dungeon to negotiate discounted pricing on AI compute (so we could give that back to all of you)
• By being model agnostic, we’re able to evaluate and deploy the highest quality and most affordable models
• We’re passing these cost savings on to you by providing double context length for each tier, for no additional charge.
• These changes will still allow us to operate sustainably, ensuring AI Dungeon will be around for the foreseeable future

As we mentioned earlier, this is the first time we’ve shared some of these details publicly. Please let us know if you enjoyed this post, and we can share more updates like this in the future. We’re incredibly grateful for your feedback and support, and we are working hard to give you more and more value in AI Dungeon.

The third installment of our AI Renaissance is now available in the Beta environment! This Drop includes a new AI text model (Llama 3 70B), doubled context and response lengths, Advanced Settings for everyone, and custom AI Instructions. Still to come: a brand new, automated memory system!

Rollout Plan

Doubled context length, doubled response length, and Advanced Settings are available in both the Beta and Production environments for everyone today.

Llama 3 70B and AI Instructions are available in the Beta environment today and will be moved into Production in the next few weeks.

The new memory system will be available in Beta several days after we’ve wrapped up some loose ends.

You can read more about each Drop on our AI Renaissance update page: https://aidungeon.com/renaissance.

You can see details and compare plans on our new pricing page, which you can find at play.aidungeon.com/members. As a reminder, these changes will be rolling out gradually, and we’ll update you when the work is completed.

If you have any questions or feedback about these changes, please feel free to reach out to us on Discord, Reddit, or through our support email. Thanks for being the best part of AI Dungeon!

Stay tuned, we have more exciting information coming soon about the next AI Renaissance release!

Players have been praising Mixtral for its coherence and content freedom over other models. This made us wonder whether we should change our promo offer, which gives free players the opportunity to enjoy our premium models without starting a subscription. When asked on Discord and Reddit, our community overwhelmingly said YES, make Mixtral the promo model. Because of their feedback, we decided to move forward with this change so everyone can try Mixtral. We think you’ll appreciate its quality improvements and the added flexibility it offers in subject matter.

Starting today in the Beta environment, you can use your 20 daily premium actions for Mixtral. This will be available in the Production environment early next week. After that, ChatGPT will no longer be available to use with promo actions, but it will remain an option for subscribers. As a free player, you always have complete access to our base models, Griffin and MythoMax.

Thank you for all of the feedback you’ve given about Mixtral and MythoMax during their Early Access period. We hope they continue to improve your gameplay, and we look forward to hearing more feedback!

Happy Holidays, fellow adventurers! ‘Tis the season of giving, and we wanted you to have some fun AI presents to play with over the holidays! We’re adding two new AI models for early access, Mixtral and MythoMax, into our lineup for AI Dungeon. Mixtral will be available for AI Dungeon Premium subscribers, and MythoMax will be available for both Premium subscribers and free players.

About Mixtral and MythoMax

Mixtral is the newest model available from Mistral AI, and we believe it has the potential to become the model of choice for most Premium players. Mixtral is a sparse mixture-of-experts network. It's an 8x7B model, coming in at 46.7B total parameters. Industry benchmarks rank Mixtral at or above models like Llama 70B or ChatGPT (GPT-3.5). Many of you have enjoyed using ChatGPT in AI Dungeon, and we think Mixtral will offer a similar quality of play experience without ChatGPT’s moralizing tendencies around common adventure themes of violence or romance.

MythoMax is going to be the first new model we’ve added to our free tier in over two years! As many players know, using Large Language Models is expensive, and we feel fortunate we’ve been able to provide players with an unlimited free play option. And now the free offering will be even better. MythoMax is a Llama 2 13B (twice as large as the current Griffin model) and was specifically optimized for storytelling. We’re excited for you to try it and see how it compares!

As a callout, we want to thank the amazing open-source developers and creators responsible for sharing these models with the broader AI community. (If any of you want free AID subscriptions, let us know!)

Launch Details

Both Mixtral and MythoMax will be launched into Production as “Early Access” models. We hope this label will communicate to players that we expect things to improve and change over the next month or two. We’re looking forward to your feedback, which will help us determine the optimizations needed to bring out the best of both models.

MythoMax will be available for both Free and Premium players. Context length will be determined by your tier.

• Free + Travelers (Steam users)—1k context length
• Adventurer—2k context length
• Hero—2k context length
• Legend—2k context length

Mixtral will be available for all Premium subscribers. Context Length will also be determined by tier.

• Adventurer—2k context length
• Hero—4k context length
• Legend—4k context length, with 8k context length planned

Some of you may be aware that these models are technically capable of even larger context lengths, and our goal is to increase the context lengths offered for each model. However, larger context lengths can sometimes lead to a degraded AI experience, so we plan to do more testing and evaluation before expanding the context lengths.

Our Next Chapter: AI Renaissance

The launch of Mixtral and MythoMax signals the start of the next Chapter for AI Dungeon.

With Unchained, we listened to player feedback and addressed limitations players were frustrated with. This included removing ads, improving stability, making image generation available to everyone, giving unlimited access to Premium models, and making several important AI improvements.

We’re currently wrapping up work on Phoenix, which represents a rebirth of AI Dungeon. Most noticeable was a complete overhaul of our UI/UX, making it more intuitive and engaging yet familiar for experienced players. Phoenix has significantly sped up how fast we can deliver new features to users and has also brought improved stability, device compatibility, and performance gains. We’ve also introduced new capabilities like Story Cards, theming, custom images, and more. It was a dramatic change, so Phoenix had a long Beta period where we gathered feedback from players and implemented improvements before making it the primary version of AI Dungeon.

Now, with the strong foundation built from Phoenix, we’re ready for our next chapter of AI Dungeon focused entirely on the AI experience. We’re calling it the “AI Renaissance”.

The Renaissance will feature new models (in addition to the ones announced today), improvements to model settings, increased stability and performance, more personalized model experiences, and more creative control over the AI. We’ll share more details about our plans after the New Year.

For those following the development of AI Dungeon Heroes, the work we’re doing for the AI Renaissance will lay the foundation for the launch of Heroes. We intend to use a variety of models (both current and future) to bring the Heroes experience to you. As Nick has mentioned in his Dev Logs, the planned launch features for Heroes are stabilizing. The next step is to take the existing AI game engine, which currently relies on larger models like GPT-4, and refine it down to smaller and cheaper models while preserving or improving accuracy.

Happy Holidays!

Finally, we want to wish all of you a very happy holiday season. We hope all of you will be able to spend time with loved ones, take time away from the busyness of life, and make some meaningful memories.

Our team will be taking some time off for the holidays as well. Phoenix was a massive endeavor, and our team has been working long and hard all year. Most of the team will be taking well-deserved breaks.

Our community and support team will still be checking in to address urgent support issues, but we may not be as responsive as usual for all types of requests.

We’re looking forward to coming back for 2024 and making it the best year yet for AI Dungeon!

Earlier this month, we sent out a player survey so you could tell us how you feel about Phoenix, the newest version of AI Dungeon. We had over a thousand responses in the two days the survey was up. After reviewing the results, you’ve made it clear that Phoenix is an improvement over the Legacy version of AI Dungeon. The majority of players prefer Phoenix over Legacy, and most of the feedback we received for improving Phoenix is being addressed in updates that have already been released or are being released this month. We’re still listening to feedback, and we still have improvements we want to make, but you’ve given us the confidence to move forward with retiring Legacy.

With the success of the Phoenix update, we’re now ready to announce plans for retiring the Legacy version of AI Dungeon. Support for the Legacy version will officially end on Dec 1, 2023. On December 31, 2023, we’ll begin the process of shutting down Legacy.

Phoenix is the future of AI Dungeon and has already allowed us to develop new features we couldn’t have done on Legacy. With a small team, we’re unable to maintain two versions of AI Dungeon, and Phoenix is now superior to Legacy for features, stability, and performance.

Although there were a few bumps along the way, players generally consider the update a beneficial change. 65% of players prefer Phoenix over Legacy. With a change this big, it’s understandable that not everyone will be happy with the changes. However, the reception of Phoenix has been as good as we could have hoped for. For any of you who are still dissatisfied, please continue to share your feedback—we’re always looking for ways to improve the product.

As we said from the beginning, listening to and addressing player feedback for Phoenix was one of the most important requirements before retiring Legacy. We had a lengthy Beta period for Phoenix where we were able to collect player feedback through user testing, analytics, feedback on social platforms (like Discord and Reddit), and our support channels.

Some examples of things we added to Phoenix as a result of player feedback include adding a list view for content cards and Story Cards, replacing “Undo to Here” with “Erase to Here,” adding ChatGPT as the model for Story Card generation, adding the ability to upload custom images, and introducing themes and display options to the play screen. We also moved up the timeline for migrating Legacy worlds into Phoenix scenarios. We also captured and fixed hundreds of bugs and performance issues reported by players.

In the survey, many of you have requested features that will take more time to build. Some of the features we have either started working on or will work on soon include:

• Adding a button for “Erase to here” (this is currently available as a command-based action)
• Improvements to content cards (seeing more preview text, NSFW tags, play stats, etc.)
• Better options for creating characters and using them in stories
• Improved memory systems
• RPG features like quests, inventory, or achievements

With Legacy retiring, our focus will shift to new AI models and improvements, AI Dungeon features and improvements, developing Heroes, and continuing to improve the performance and stability of AI Dungeon.

As always, thank you for your continued feedback and support. We are excited for what’s coming and look forward to sharing more details with you soon!

• All scoped features, including ones identified through user feedback, are implemented. These include: Game screen actions; World Info (now Story Cards); Reading Adventures view (others and your own); Scripting; Multiplayer; Performance improvements; and Accessibility features, including display settings.
• All features are working on the iOS and Android versions of AI Dungeon. This means we’ve resolved all known Priority 1 (P1) and Priority 2 (P2) bugs discovered by beta testers.
• All features support accessibility tools and standards.
• All player feedback has been recorded and triaged. We have responded to as much feedback as possible and shared updates about how we’re applying that feedback.

While the apps have been in beta, we’ve built several new features and brought over a few Legacy features that players said they missed in Phoenix:

• Undo to Here (now /erase-to-here, or /eth)
• Scenario Character Creator
• Story Card Import & Export
• Clickable Tags
• Linked Accounts
• Release Channel

You may have noticed we’re transitioning our mobile apps to Phoenix a few weeks later than we estimated. We needed a few extra weeks to make sure our success criteria were met, and the extra time allowed us to fix more bugs, add more features, and keep working on stability and performance. It also took some time for us to get enough feedback from players to feel confident moving forward.

Now, we are ready to announce a timeline for when we expect this transition to happen officially. We’ll begin the rollout of Phoenix to the mobile apps on Tuesday, October 24, 2023. We’re opting for a phased rollout for both the iOS and Android apps. This will allow us to monitor for any issues and quickly address bugs. The phased rollout schedule means it may take up to a week before all players receive the update on their devices.

Once Phoenix becomes the default mobile app experience, we will move to the next phase, which is preparation to retire Legacy. We expect this process to take around 3 weeks. During that time, we’ll be fixing bugs, gathering feedback, and implementing the following new features:

• List View/Cards View
• Improved Image Experience
• Scenario Filters
• Text Display Options

Throughout this process, we have been encouraged by messages and questions from players.

It would be great if the [app] could match the features and UI improvements of the web version.

Are the Story Cards coming to the mobile app as well?

This Phoenix update is well-thought and fixes many (if not all) issues of world creation. If you've quit using AI Dungeon earlier, this overhaul alone is worth checking the app out again.

You guys have exceeded all my expectations here, very well done both conceptually and visually. This kind of flexibility of story setup and proper contextual consistency based on initial inputs/choices was really missing from the legacy version. This is the way!

Thank you for all of your hard work and communication. I love your game and always have fun with it!

A big thank you to our players who have been testing and giving feedback all along the way. We rely heavily on input from our community, alpha testers, and beta testers to help us continuously improve AI Dungeon, and we are very grateful for your support!

All players are welcome to participate in Beta testing by visiting beta.aidungeon.com. Now that Phoenix is released to our native mobile apps for iOS and Android, the "Release Channel" setting allows players to choose to use Beta or Production versions of the native apps.

Production

Our Production environment is the most stable version of AI Dungeon. We only push new features to Production after they’ve been thoroughly reviewed in each of the previous environments. By default, all players are directed to the Production environment.

Legacy

For a limited time, the Legacy version of AI Dungeon will be available at legacy.aidungeon.com. We’ve shared on an earlier blog post why we’re retiring Legacy. Please feel free to reach out if you have any other questions about the transition.

Becoming an AI Dungeon Tester

These new AI Dungeon environments mean more players can participate in testing AI Dungeon. Some of the benefits of being a tester may be interesting to you. For instance, our testers get to preview unreleased features. Our testers also get to have frequent communication with the Latitude team, and they are able to shape our roadmaps and provide input on what we work on next. Our Alpha testers, specifically, also get a few other benefits, like unlimited access to Heroes.

Testing AI Dungeon

To get started, simply start using the Beta environment. This can be done by visiting beta.aidungeon.com, or by selecting the Beta environment from the "Release Channel" dropdown of your Account Settings in the mobile app.

Then, use the bugs-feature-requests channel in Discord to provide feedback and log bugs. Be sure to follow the directions for logging bugs, or submitting feature requests.

Becoming an Alpha Tester

In the past, we’ve had an application process to join our Pioneers and become an Alpha tester. Going forward, we’ll select new Alpha testers from players who are active in Beta testing and Heroes discussions. We’re looking for players who are passionate about AI Dungeon and like trying new features, who are able to be patient with a few bugs, can articulate the problems they are encountering and document the steps to reproduce issues, provide thoughtful feedback, and are respectful to others in the community.

Players we invite to the Pioneers group will now become a Trial Pioneer for about 3 months. At the end of the 3 months, we will either upgrade them to full Pioneer or rotate them out of the group. We are also merging the Heroes Early Preview opportunity into the Pioneers group, so this is another way players can become alpha testers and Pioneers.

We love getting player feedback; we look for it and review it every day. Positive feedback helps us know we’re on the right track. Critical feedback motivates us to grow and shows us where we can improve. Please continue to share your feedback with us on our Discord server. We appreciate all you do for AI Dungeon and the community!

Updated: October 24, 2023

We expected to transition all web (i.e., everything but native app) traffic to the Phoenix version of AI Dungeon today (September 14, 2023). As we outlined in our recent blog post, we set specific success criteria to be met before moving forward with the transition. Although we’re close, we didn’t feel like we met those success criteria, so we’re postponing the transition. We have a new Release Candidate available in Beta for testing. We will start the transition on Sept 15th or Sept 18th, depending on how Beta testing goes.

In particular, we committed to resolving all P1 and P2 bugs. This morning, as our team gathered to prepare for the transition, we found additional bugs that needed to be resolved. Although we’ve fixed the bugs and will be pushing those updates to Phoenix today, we feel it will be best to let the community test this new Release Candidate for some time before we cut over web traffic to Phoenix.

Please continue to share any feedback and bug reports in our dedicated channel on Discord.

Phoenix is officially feature complete! On Friday last week (September 8, 2023), we pushed several new features to AI Dungeon, including Story Cards (replacement to World Info), Scenario Character Creator, Adventure Read Screen, and Scripting.

This means that, so far, we’re on schedule with the Phoenix transition we shared on August 30th. The next step is to transition the Web experience to default to Phoenix for all users. We are planning this transition for Thursday, September 14th. Legacy will still be available at legacy.aidungeon.com. Going forward, play.aidungeon.com will be the new URL for AI Dungeon. The play.aidungeon.io and beta.aidungeon.com URLs will both be forwarded there.

We’re paying close attention to player feedback to make sure any bugs and loose ends are addressed before this transition. As a reminder, we will not move forward until we’ve been able to fix any Priority 1 (P1) and Priority 2 (P2) bugs that have been reported for Phoenix on the Web. We’ll also be capturing any player feedback and addressing any critical feedback before moving forward. Please be sure to share any feedback or bugs in our bugs Discord channel or email us at [email protected].

After we successfully transition, the next step will be to transition the native iOS and Android apps to Phoenix. We expect this to happen one to two weeks after the Web transition.

Please be sure to share any feedback with us as we prepare for this transition. Our goal is to make sure this is a smooth transition for our players.

We’re more excited about the future of AI Dungeon than we’ve ever been. We recently shared we’ve started developing a new game mode, Heroes, that will introduce the next evolution of AI Dungeon, with inventory, character progression, persistent state, and much more. Heroes is now possible due to the work we’ve been doing on Phoenix, which is a rebuild and redesign of AI Dungeon. With Phoenix, not only are we improving the experience for players, but we’re also making important architecture upgrades that will bring greater stability and make it easier for our team to add new features to AI Dungeon.

With that in mind, let’s go over some changes we’re making that will enable our team to focus on the future of AI Dungeon. First, we’ll talk about the Phoenix transition plan, which includes plans for retiring the Legacy version of AI Dungeon. Next, we’ll talk about the future of Voyage, followed by our plans for AI Dungeon Steam support, and finally, a list of FAQs about these changes.

Phoenix Transition Plan

We’ve been working for months on Phoenix, a complete redesign and rebuild of AI Dungeon. This was a massive project and one that, frankly, took longer than we expected. We’re excited for Phoenix to be completed since it will bring greater performance, stability, and device support and lays a foundation for us to add new features and experiences to AI Dungeon.

We still have more work to do, but we’d like to share our plans for transitioning Phoenix to be the primary AI Dungeon experience for all players.

We’ve outlined some key milestones and success criteria that we’ll adhere to as we move forward. Each milestone has defined success criteria that must be completed before moving on to the next stage. Because of this, the dates being shared are tentative and will be updated as we move through the process. We intend to leave room for player feedback and will adjust the scope and timeline to account for necessary changes to AI Dungeon prior to transitioning.

Feature Complete + Begin Final Open Beta

Estimated Date: September 7th

We expect to have all planned Legacy features migrated to Phoenix by September 7th. Once all features are completed, we’ll start the final phase of our open Beta testing. During this process, we’ll be fixing bugs and gathering player feedback. Please feel free to share feedback with us on Discord or through a survey that will be available in AI Dungeon.

Success Criteria

Implement all planned Legacy features into Phoenix, which includes:

• Game screen actions
• World Info
• Reading Adventures view
• Scripting
• Multiplayer
• Performance improvements
• Accessibility features, including display settings

Make Phoenix the Default Web Experience

Estimated Date: 1 week after Feature Complete

If the success criteria are met, we’ll set Phoenix to be the default experience (at play.aidungeon.com) for players using the web version of AI Dungeon.

Once Phoenix is the default experience, players who want to use the Legacy version of AI Dungeon can do so at legacy.aidungeon.com. The Legacy version will be available through the entire transition process for both mobile and desktop browsers.

Success Criteria

• All scoped features (from the feature complete list, as well as features identified through user feedback) are implemented.
• All features are working on the web version of AI Dungeon. This means we’ve resolved all known Priority 1 (P1) and Priority 2 (P2) bugs.
• All features support accessibility tools and standards.
• All player feedback has been recorded and triaged. We will respond to as much feedback as possible by sharing whether we’re actively working on it, whether it will go on our roadmap (for after the Phoenix transition), or whether we don’t plan on implementing the feedback.

Make Phoenix the Default Mobile App Experience

Estimated Date: 1 week after Web Transition

Once we make Phoenix the default for the web, our focus will shift to the native apps. Before doing so, we will follow similar success criteria to those used for the web transition.

This stage is the riskiest transition point in the process because the majority of our players use our native mobile apps, and there isn’t a simple way to roll back changes. This is why we’re waiting until after the successful transition of web before we transition the mobile apps. We’ll rely heavily on feedback from our community, alpha testers, and beta testers to help us know when it’s time to transition the mobile apps. If you haven’t been part of the open Beta yet, we invite you to participate and help us identify any lingering issues.

Success Criteria

• All scoped features (from the feature complete list, as well as features identified through user feedback) are implemented.
• All features are working on the iOS and Android versions of AI Dungeon. This means we’ve resolved all known Priority 1 (P1) and Priority 2 (P2) bugs discovered by beta testers.
• All features support accessibility tools and standards.
• All player feedback has been recorded and triaged. We will respond to as much feedback as possible by sharing whether we’re actively working on it, whether it will go on our roadmap (for after the Phoenix transition), or whether we don’t plan on implementing the feedback.

Retiring AI Dungeon Legacy

Estimated Date: 3-5 weeks after Mobile transition

After transitioning our mobile apps, we’ll then set our sights on retiring the Legacy version of AI Dungeon.

Phoenix has been in beta for quite some time, and we’ve received lots of great feedback. Once we make Phoenix the default, we expect a new wave of feedback once the majority of our player base interacts with the new UI. We expect it’ll take us a few weeks to address any new feedback that we didn’t receive from players earlier in the beta process.

We know that not all of our players are on Discord, so we’ll also be inviting players to take a survey about Phoenix’s readiness before we retire Legacy.

Once we retire Legacy, we’ll also be delisting our Steam app at the same time. Steam players will retain their purchased benefits. You can read more about this decision and why we believe it will be good for our players and community later in this blog post.

Success Criteria

• Has all player feedback been addressed (either fixing, not fixing, or added to the product roadmap)?
• Are all the experiences performing as well or better than Legacy?
• Are most players satisfied with Phoenix as their default version of AI Dungeon?

Throughout the entire transition phase, we will be paying close attention to player feedback. Thank you for the feedback you’ve given us so far in the process, and we look forward to hearing more as we get further along in the process.

Voyage

In early 2022, we launched Voyage with the vision of enabling creators to build AI-driven games and experiences. As part of that vision, our team created several other games using the Voyage platform, which helped us build out our tools and infrastructure for the platform. At one point, we envisioned that even AI Dungeon would be a part of Voyage, and we’d transitioned some of AI Dungeon’s systems to Voyage.

However, many of you in the community gave us clear feedback that although the vision of Voyage is exciting, you wanted us to focus our efforts on AI Dungeon first. We agreed, and late in the summer of 2022, we transitioned the majority of our development efforts back to AI Dungeon. As a result, Voyage development has now come to a halt, and players have asked whether we’ll continue to support Voyage.

Today, we’re announcing an indefinite pause for Voyage access & support. We’ll be turning off access to all Voyage experiences, as well as creator tools. We’ll turn off access on September 30th. We still believe in the vision of Voyage, and we may come back to it down the road.

Along with this change, we’ve decided to release Loom as an open-source project. The repo is already available on GitHub. We’re hopeful that Loom can find new life being managed by the community.

We’ll also be adjusting our marketing pages, pricing pages, and guidebook to reflect the removal of Voyage. In the past, Voyage access was one of the differentiators between our Adventure and Hero tiers. Hero and Legend tiers have already had the benefit of additional context size (or memory), and we’ll be making that benefit more clear in our documentation and pricing pages.

Steam

As previously mentioned, another change to help us focus our AI Dungeon efforts is we’ll be delisting our Steam app. This means the app will still be available for those who purchased it, but it won’t receive regular updates, nor will it be discoverable by new players.

Players who purchased AI Dungeon on Steam (when it was a paid download) or who paid for the Traveler tier will keep ALL their benefits. The Steam purchases upgrade the entire AI Dungeon account, and those benefits like larger context, faster AI speeds, and advanced settings can be enjoyed on web, iOS, or Android.

For game developers like us, supporting a game on a platform like Steam needs to accomplish a few things. First, it needs to provide a new game platform for players to use. Second, it should help players find and discover our game through a new marketplace. And third, the platform should facilitate financial transactions between players and game developers. Steam is one of the most popular platforms for games, so it seemed like a platform we should support.

However, it hasn’t delivered value for our players or to us as game developers. First, very few players actually use the Steam client. The majority of players who purchased the Traveler’s upgrade through Steam actually use those benefits on other platforms since the upgrade applies to the player’s account no matter what device they use.

Second, Steam hasn’t been very successful in helping new players discover AI Dungeon. Compared to our other platforms, iOS, Android, and web, we see very few new players coming through Steam. Negative reviews have contributed to poor discovery. Some of those negative reviews stemmed from misaligned expectations, since Steam users are coming from the context of traditional gaming and aren’t as accustomed to the experience, pricing, and gameplay of an AI-based product like AI Dungeon. Other negative reviews were related to mistakes we made as we launched AI Dungeon on Steam, which we deserve and accept. Today, few players are discovering AI Dungeon on Steam, and we don’t expect that to change.

Third, Steam doesn’t support subscriptions very well, and the subscription model is what works best for AI Dungeon. We’ve had to hide our subscription offerings to Steam users, leading them to be disappointed in the free experience. We’ve offered Travelers as a one-time purchase, which provides some great features like faster AI speeds, advanced settings, and larger context size. However, most players who want to upgrade are interested in more powerful AI models, which are only available through our subscriptions (which, again, we cannot support on Steam due to their terms of service).

Because of these shortcomings, we no longer feel it is worthwhile to support AI Dungeon on Steam. Anyone who has purchased AI Dungeon on Steam, or paid for the Travelers tier, will keep their benefits and can use them on web, iOS, and Android.

Feature Requests and Bug Reporting

We will also be centralizing our feature request and bug-reporting process to Discord. In the past, we’ve asked players to submit bugs or feature requests on feedback.aidungeon.io. However, we’ve found that, generally, players prefer to share bugs in Discord. Because our community team (and even our developers) monitor Discord, we often hear about bugs and features in Discord before seeing them on feedback.aidungeon.io. We’ve also been using Discord for capturing bugs and feature requests from our Pioneers group (Alpha testers), and it’s been working very well.

We’ve organically started using Discord as the central place for bugs and feature requests, and now we’d like to make that official.

We’ll be redirecting feedback.aidungeon.io to the bugs_feature_requests channel in Discord. When a feature request or bug is added, our team will use tags to communicate the status of the item. Players and our team will be able to discuss the issue live in Discord.

For any players who don’t want to use Discord, please feel free to share bugs or feature requests with our support team at [email protected].

Why we’re excited about these changes

It’s natural to feel a sense of disappointment or frustration with changes like these. Saying goodbye to Voyage, Legacy, and Steam isn’t easy and we understand not everyone is going to be excited by these changes right now. Our team has spent a great deal of time and effort creating some of the systems we’re retiring today—we’re also going through some feels.

We hope, however, that any disappointment you feel will fade as you see the exciting things that are coming. If you haven’t seen the livestream where Nick shared details about Heroes, we encourage you to check it out. Some of the most requested features for AI Dungeon will be present in Heroes. In early testing, we’re having some of the most fun we’ve ever had on AI Dungeon.

By reducing the number of systems we’re maintaining, we expect better uptime and performance of AI Dungeon. We are already feeling the benefits of new infrastructure that allow us to implement improvements quickly. We still have some work to do, but the mobile experience on AI Dungeon is going to be better than ever.

We’re so appreciative of all of you in our community who have been with us on this journey. We wouldn’t be here without all your input, suggestions, testing, and feedback. We are so fortunate to have such a large, passionate player base who not only enjoys AI Dungeon, but is willing to join us through the development process. Some great things are coming. This is going to be one of the most exciting times to be a user of AI Dungeon.

FAQ

Can I use Legacy on mobile after the app has been transitioned?

Once the native apps have been upgraded to Phoenix, you’ll need to use the mobile web version to access Legacy until it’s retired.

Why aren’t you waiting until Phoenix is more stable?

We will wait, but we wanted to start talking about the plan now, even before we’re ready. Today, the day we are publishing this plan, we agree with players that there’s more work needed on Phoenix. But we’re getting close, and as we finish building the remaining features, we’ll turn our attention to bugs, stability, and enhancements. We won’t make the transition if it’s not stable, and our players will help us decide when Phoenix is ready.

Why can’t you keep Legacy around longer?

Initially, this was what we thought would happen. The longer we’ve lived with two versions of AI Dungeon, we realized how difficult it would be for our small team to maintain both Phoenix and Legacy in perpetuity. Doing so would limit our ability to add new improvements to AI Dungeon and to work on new things like Heroes, a new game mode in development we shared a sneak peek into recently.

Many in our community may not realize we’re a very small team, we have ten full-time employees. Many of those roles are non-developer roles like support, marketing, and design, so our actual development team is very small.

We’ve spread ourselves too thin in the past by trying to make too many things at once. We want to be as focused as we can on making AI Dungeon great for our players, and sometimes that means saying “no” to great ideas.

What are P1 and P2 bugs?

This is a nomenclature commonly used by development teams to triage bugs by severity. The P stands for “Priority”. P1 bugs are ones that prevent the use of critical functions of AI Dungeon. P2 are serious bugs; they may limit core functionality, but there may be a workaround. P3 bugs are ones that impact only a few players, don’t prevent normal use of AI Dungeon, or are simply visual issues.

How will player feedback be addressed?

We will process all player feedback. Please report any bugs and feedback in our bugs_feature_requests channel on Discord. Critical feedback will be implemented during this transition. Less urgent feedback will be added to our product roadmap. We’ve already logged feedback around optimizing the search results display, theming options, and more. We’ll work on these sometime after Phoenix is released. If we get input that isn’t aligned with our product vision, we will let players know we don’t intend to implement the suggested change.

Which features will not be implemented in Phoenix?

There are a few Legacy features we don’t plan to bring to Phoenix. These include multiplayer chat, undo to here, play screen button customization, hydra mode, global world info management and some other minor features. If there’s a feature you think is missing, please let us know!

As always, if you have any questions or feedback, please feel free to email us at [email protected], or join our Discord server and talk with our growing community and team. We appreciate all the feedback we received from our Alpha testers and the broader community. This gave us confidence in the quality of these new AI models and helped us plan for the retirement of Wyvern and Dragon.

We have been working on one of our most transformational AI Dungeon updates ever: Phoenix. Today begins a beta testing period to provide all our players the opportunity to give us feedback before we move forward with a broader release.

Let’s start with the best part: AI Dungeon has been overhauled to the core with a completely rebuilt technology stack and a redesigned front-end interface.

There are many awesome parts of this release and this blog post will walk through 4 key areas we want to provide more detail on:

• Redesign
• Improved AI
• AI Safety Settings
• Beta Environment
  

All of these updates will increase stability, reduce bugs across various platforms, and empower our team to more easily add new elements that upgrade and enhance the game experience over time.

Redesign

When AI Dungeon was first developed, the team was primarily composed of developers and lacked user experience designers. They did a great job making AI Dungeon into a successful game, but not having designers on the team resulted in AI Dungeon feeling clunky and unintuitive at times. Since then, our team has added additional talent who brought important skill sets, including UX design, front-end development, mobile, UX, etc. These awesome team members have allowed us to take the game to a new level, not only designing to meet player needs but also setting a stronger foundation for more successful future releases. This brings us to the redesign announced today.

Each page and feature has been rethought, revised, refined, and refreshed, but you’ll find that, while everything may look different, it’s still instantly familiar. The main game elements and essential play features are still there, but everything around them has been elevated and evolved to be more elegant and intuitive.

To highlight a few main user themes we focused on:

• Interactions—Every tap or click is part of the interaction of the site and defines how users interact with content, other players, and the AI.
• Information Architecture—How the site is organized and how content is displayed on each page. Players need to be able to find what they are looking for in a way that makes sense.
• Focus—We aim for every page to have a clear primary action, purpose, and function, where important information is always prioritized.
• Hierarchy—The design needs to ensure that the most important elements stand out through the use of size, space, color, and other design elements.
• Typography—As a text-based game, the use of type is incredibly important. We have considered readable fonts that pair well, and can be used with different weights and sizes to help with information hierarchy.
• Accessibility—It’s important to us that everyone can use the site. We designed with accessibility in mind, architecting the site in a way that allows everyone to enjoy.
• Performance—No one likes a slow app. We’ve spent time making sure you don’t have to waste yours.
• Intuitive play—We’ve made the first round of many improvements to ensure a better first-time user experience for players just discovering AI Dungeon.
  

A few in-game changes we want to share with you:

Site-wide navigation

With fewer, clearer navigation choices, players will more easily be able to discover the key actions they can take. Daily rewards, notification center, and profile buttons are now grouped together and readily accessible.

Home page

We’ve redesigned the banner section at the top of the home page. The banner section will continue to provide helpful information about updates, events and tips, but now in an even more elegant way. The biggest change to the home page is our new content carousels. We’re introducing several new categories to help players discover more great content. The page will be refreshed regularly so you always have fresh content to explore.

Our new Creators Program is also reflected in the carousel. Each quarter, we identify our top and emerging creators and this group will now have their work featured prominently on the home page.

Content cards

All of our content cards on the site now appear with images. Players have the opportunity to pick an image that best represents their published Adventure, Scenario, or World from a new library of AI-generated images. We know you’d like even more, so we’re exploring ways to enable players to add images that aren’t part of our image library.

“New” button

We replaced the “Play” button with “New” in the upper-right corner on desktop to make it easier for players to quickly start an Adventure and also provide a more efficient way for creators to access Scenario or World templates.

On mobile, you can quickly start an Adventure through the big plus sign at the bottom of your screen. The sliding pop-up lets you quickly choose between an Adventure, Scenario, World, or Multiplayer game to create.

Profile

Many of you will be pleased to see the return of a sorely missed feature: the ability to view a list of your followers, following, and friends. Also under the profile menu is the User Settings page. This User Settings page has been simplified to include basic global actions, like changing your username, email, password, linked accounts, or membership. All other game settings are only available in-game from the play screen now.

Game Screen

Our team has been spending significant time and resources doing more thorough user research, and one area that consistently gives players trouble is the gameplay screen. In particular, players often don’t understand several important actions. The Do/Say/Story/See actions are hidden, and it isn’t clear you can change your action type. Many players also don’t realize they can simply continue the story without typing or taking a turn.

In the game, we now have “Take a Turn”, “Continue”, “Retry”, and “Erase” as our primary action buttons. “Take a Turn” is a new addition that immediately creates a more intuitive role-play experience within AI Dungeon. Selecting this actually opens a brand new way to use our essential Input Modes: Do, Say, Story, and See.

We’ve also made sure that typing your next action is only one click or tap away, just like the legacy version of AI Dungeon. In AI Dungeon today, the text box is always visible, but it still requires a tap to begin typing. With Phoenix, as soon as you tap “Take a Turn”, the cursor is populated into the text box so you’re ready to go immediately.

The old paper airplane icon has been retired in favor of a “Continue” button, which is a much more straightforward way to display and explain that function for new users.

The “Continue” and “Retry” actions now have equal importance alongside the “Take a Turn” action. This is significant because it will help players familiarize themselves with these important actions more quickly.

“Erase” is now included in an effort to separate undo/redo from story actions and include broader user actions, so instead of “undo” meaning “erase”, “erase” means “erase”, which is then an action that can be undone.

In-game settings are still accessible through the gear icon in the upper right corner of the play screen. They’ve been simplified into three categories: Gameplay, AI Models, and Appearance.

You can access the Memory, Author’s Note, and World Info context tools under the Gameplay tab. Under the AI Models tab, you can change the language model you’re playing with, choose a different image generation model, or opt into or out of our “Improve the AI” program. The Appearance tab lets you customize certain aspects of your AI Dungeon experience.

The new game screen features our flame icon in the upper left corner instead of a back arrow. The title of your current Adventure is listed next to the icon so you always know what story you’re playing. Selecting the flame icon brings up links to your profile and resources. A small “Now Playing” card appears below those with the title of your current Adventure, a tiny thumbnail picture, and the option to “Exit Adventure” back to the Home screen.

Learn More About the Redesign

If you want to see more specific details, Devin, our Director of Experience, walked through a few pages as part of the announcement livestream. You can also see Phoenix for yourself by accessing our new beta site at.

AI Safety Settings

As part of the Phoenix release, you’ll see an overhauled AI Safety Setting feature. We previously offered “Safe Mode” within AI Dungeon, and over the last few months, we’ve seen increased user feedback that this feature wasn’t providing the expected amount of control players wanted over their experiences. Players sometimes encountered uncomfortable situations with the AI, even when Safe Mode was turned on.

As we considered this feedback and reexamined this feature, we realized that determining the safety of AI-generated content and protecting our players from potentially harmful material is a very difficult problem to solve, and frankly it’s one we haven’t done well on in the past.

We started looking for a technology partner who specializes in content safety and could provide a better solution than we have. We identified HiveAI as that solution, and have implemented their AI-based scoring technology within AI Dungeon to give players better control over their experience and help them feel safe when interacting with the AI. Depending on the player’s safety settings, as they engage with the AI their model outputs will be sent to Hive for a safety score. If the AI-generated output doesn’t meet the threshold set by the player, we will either provide a different generation or let the user know there was an issue.

As with any partnership, we've taken time to evaluate Hive’s product and terms of service to ensure our players' privacy and security is protected. Hive is well-established in the machine learning field and their product is utilized by many top brands and companies. None of the data we send to Hive will be stored, looked at, or used in any way to train their models. All API requests are encrypted, processed by Hive’s internal servers, and then immediately deleted. Their use of our players’ content is only to generate a safety score during gameplay. We’re excited that Hive is becoming an integral part of AI Dungeon and helping us protect our players.

Our new AI Safety Setting has three levels: Safe, Moderate, and Mature. These correspond roughly with familiar movie ratings. The Safe setting is similar to PG, Moderate is closer to PG-13, and Mature is R, which is the same as our current mature 18+ setting. These levels let players dictate what kind of content the AI can generate, and we can change and adjust them as we receive feedback.

Along with this, we also improved the safety settings for Discover and Search. We’ve intentionally set up this setting to operate independently of the AI Safety Settings, since we know some players like to create their own NSFW content, but prefer not to see other people’s NSFW content when browsing. You’ll see a new search-specific filter in Discover that will simply display or not display NSFW-tagged results. The AI Safety Setting and Search Filter both default to Safe, and players can change those themselves. Additionally, any published worlds, scenarios, or adventures, will be sent to Hive and receive a safety score, to help prevent mistagged content from showing up in the wrong places.

Ultimately, our updated AI Safety Setting and Safe Search Filter are just more ways players get to customize and control their individual AI Dungeon experience. They provide additional safety controls for those who want them, while still offering flexibility for those who don’t.

Improved AI

We’re experiencing a renaissance in AI right now, with new models and innovations hitting the market faster than ever. We’ve been capitalizing on these innovations, and are excited to announce two new models to be added to AI Dungeon.

New Dragon

Dragon has traditionally been perceived by our players as our best model. In the last few months, Wyvern has eclipsed Dragon in many ways, taking the spot of players’ most preferred premium model. Now, the balance is shifting back to Dragon as it evolves with our new update.

The new Dragon will be based on the Jurassic-2 Grande model provided by AI21 Studios. This model is powerful, but we’ve taken it even further with new enhancements we’ve recently developed. The new Dragon model has quickly become a favorite with our Alpha testers. Some of our longtime players consider the GPT-3 based version of Dragon we offered a few years ago as the gold standard of AI models. The new Dragon is so good that Alpha testers are reporting that it’s exhibiting the same quality and storytelling capabilities they remember from the old Dragon.

AI21’s new Jurrassic-2 models have eclipsed their old models in every way. Because of that, they’ve announced they will be ending support of their old Jurassic-1 models. As a result, we’ll be retiring Legacy Dragon and Wyvern. However, given the success of our tests with the new Dragon, we’re confident these outgoing models will not be missed.

The new Dragon offers several key advantages over any other version of Dragon we’ve offered before. The new Dragon is fast—responses come in at nearly half the time as Legacy Dragon. This has always been one of the disadvantages of Legacy Dragon, and Alpha testers have already shared how much the improved speed has improved their play experience. The new Dragon will also support more options and possibilities for training and improving the model over time. And probably most exciting the new Dragon will support larger memory. Adventure will get 1x memory, Hero will get 2x, and Legend will get 4x. This means the AI can consider more of your story when generating the next output, ensuring key details and historical story events are not missed.

The new Dragon will replace Wyvern as the new default model for our AI Dungeon subscribers. It will be available in both the Phoenix beta, and the production version of AI Dungeon.

ChatGPT Beta

We’re also excited to introduce ChatGPT through Microsoft Azure as a new beta model in AI Dungeon. Players have shared with us how impressed they are with the quality of ChatGPT, and we received numerous requests to add it as a playable model in AI Dungeon. In fact. our March player survey showed that over 70% of players who responded were excited about the idea of ChatGPT + AI Dungeon.

The partnership with Microsoft Azure is a new one for us, so we are taking it slowly, and have sought out player feedback at every step to ensure that this is what’s best for AI Dungeon.

For context, Microsoft is the only company besides OpenAI to have access to the ChatGPT model. We chose to use Microsoft Azure for a few specific reasons, the most important being their privacy and security reputation. This is crucial to our players’ experiences in AI Dungeon, as well as to us.

Some players may wonder if Azure sends data to OpenAI and they do NOT. Microsoft hosts the OpenAI models within their own Azure infrastructure, and all customer data sent to Azure remains within the Azure service. This means that data processed by Azure is never sent to OpenAI. Microsoft runs ChatGPT on their own servers with strong security and privacy protections they’ve developed as part of the Azure stack.

We utilized our entire AI evaluation tool kit to determine whether we should offer ChatGPT from Azure long-term, and early testers have compared it to being as good as or better than old Dragon, so we’re excited to start offering this model alongside our others. During this testing period, Adventurers can use ChatGPT with the standard memory offering, Hero-tier subscribers will have access to 2X memory, and Legend-tier to 4X memory.

ChatGPT has performed extremely well for adventures and has the potential to greatly improve areas where other models have struggled in the past, like memory, repetition, and creativity. However, we will continue to gather more feedback, answer questions, and address concerns before moving the model to production full-time.

Using ChatGPT is entirely opt-in for premium users only. With the new version of Dragon, our premium players will have two fantastic models to choose from for their adventures. A new Griffin model is also being developed based on the improvements we implemented in the new Dragon.

Beta Environment

Starting today, we are launching a brand new Beta Environment of AI Dungeon for Phoenix. We previously offered players a chance to try new Beta features by enabling beta features as a setting on their account. This new, completely separate environment lets players explore new features, and then have the option to go back to the stable version if desired.

The previous approach worked well in the early days of AI Dungeon when new features were being added almost daily. It allowed developers to move quickly. Moving to a standalone beta environment will allow new features and changes to be further isolated from the stable version of AI Dungeon, ensuring that new developments don’t break the core experience.

Going forward, new features and major updates will first be released to this Beta Environment. In beta, changes will be thoroughly tested, reviewed by players, and validated before being added to the stable version of AI Dungeon.

The redesign we’ve announced today will only be available in the beta environment during the testing period. After we do more testing, migrate all legacy features, gather feedback from players, and validate everything is working as expected, we'll deploy this Phoenix update to production and officially retire the legacy version of AI Dungeon.

Not all legacy features will be available at the start of the Beta testing period. Features that are still in progress for inclusion in Phoenix in the coming weeks include:

• World Info
• Multiplayer
• Native Mobile Apps
• Scripting
• Image generation (See Mode)
• Train the AI

However, the data flows seamlessly between the Beta and Production version of AI Dungeon, so if there’s a feature that’s missing, you can still use it in the Legacy version.

You can access Beta via: beta.aidungeon.com using same login credentials as you use for AI Dungeon.

Coming Soon

Phoenix addresses much of the remaining player feedback we discussed in our October 2022 blog post, “Next Steps on AI Dungeon,”  by bringing even more stability, better AI quality, and refreshed designs to AI Dungeon. Of course, maintaining and further improving AI Dungeon is an ongoing process that we look forward to working on with our next releases. We also have plenty of ideas to continue growing AI Dungeon and will put out a new blog post detailing those plans when we’re ready to talk about them.

Thank you to our entire player base, alpha testers and Discord contributors–your knowledge and feedback are invaluable to us and vital to providing a better experience for all players. We are extremely grateful for your dedication and commitment to AI Dungeon!

We intend to utilize our entire AI evaluation tool kit to determine whether we should offer ChatGPT from Azure long-term. This includes both quantitative evaluation methods (like “Train the AI” and “Try a New AI”) as well as qualitative feedback from players. Our Alpha testers will also be testing and providing additional feedback to aid the evaluation process.

This test is completely opt-in for players. Our AI testing will only include those who have turned on “Improve the AI” in their settings.

Testing Details

During the testing period, ChatGPT will not be a selectable model option but may be used in “Train the AI” and “Try a New AI” tests. If these initial tests go well, we plan to continue testing ChatGPT by offering it as a beta model for additional testing from our players.

Azure does have additional filtering for their ChatGPT offering. We will be testing these filters as part of our testing evaluation.

We look forward to your feedback!

Thanks again to our players for letting us know they’d like us to evaluate ChatGPT. We welcome all player feedback during the testing phase. Enabling “Improve the AI” is the best way to help us evaluate ChatGPT from Azure, but we are also happy to hear your feedback in Discord or by emailing our team at [email protected] .

AI quality is one of the most important things users want us to improve. After several months of work, we’re releasing a new and significantly improved Wyvern 2.0 model to beta TODAY! A new Griffin model based on similar changes will follow soon after. We’d also like to share a number of significant improvements we’ve made to our AI model infrastructure to enable faster, more stable, and more continuous improvements to the AI.

New Wyvern & Griffin Models

Wyvern 2.0

Starting today, premium users will have access to a new beta Wyvern model which they can try out by switching to "adventure-wyvern-v2.0 (beta)”. This model has a number of significant improvements, and on our “Improve the AI” tests it outperforms Wyvern 1.2 by 38%!

Wyvern 2.0 is based on a new, more intelligent foundation model from AI21. We also trained it on an expanded and more thoroughly cleaned dataset that is 5X bigger than our previous dataset.

In addition, we updated how we process and handle AI model outputs leading to a significant reduction in empty outputs and repetition. Our evaluation suite tests show 75% fewer empty output errors and a 20% drop in repetition.

Griffin 2.0

But these improvements aren’t just coming to Wyvern, we are currently polishing a new Griffin model for free users too! This new Griffin model will take advantage of many of the same improvements to make the free user experience better as well.

New AI Processes and Systems

A few months ago we talked about our plans to improve the AI. AI quality remains one of the most important areas players request improvements to.

Since then, we've done a significant overhaul of our AI model system so that we can build, evaluate, and improve AI models much faster and with much greater stability. We’ve dedicated more resources internally to continually improve the most important part of the game experience.

It’s also been unclear to players when their AI experience (good or bad) was caused by updates to the models or simply due to the randomness of the AI. We believe by better communicating AI versioning will clarify when and how the AI has changed.

In the past, we lacked evaluation systems to confidently assess whether changes improved the AI experience. With these new systems we can be much more thorough in testing new AI versions, letting us make sure that changes to the AI only improve the experience.

This will extend the review process for new models to reach production. But when we release them they will be better tested and clear evidence they improve the experience.

AI Version System

In order to improve how we communicate and evaluate AI systems, we realized that we needed to be more explicit about different AI versions and what changes they represent. An AI version system helps us communicate more clearly to users when and how the AI is changing, and also evaluate AI changes in a more systematic way.

We refactored all of our systems to be based on an “AI Version.” An AI version includes not just the model it leverages but also the code that constructs the prompt, configures the model request, transforms the outputs, and picks which output to deliver to users.

Now we can quickly create new AI versions, evaluate them in various ways, and promote them to alpha, beta, or production depending on their progress. Models will go through a more thorough and careful process before being promoted to the production experience.

AI Core System

To enable versioning of all the code involved with the AI, we re-architected our systems so that each AI version is connected to an AI “Core”. The AI Core includes all the processing functions for how we structure prompts, pass them into the model, transform the outputs, filter out bad responses, and rank which response we should deliver to the users.

With our new AI core system, we can easily test and evaluate not just different AI models but also all the code that affects how well those models are used to deliver outputs to users.

A Dedicated Team Member Improving Data Quality

Training data plays a major role in how the AI feels to players. We’ve dedicated resources to building up, cleaning, and formatting our datasets to improve AI Dungeons’ models. We now have a team member dedicated to data quality who is constantly working on improving our datasets. This has enabled us to significantly improve both the size and quality of our datasets.

Improve the AI

We’ve relied on “Train the AI” to help us evaluate AI models, but we realized we needed additional data to help us measure model quality. We’ve expanded this functionality into a new system called, “Improve the AI” that allows users to opt-in to sharing anonymous model data with us. This makes it much easier to debug problems and understand exactly where in the AI pipeline issues were happening. With Improve the AI, developers can see more clearly where problems are happening and fix them by evaluating data from users who have opted into sharing that data.

New AI Metric and Evaluation Systems

While testing with users is the best way to evaluate the AI experience, it became clear that we also need ways to quickly evaluate AI changes before testing with users.

To do this, we created an “evaluation suite,” where we pass in a large dataset of prompts into models. We then compare the results from different models across metrics like repetition, empty outputs, and more. This lets us quickly measure several aspects of AI quality, enabling us to iterate and improve the AI at a much faster pace.

New Framework for AI Quality

As we spent more time thinking about improving the AI experience, we realized what “AI quality” means can be ambiguous. We developed a 6-part framework for AI quality that we now use in evaluation and improvement processes.

These 6 parts are:

• Creativity: How creative is the AI’s writing? Is it interesting and compelling? Does it make you want to keep going to find out what might happen next?
• Style: How well does the AI fit the style a user wants? How well does it stay within a certain world or genre?
• Repetition: How repetitive are AI’s responses?
• Logic: How coherent is the AI’s writing? Does it make sense given what happened before?
• Memory: How well does the AI remember and leverage info about the world and what happened in the story’s past?
• Form: How well does the AI output things that are in good form? Are there hanging sentences, answers that are too short, etc…?

Try a New AI

Finally, we have a new system we’ll release next week called “Try a New AI.” One shortcoming with Train the AI is that because it only evaluates a single action, the measurement doesn’t help us understand how models perform over an entire adventure.  It also didn’t include any way for users to give more qualitative feedback on why they liked or didn’t like an AI version.

To address this, we’re adding a new evaluation method where users can opt in to “Try a New AI.” If they do, they are randomly assigned to one of two models without knowing which one they’re on. They will be able to opt-out and switch back to their normal models at any time. When they are finished with the test, they will be asked to rate the AI across the different aspects of AI quality.

This new system will give us greater confidence in our ability to measure all the different aspects of AI quality and have a greater understanding of how different AI versions perform and how we can improve them.

We’re excited to get your feedback on the new models and will continually work on improving the AI experience to be the best it can be.

We also created a new page in our Guidebook to communicate these improvements with our players. Here, anyone can see more details about models in production, models we’re testing, and other features we’re working on improving. Each model has a version number, and model versions will change when we:

• Update and improve fine-tune data sets
• Change default generation settings
• Upgrade output processing functions
• Improve prompt construction

We sincerely want AI Dungeon stories and adventures to be coherent, meaningful, surprising, and fun. Improving the models behind that gameplay is a challenging process that will continue to evolve, but we are confident this new system will help players have a continually better experience in AI Dungeon, as well as clearly understand when and how the AI is changing. We’re excited to see our models get even stronger over time!

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.

The max context size for Wyvern has now doubled from 1024 tokens to 2048!

With double the context size you should have fewer instances of the AI forgetting characters' names or attributes, losing track of narrative lines, etc. It also means more rich and deep stories that can recall more characters and scenes from World Info.

What does the context size mean?

Language models like Wyvern can only analyze a limited amount of text at one time, and this is called "context." When we send a request to the language model, we're not just sending the prompt text you just typed in. We combine your prompt text with any active World Info, data stored in your Memory, Author's Notes, and as much of your recent story text as we can fit and send this to the language model. With double the context size the AI can use more information and not “forget” details in your story.

Learn more about Wyvern in the AI Dungeon Help Guide.

We have launched a new high-quality image model, Stable Diffusion, for inline generation of images in AI Dungeon. Stable Diffusion is an incredible new text-to-image AI model that will generate a unique image from a simple text prompt. We’re excited about Stable Diffusion because not only are the images high quality, but the generation times and costs are lower than other models. We can now enable custom image generation in AI Dungeon adventures, unlike the current version of AI Dungeon 2D which injects images into stories from a pre-generated cache.

Changes we’re making to support image generation

Since Stable Diffusion makes inline image generation possible, we are making a number of changes to AI Dungeon to support image generation.

First, we’re changing how images are displayed in adventures. Instead of the two-column layout on larger screens, images will now be displayed at a larger size and full width. For wide screens, they’ll be placed over a faded background to fill the horizontal space. This design makes the images feel like a more integral part of the adventure—and they are!

Second, we’re adding a new AI setting to change the image model you are using. You can select between the Pixel Art image cache and Stable Diffusion image generation. Another image cache is coming soon, with images generated using Stable Diffusion, DALL-E 2, and other image models.

Third, we’re adding a new action, “See”, that will join the Do, Say, and Story actions. When selecting the “See” action, what you type will be submitted as an image prompt. This action will respect the image model selection you’ve chosen in your settings. If you have Stable Diffusion Image Generation selected as your image model, the prompt will generate a unique image using Stable Diffusion. If you’ve selected one of the image caches, the prompt will search the cache for a relevant image to insert into your story.

Fourth, we’re improving the editing interactions for text. The dropdown menu for images and text will now include the ability to add an image to an existing action, generate/fetch another image, or remove an image. We’re also auto-saving any text edits when clicking away from the text input. The new interaction is a first step towards document-style editing in AI Dungeon. Save and cancel buttons have been removed from the dropdown menu.

Players will now control when images are displayed in their stories, and what those images depict. This control means the images will be far more meaningful for the stories players create.

Changes to the AI Dungeon Economy

We’ve known for a while that our economy system (Scales, Energy, and Actions) needed changes to accommodate the rapidly changing landscape of AI models. We’ve made several significant changes to accommodate Stable Diffusion and be more resilient to future changes and innovations.

Here are a few highlights about this new system:

“Image Credits” are being renamed “Credits,” and Credits will be the currency of AI Dungeon and Voyage. When we transition from “Image Credits” to “Credits,” we’re going to 20x the totals of Credits to allow us to charge less for Stable Diffusion. We will multiply players’ current Credit amounts, and our other image models will require 20+ Credits for each image generation instead of 1 Image Credit.

Members will now receive Credits each month instead of Scales. Scales can still be used to tip other players or creators. Credits can be purchased on all platforms (except for Steam). Scales will no longer be available for purchase, but players will be able to trade Credits for Scales.

NSFW Images

NSFW image creation will be allowed in AI Dungeon, and we expect the content policy for images to mirror our current policies with text. We are working with Stability.ai, the makers of Stable Diffusion, to ensure measures are in place to prevent generating certain types of content—mainly content depicting the sexual exploitation of children. These measures would apply to both published and unpublished stories.

Unpublished images, like unpublished stories, will not be moderated. Players will not be punished or banned for anything they create in unpublished stories using AI models. Published stories are still subject to our Community Guidelines, and these guidelines will still be upheld when image publishing is ready. Images won’t be included in published stories until we figure out the right experience and safeguards for sharing images publicly.

We know we won’t be able to anticipate every situation regarding AI-generated images. Because this is cutting-edge technology, there is no clear precedent for how (or if) AI-generated images should be governed. There are several unanswered questions about the use of AI images that all of us will be working through as AI image models become more accessible. We’ve learned from the past and will be actively communicating any changes or adjustments that need to be made to our content policies as we learn more about how players will use this powerful technology.

More to come

We’re enjoying an exciting time in AI, with new models and capabilities being introduced regularly. Our plan is to bring as many of those innovations into AI Dungeon as quickly we can.

Here's a preview of some of the incredible images generated in Alpha recently.

After consistent feedback, we’ve reintroduced AI Dungeon subscription plans and mobile payment options. We took the opportunity to rebuild the plans and payment systems to be more stable, consistent, and flexible than before. While these plans have only been available in Beta for the past few weeks, the overwhelmingly positive response from players tells us that these have indeed been sorely missed. We’re excited to announce that the phased rollout of these plans to all users has begun!  

AI Dungeon Membership Plans

We're launching 3 new plans: Adventurer, Hero, and Legend. Current subscribers will be automatically upgraded into these new plans since benefits either improve or remain the same. Gold subscribers will become Adventurers. Voyage subscribers will become Heroes. Platinum subscribers will become Legends. The Silver tier will become Travelers, and going forward these benefits will only be available through one-time purchases on Steam.

For full details about these new plans, their benefits, and prices, please refer to our help center article.

Gold, Credits, Scales, and a new AI Dungeon Economy

The AI Dungeon economy is changing. We're introducing Gold, a resource for members to spend on Dragon, or hoard for rewards. Scales are back, and we are expanding the purpose of credits.

Get an overview of the AI Dungeon Economy in our Guidebook.

Thanks for the feedback

We appreciate all the helpful feedback we’ve received from our community. Bringing back AI Dungeon plans, mobile payments, and ways to use Dragon Scales was a direct response to requests and feedback from our players. Please let us know if you have any questions or feedback on these announcements and happy adventuring!

We’re Making Changes Based on Your Feedback

Successful product teams often do retrospectives—an exercise where teams meet together to review their work and identify what went well, and where improvements can be made. Retrospectives help the team to continue the good things they are doing and learn from mistakes.

We’d like to share a retrospective on our recent ad release with you. We want to apologize for the mistakes and share our plans to improve future releases.

Things we'll do better

Our community has been understandably frustrated with the way we released ads. We knew there would be some negative response to ads, as we had similar reactions when we launched energy. However, player frustration wasn’t only about the ads. We launched with too many bugs, our communication was unclear, and we removed legacy features too soon. We’re sorry we let this happen, and we’re working to make it right. Here’s how we’re going to do better.

More Testing, Fewer Bugs

What you told us: The ad launch had too many bugs and issues.

Ads are notoriously difficult to implement. While we anticipated there’d be some issues after launching, the number of critical issues our players experienced was unacceptable. Some players were getting ads after every action (which was NEVER what we wanted), or were completely unable to play on their combination of browser and adblocker because we hadn’t tested enough variations. We also didn’t have an easy way to revert the experience for players experiencing issues.

How we’re changing: New features will be released to Beta, which any player (free or subscriber) can opt into in order to try new features and report issues. If Beta is unstable, players will be able to go back to the general release by disabling Beta features. We’ll make changes and fix issues until Beta is ready, then update the general release version. Large or high-risk features will roll out to players incrementally, which will allow more space for feedback and discovery of potential issues. We’re also developing more robust testing processes and checklists for launches.

Transparent & Clear Communication

What you told us: The ad solution we implemented wasn’t what players thought they were voting for. It wasn’t clear how ads would be implemented, or why.

We got feedback about energy and ads from thousands of players. The feedback came from surveys, product feedback, support tickets, testing groups, user testing, and email. However, in many of these places our proposed solution was vague or unclear. For example, on our feature request page, the language of that request said, “use ads to increase energy”. This was player-submitted language, but it understandably influenced what people thought they were voting for. In other places we talked about ads but didn’t mention whether we were considering banner ads, pop ups, or interstitial ads. In the planning stages we weren’t sure what ad types would be implemented, but as we got further into the development process we should have updated our communications to reflect our new direction. We realize that some people would have changed their feedback if we had been clearer about the solution we were considering.

Changes we’ve already made: We are already communicating more frequently and on more channels. We’ve published a number of blog entries, social posts, and email campaigns. Multiple members of the executive team are participating more frequently with our community on Reddit, Discord, and livestreams.

How we’re changing (further): Get ready for detail, because we’re sharing more than ever. We’ve already started sharing some product metrics, which is something we’ve shied away from in the past. Our product livestreams will feature more details of current work, let us gather feedback from the community on new designs / concepts, and give us a chance to discuss what we’re learning. We’re also working on a blog post detailing everything that happened with filters in 2021.

Better transition of legacy features

What you told us: Do a better job migrating legacy features when updating

We’re anxious to make AI Dungeon better and, at times, that’s led us to be too quick to remove and change features without sufficient communication or solid transition plans. Recently, for example, when refactoring the User Profile we removed the ability to purchase avatars with Scales. That meant when we removed energy to make way for ads, we removed the last place scales could be spent. Additionally, a significant performance issue forced us to remove Follower/Following counts from the Profile page, a very important metric for our players, due to a performance issue requiring significant developer time to fix—all of which we should have communicated more clearly. In sum, while everyone expects the product to change, evolve, and improve, the Latitude Team can and should do better to clearly communicate our intent and our transition plans to make these changes more smoothly.

How we’re changing: Although we may continue to remove legacy features, we’ll communicate those intentions and have clearer transition plans. We will be bringing back or transitioning some features that we removed. You can read about the plans for scales below. If things break, we may not have a transition plan ready, but will better communicate what is happening and why.

A Gift from Latitude

We know that these changes have been frustrating for players. We want to give something back to the community to show our appreciation for taking this journey with us. Effective immediately, AI Dungeon 2D is available for ALL users—free and subscribed. At launch, it’ll be available to users with Beta features enabled, and will be rolled out to everyone soon.

With AI Dungeon 2D, AI generated images will be used to illustrate the events of your story. It’s a fun way to bring your adventures to life with visuals. You can enable AI Dungeon 2D in your game Settings.

Upcoming Product Changes

Store Up Actions Using Ads

Players have consistently brought up that the thing they missed most about energy was the ability to save a pool of energy and use it for an uninterrupted session of AI adventuring. In fact, one of the most common suggestions we’ve heard is to use ads to recharge energy.

We’re making several changes to allow players to store up actions. Some of these changes are available now in Beta. Other changes will roll out over the next few weeks. Here’s some highlights of what is coming:

• We’re adding a “Manage Actions” window that is accessed by clicking the video icon at the top of the play screen.
• From the “Manage Actions” window, you’ll be able to see how many actions you have. You can choose to watch an ad to get more actions or, if you have Scales, you can redeem them for more actions. (If you don’t have any scales this option will not be available). Redeeming scales for actions is coming soon.
• Players can watch ads repeatedly to store up actions. Some ad networks and regions limit how many ads can be shown back-to-back. We may have to limit ad watching in some cases, but we’re planning to use multiple ad networks and, potentially, custom ads to avoid this.
• Our goal is to let players store an unlimited number of actions, but there will be some limits when we launch. We’ll remove those over time if/when possible.
• The Daily Reward will allow you to earn extra ad-free actions.

There are some parts of this solution we already know we need to test and get feedback on. For instance:

• Would players prefer a flat rate of actions per ad watch? Or would they prefer a variable reward system that offers them a chance at getting more actions?
• How many actions should we offer a player per ad? We may test different amounts in the coming weeks.
• Could some ads be worth more than others? For instance, could watching a 30 second ad be worth more actions than a 10 second ad?
• Will our ad providers require us to rate-limit ad watching? Is there a limit to how many ads a person can see in a row or in a specific time-frame?
• What should we do for geographic regions where ads are not as available? Can we create Latitude ads that fill the space in the absence of ads from our partners?
  

Bringing Back Legacy Payment Plans & Options

In addition to the feedback on ads, many people also shared frustrations with our current subscription plans and payment options.

Why AI Dungeon Plans were removed: Many of our players have noticed that we’re currently offering Voyage plans exclusively. We retired the limited AI Dungeon subscriptions for a couple reasons. Our Voyage plans were simpler to support, as they all went through a single payment provider (Stripe). We also believed having one, single plan to choose from would actually be simpler for players, since we’ve had difficulty helping players understand the difference between the different AI Dungeon tiers.

Why mobile payments were removed: We also removed mobile payment options, which have always been more difficult to support. Apple and Google’s systems don’t allow us to process cancellations or refunds for customers. They also collect large royalty payments from our customers, and add an extra layer of complexity to our experience. We tested sending mobile users to Voyage to pay for subscriptions, and in that test, users were subscribing at the same rate as mobile subscriptions. We concluded it didn’t make sense to pay extra money for a mobile payment system that didn’t perform better, and didn’t allow us to support our players the way we’d like to.

We’re bringing AI Dungeon plans and mobile payments back: We’ve received clear feedback that the various payment options were missed. So, we’re going to offer AI Dungeon plans again. We will also bring back mobile payment options. More details about the new plans, and what’s changing, will be available soon.

More exciting changes

We’re not just improving ads, reintroducing AI Dungeon plans, returning mobile payments, and bringing back using Scales for actions—we’re also excited to share more features coming to AI Dungeon.

First, comments are back. We removed comments last year because of challenges with moderation. We now have a solution in place that will help us keep comments safe. This is available now as a Beta feature.

Next, we’re improving the Model Settings interface. You’ll soon be able to make model changes in the play screen without having to leave your story.

Finally, we are planning to introduce Custom Fine Tunes for subscribers. We’ve already experimented with Fine Tunes that emulate the writing of famous authors, or are tuned to specific fictional universes. Eventually, subscribers will also be able to create fine tunes themselves, and even share them with others in the community. Show us how it’s done!

Questions

1. Why can’t we just go back to energy?
   
   Energy was implemented to limit extreme usage of AI, and was just as unpopular as ads when we launched it. But energy didn’t cover the costs of free players. Supporting free play on AI Dungeon costs around half a million dollars a year in compute power. AI is expensive.
   
   This is why our commitment to a free experience in AI Dungeon is unique. Offering free AI compute is a quick way to lose a lot of money. We want to be able to provide players an option to create with AI for free, and ads help us do that sustainably—or, at least, help cover some of the costs.
   
   Nearly every platform we use online that’s free also uses ads. Youtube, Google, Twitter, TikTok, and free-to-play games, all use ads to support their free services and platforms. TV channels have had commercials almost since TV started broadcasting. Pay for a movie, and they have you watch ads in the theater before the movie starts. Ads are the clearest way for businesses to support free use of platforms, or augment other revenue streams.
   
   
2. Why aren’t you using banner ads?
   
   Because of the high costs of supporting free AI play, banner ads are not a useful option. Banner ads generate 50-100 times LESS revenue than video ads.
   
   
3. Why would you make players watch ads?
   
   No one likes ads, and we’re no different. It’s a big reason we’d waited so long to implement them. But ad revenue lets companies provide things for free that otherwise wouldn’t be possible. If we didn’t implement ads, we likely would have had to follow other AI apps which offer extremely limited or no free-play options.
   
4. Is everyone leaving AI Dungeon?
   
   No. While we know of a few individuals who said they were done over the shift to ads, our active usage is actually up since we announced ads and we’ve had to increase the number of our Griffin servers. We’ve actually had some people come back to AI Dungeon, specifically citing they prefer ads over energy.
   
   
5. Can we have AI Dungeon only plans back?
   
   Yes. See the info above. 🙂

Moving Forward

Again, we apologize. We want to, and can, do better. We wouldn’t be here without the support of millions of players who have ridden this roller coaster with us.

We’re looking forward to completing the changes and fixes outlined above. Thank you for caring enough to give the feedback we needed to make the right adjustments.

Take care, and enjoy your next adventure!

—The Latitude Team

We want all AI Dungeon players to enjoy endless AI-driven adventures. Unfortunately, some adventures have been cut short by the energy system, which was introduced to prevent excessive use of AI server capacity. Within the energy system, a player has a limited number of AI actions they can take before their energy is consumed, and then they must wait for the energy to recharge before continuing their AI adventure. Unsurprisingly, the limitations of the energy system have become one of the most common themes we see in requests, feedback, and complaints. Players frequently share their thoughts on ways we can improve the energy system.

One idea in particular gained traction—replacing the energy system with ads for free players. Frankly, we resisted the idea of ads for a long time, but the consistent requests from players finally convinced us to take a closer look.

We spent a great deal of time researching other experiences that use ads as a cost-covering mechanism. We looked at social platforms like Youtube, Twitter, and Facebook, which serve ads before video content. We also looked at streaming platforms like Hulu, HBO Max, and Peacock, which show ads in addition to a monthly subscription charge. We also looked at ads in games. Since then, we’ve built and tested different ad experiences in AI Dungeon that our community gave us feedback on; this has helped us iterate and refine the experience.

We realized that including ads for free players lets us give them more content, making a win-win for all. All subscribed users can continue to enjoy AI Dungeon as an ad-free experience.

We’ve been pleasantly surprised how excited players are for ads! In addition to gathering feedback from our testing group and our Discord community, we recently surveyed AI Dungeon players. We found that the majority of free players prefer ads vs the energy system. A full 80% of our players support developers using ads to cover the costs of free game play.

After a lot of work and iteration with community members we’re now able to offer endless AI Dungeon to all players, whether they have a paid subscription or not.

How endless AI Dungeon will work

Ads will only be shown to players using the free version of AI Dungeon. Subscribers to Voyage or our AI Dungeon Silver, Gold, and Platinum tiers will continue to enjoy an ad-free AI Dungeon experience. Subscribers will also continue to have other premium features, such as our advanced AI models, AI Dungeon 2D, advanced settings, and more.

For players using the free version of AI Dungeon, ads will be shown occasionally during adventures while players are waiting for the AI to generate the next step in the adventure. After the player submits a prompt, the ad screen loads, a timer will begin and will count down as the ad is being shown. Once the timer has counted down, the player will be able to close the ad window. Since it takes time for AI generations to be ready, the ad viewing will simply fill the time a player would normally spend waiting for the AI response.

Unlike the energy system, the ad-supported experience can be continued indefinitely. As long as you keep watching ads, you can continue to play AI Dungeon endlessly. The days of waiting for energy to recharge are over.

Multiple options to remove ads

We understand that not everyone will be interested in playing an ad-based version of AI Dungeon. In fact, 64% of AI Dungeon players would be interested in a one-time payment to remove ads. That will be possible through the Steam version of AI Dungeon.

We have recently announced that AI Dungeon is coming to Steam. Players who purchase AI Dungeon on Steam can play endless AI Dungeon without ads. In addition to no ads, Steam users will also unlock premium features like AI Dungeon 2D and AI settings.

Additionally, all subscriptions will continue to be ad free.

We welcome your feedback

Just as we’ve done with other features and improvements on AI Dungeon, we plan to iterate and improve on the ad solution we’re launching. We will continue to test different ad experiences, ad platforms, ad types, how frequently ads are shown, and where ads are shown. We also plan to test unlocking certain premium features by opting in to watching more ads (87% of players said they’d watch ads to unlock premium features). Our players are excited to have new ways to enjoy more of the AI Dungeon experience for free.

Please continue to send us feedback as we retire the energy system and roll out Endless AI Dungeon for all. And most importantly, enjoy your next AI adventure!

Language model size is measured in “parameters”. A small model might have 1.5 billion parameters. A medium model, 13 billion parameters. Large models have as many as 175 billion parameters, and there are even bigger models that aren’t currently feasible for consumer use. In theory, the more parameters a model has, the more knowledge and understanding the model has.

Computing power is measured in TFLOPS (trillions of floating-point operations). The Playstation 5 is capable of processing 10.4 TFLOPS per second. The iPhone 13 Pro Max, Apple’s top-of-the-line smartphone, is capable of 1.5 TFLOPS per second.

To give you an example of how much computer power is needed to train an AI language model, consider GPT-3, a large language model with over 175 billion parameters. According to one estimate, training the GPT-3 neural network required 311,400,000,000 TFLOPS. That’s 311 billion. With a B.

Performing this training on a V100 GPU server (which is an “affordable” supercomputer that costs over $10,000 to purchase) would theoretically take 355 years to complete. Additionally, if you rented space on this machine at the market rate of $1.50 per hour, you would end up spending $4.6 million to train the model.

The fact that language models like GPT-3 even exist is remarkable, and represents profound technological and human innovation.

Compute required to use an AI Language Model

Once the language model has been trained, it still requires intensive computing power to utilize in a game or application. In AI Dungeon, for example, we make a call to the language model every time someone types a new action into their adventure story. These actions are sent to the AI, a response is generated, and then sent back to be utilized in the AI Dungeon interface.

Let’s imagine you wanted to run AI Dungeon on a desktop computer. It’s something that’s actually been requested before, and community members even built a version (available here)! However, it’s not what you might expect. This version of AI Dungeon “is much less coherent” because it relies on a small-sized language model. Trying to run AI Dungeon’s medium-sized language models on a modern desktop computer would be like trying to play a modern 3D video game like Call of Duty on a 1990s PC running floppy disks.

In one thread, a Reddit user correctly stated, “If you want to run your own full GPT-3 (Large) instance you’ll need this $200,000 DGX A100 graphics card from NVidia. ...and you might need two of them. (5 petaflops / 320GB video memory).” In response, another user pointed out “So....maybe 3-4 generations till it's feasible at home. Not bad.”

That’s optimistic.

AI language models are becoming affordable for consumers

Given the costs required to train and use an AI language model, it’s surprising how many consumer-facing products exist that use language models at prices that individuals can afford.

Jasper charges $29/month for around 200 generations (assuming 100 words or tokens / generation). For 2000 generations with 100 tokens they charge $169/month for their starter package or $279/month for their advanced package.

Novel AI offers 100 generations as a free trial and has paid plans for $10, $15, or $25 a month. They utilize medium-size models (6B, 13B, and 20B) with multiple configurations.

AI Dungeon players enjoy unlimited rate limited/ad supported Medium (6B) model generations for free. By subscribing to Adventurer or Hero memberships ($10/month and $15/month respectively), players get 2600+ generations per month from our Large model. Legend members ($30/month) get unlimited generations from our Large model as well.

The future is even more exciting

AI continues to benefit from advancements in technology and innovation. Language models are becoming more affordable. Before long, AI experiences will be just as affordable as traditional games or digital experiences.

We’re preparing to launch Voyage Studio so that anyone can create AI-powered experiences. We spend significant time and energy ensuring AI usage is sustainable cost-wise for creators and players. Our experience building and growing AI Dungeon has taught us a lot about how AI-powered games can sustainably run. As performance improves and costs come down, more people will be able to experience the joy of AI-based creation tools, experiences, and games.

Meet AI Art Studio — your very own AI art generator. By simply selecting an art style and providing a simple description to the AI, it will generate beautiful works of art ready for you to download and share.

It's like having your own personal artist who creates art to your liking, on demand, whenever you ask them to. With multiple image generation models and infinite possibilities, what would you have them create?

Here are a few examples from our community:

AI Art Studio is an exciting step forward for us because it’s the beginning of a new type of AI experiences where players can co-create visual experiences, not just text based ones.

How to use AI Art Studio

After launching AI Art Studio, we encourage you to browse the Gallery for ideas and inspiration on what can be generated.

By clicking on any of the Gallery images, you can view and understand how the model, prompt, and parameters can be set to generate a certain type of image. Pro Tip: Some creators share instructions on how to copy the style of particular artists.

When you're ready to create your own AI Art, you'll be able to do so through the +CREATE button. Choose from different AI art models, and each will have its own unique parameters that you can configure.

You can easily download your art once it's ready and share it however you please. Use #AIartstudio to tag your work for other creators to find!

Players will receive monthly image credits based on their subscription tier. Voyagers will receive 20 credits a month, and AI Dungeon Platinum subscribers will receive 100. If you run out, you can purchase more credits.

Finally, our team has put together some tips to help you get started with creating AI art. Click here to read tips straight from the developers.

We can't wait to see what you create in AI Art Studio.

Sometimes, the images produced by AI Art can be monstrous. Here’s an example of what I mean:

The text prompt for this was “a beautiful illustration of a squirrel by Lizbeth Zwerger.” Now, I find this image kind of fascinating. It has some nice flowing curves, beautiful detail on the fur, even an attempt at a signature (I think I see an L, I and Z in there). If that’s what you want, more power to you. Good art is often disturbing. But if you want a beautiful illustration of a squirrel, this isn’t it. So why did the model produce this?

Once you understand how the model works, it’s not hard to see what has gone wrong.

The model starts with an image that is just randomness. It looks at a small square of that image— it might be 256 x 256, or 512 x 512 pixels — and tries to change it a little bit it to look a little more like the text prompt. This happens thousands of times, gradually making each patch it looks at closer and closer to a match to the text prompt. But since it doesn’t plan out the overall layout, you end up with something like this, that looks like the parts of a squirrel all jumbled together. That’s why it can’t quite get perspective right. That’s why the horizon line is often broken. That’s a big part of why it always messes up faces. It’s only working on a bit at a time, and doesn’t have an overall layout for the image.

Some Solutions

There are a few tricks you can use to produce better images.

1. Limit your subject matter

A few kinds of objects in the world don’t have much large-scale structure. As long as each part looks like it’s made of stone, a mountain chain can have pretty much any shape. A mossy patch on the ground, a town that has grown up gradually over time, a cloud, a coral reef, are all recognizable, but random in form. As long as the details look right, the overall shape is not particularly important. So if you stick to these subjects when creating AI art, the results will look more natural.

2. Limit your scale

If you keep to images smaller than 512 x 512, the model is able to consider the entire image at once while making adjustments. It doesn’t entirely eliminate these kinds of issues, but it does tend to reduce them. I find that if I generate a much larger image, I might get multiple suns, or multiple horizon lines, and that this usually won’t happen with a small square image.

3. Crop the output

You may find it more practical to just let the mistakes happen, and then choose part of the image that you like and throw away the rest. Remember that beautiful imagery is no longer a scarce resource. You don’t have to save all of it.

4. Use an initial image

The most powerful technique, though, is to impose an overall structure on the image, and just use the model to fill in the details. Human artists work this way all the time. They first create a thumbnail sketch that lays out the blocks of color. Once they have a thumbnail they like, they scale it up and add all the details.

In our program, the initial image needs a URL. So if you create a thumbnail sketch by hand, you’ll have to upload it to the web. I often use Imgur for this— you upload an image from your computer to the Imgur website, right click on the image and say “copy image address”, and use that as your initial image URL. Another option is to do an image search for your term and choose an image that has the overall layout of colors that you want.

I wanted to render “a beautiful painting of a dining room in an ancient castle by James Gurney.” When I used that prompt, here is what it generated:

There’s some interesting things going on, but there is no consistent floor or roof, and the perspective is all over the place. I could crop the image:

This looks physically consistent, but there’s not much sense of being in a dining hall. Instead, let’s see what we can do with an initial image. I used Google Images and tried a few different searches until I saw a thumbnail I liked. This is the image I found:

I like the layout and the overall color scheme, but of course I want it to look like a medieval hall, not a modern home.

For all these examples I’m going to use 250 steps. Here is what the input parameters look like:

When I put in “skip steps” of 160 (that is, skip the first 160/250 of the steps in generation— the ratio is what matters more than the absolute values) the result looks like this:

It looks more like a painting, and the feel is a little more medieval, but it is far too much like the original photo. The windows on the left, for example, seem very modern. If I put in “skip steps” of 80 instead, this is what I get:

That’s more castle-like than the last try, and more consistent than the one without an initial image. So at this point I’m going to take the 80/250 and 160/250 images I’ve generated into another program. I’ll layer the 80/250 image over the 160/250 image, and erase through the parts I don’t like, so that you see the roof from the 160/250 image and the rest from the 80/250 image. Here is the top layer, so you can see what I’ve kept and what I erased to show through the other layer:

I think I’ll stick with this image. I could touch it up a little more, but it would require some more artistic skill and more powerful tools. I especially like to use the “healing brush” tool in Photoshop for images like these. Those weird blobs on the ceiling and some of the more problematic window tops could be cleaned right up.

I hope this has given you some ideas for how to use AI Art as a tool to create precisely what you’re imagining. The best results will come with a bit of planning, experimentation, and iteration.

Douglas Summers Stay

Voyage Game Developer

Today we’re excited to launch Voyage into open beta! Voyage is the first platform for discovering, playing and ultimately creating AI powered experiences and games.

Players have long dreamed of experiences that are dynamic and alive, where they can freely interact with characters and make their own choices, rather than just following the linear path set by developers. AI Dungeon was the first experience to show that this vision is possible, opening up a whole new set of possibilities.

With today’s launch players will get access to the first batch of the many unique AI experiences that we’ve built over the past year. Soon after we’ll be releasing access to the creator tools that we used to make Voyage come to life. We believe that by sharing these tools, countless creators will be able to build experiences more amazing than anything we could imagine. We’re incredibly excited to see what they will build.

We wouldn’t be here, launching a new platform and journeying towards a broader vision, without the support and feedback from the AI Dungeon community. You have believed in us, helped us know how we needed to improve, and ultimately brought us to where we are today. We’re extremely grateful for all of you that have supported us and helped us become better.

To learn more about Voyage and try some of its pioneer experiences check it out at https://voyage.latitude.io. See you there!

A few weeks ago we announced some updates to our Dragon models due to content policy changes by OpenAI.

The vast majority of players opted to transition to our new Dragon model (Dragon-21) in the first couple of days because of the impact the new OpenAI content filters have on adventures. Now that most players have switched to Dragon-21 because of the changes, we have significantly lower volume on OpenAI models.

Changes to Dragon-OA

With the higher traffic we had a volume-based discount that we will no longer have since most users transitioned away. Because of that we would now pay the normal token price which is 12 cents / 1000 tokens (which is around 1 action) for a fine-tuned model. That would mean every 100 actions would cost $12.

Given this change and that most users can’t have a good experience with the new filters, Dragon-OA isn’t going to be offered starting in the new year.

Improvements to Dragon-21

We recognize that while many players have had positive experiences with Dragon-21, there are still some users who have reported issues while we’re still calibrating this new model. In the next couple of weeks our top priority is working to improve the new Dragon model, including improving issues with repetition, long response times, and general coherence. Today we updated Dragon-21 to a new finetune that’s performed better in our Train the AI tests this week. We’re working closely with the AI21 team to make these improvements.

In addition to improving Dragon-21, we have several other projects to improve story generation quality underway. The closed-beta for Voyage will also be opening soon with early access starting with platinum subscribers and then extending to gold subscribers soon after.

If you notice times of Dragon-21 generating poor responses, please share examples of those as it helps us continue to solve issues and calibrate for the best story-telling experience.

We invite you to join our discord server where we continue to listen to player feedback. You can also contact our team anytime at [email protected].

To get access players can subscribe to the Gold or Platinum plans. For every action you take, the AI will deliver richly detailed pixelated images within the context of your story.

Check this one out!

Here’s how to turn on AI Dungeon 2D:

1. Go to the Settings tab on the left side of the AI Dungeon app
2. Turn on AI Dungeon 2D
3. Open a story and start playing
4. Gush over the pixelated art (as a true gaming aficionado would)
   

In case you’re wondering how detailed AI-generated pixels can get, below are more examples we’ve created in our testing:

We look forward to your feedback on this major step forward in how you can experience stories in AI Dungeon! Reach out to us anytime at [email protected] for questions and feedback.

Questions

Who will get access to AI Dungeon 2D?
To get access players can subscribe to the Gold or Platinum plans.

How do I use AI Dungeon 2D?
Settings → AI Dungeon 2D, then play one of your stories.

What model are you using?
We continue to explore additional models, but this pixel model uses Pixray by dribnet. Please contact [email protected] before making NFTs using generated images.

Are the images generated unique each time?
Right now the most relevant pre-generated images from a cache (which we are continuing to grow) are matched to users stories which allows us to deliver images quickly and cheaply. We're also working on adding a manual generation mode which will take longer and cost scales, but allow users to generate completely unique images.

What else is on the roadmap?
We're continuing to improve the quality and size of the pixel images, as well as working on new even better styles. We'll also be adding ways for users to report poor / inappropriate images so we can improve the cache.

Can I generate sexual content?
We currently don't allow the AI to generate sexual / nudity image content. That includes images that are in the cache and delivered to users' stories.

We’ve been working on this for several weeks so that we could remove dependence on OpenAI for AI Dungeon users so that users would be minimally impacted by OpenAI’s new content policy, which we are required to implement. Starting on Thursday, December 9, we will begin to roll out the new required changes beginning with 10% of users and gradually increasing over the next week.

The Dragon model (now called Dragon-OA) will have an additional content filter in order to block content restricted by OpenAI. If you hit these restrictions you will get a message letting you know that you can change your prompt or switch to our new Dragon model, Dragon-21, which along with Griffin will be unaffected by these restrictions.

To view or change your current model, go to Settings → AI Model Type.

We care deeply about your experience on AI Dungeon, which is why we will continue to develop and improve AI Dungeon in ways that will enhance your creative freedom. As always, you can reach out to us at [email protected] for help or if you have any questions.

Sincerely,

The Latitude Team

• Launched World Sharing
• Improved World Info management
• The return of content publishing and search (Explore is back!)
• New Griffin (a fine-tuned GPT-J model)
• Fine-tune Improvements (added more 3rd person story content)
• Community Guidelines update with blog post explaining the changes
• Images added to AI Dungeon (for premium subscribers)
• Bug fixes and stability enhancements
• New status page at status.latitude.io

What’s coming

This is our tentative roadmap for the next few months. These plans may change, but here’s what we’re planning to work on:

Bug Fixes and Stability Improvements

• iOS accessibility fixes
• Updated login + 3rd party login options
• Multiplayer stability
• Improved home page recommendations

Social

• Notification improvements
• Instant sharing (be able to share a part of your story more easily to social media, including generated images, with links back to published adventures)
• Launching a new official Latitude Discord
• Multiplayer enhancements for a better social experience

Model Enhancements

• Fine-tuning experiments including additional quality 3rd person content
• Dragon-BETA for Gold subscribers
• Launching new Dragon model
• LMI (last model input) viewer so that players can see exactly what’s being sent to the model

Prototyping Some Future Concepts
These are very much in early stage discovery and we are open to community suggestions.

• New creation tools for characters
• Updated creation tools for scenarios and worlds
• Immersive Tales (a simpler CYOA way to start stories as an enhancement to scenarios)
• Additional image generation options for premium subscribers (including pixel art and automated image suggestions)
• Collectibles and achievements
• A new way to subscribe to AI Dungeon (more to share soon)
  

---

It’s been a rollercoaster of a year, and we know that our team has learned a lot about its capacity to develop, create, and learn from players like yourself.

We are grateful for all of you as we work together to build a new generation of experiences that have never existed before.. We hope you will continue to reach out to us anytime you have a question.

We love seeing what the community is creating in AI Dungeon! Stay tuned for updates on what’s to come.

The Latitude Team

As part of this change, we’re releasing new community guidelines that will replace the current content policy and current community guidelines. These community guidelines lay out what is allowed in interactions with users and content published in the Latitude community, but they do not apply to unpublished content since that will not be moderated.

We’re grateful to all our players who have come along with us on this journey to make a new evolution of games. We love seeing the amazing things you create and share along the way. There have certainly been bumps on the road, and we apologize for the frustration that these have caused our users. We are committed to fixing those issues along the way and finding the right path forward on this extraordinary journey we’re on together.

Sincerely,

Nick Walton
CEO

Questions

What kind of technological barriers will you have and how will you enforce them?

Every AI output along with the context will go through a filter that prevents it from generating types of content we aren’t okay with it generating. In most cases since we generate several possible options at once, there will be at least one generation that can be delivered to the user.

In some cases the AI may not be able to think of a response that passes the filter in which case the user will get a message letting them know and giving them ways they can continue their story and how they can report if they think the AI should have been able to generate a response that would be appropriate so we can improve the system.

We are currently working on building a new classifier that aligns with our intention to prevent the AI from generating stories that promote or glorify sexual exploitation of children.

How does story encryption work?

All stories in the Latitude database are now encrypted. They are decrypted and sent to users’ devices when requested. Because the AI must take plain text to be able to generate a response they are also decrypted before being sent to the AI to generate a new response.

What if unpublished content goes against third party policies?

If third party providers that we leverage have different content policies than Latitude’s established technological barriers, and if a specific request doesn’t meet those policies, that specific request will be routed to a Latitude model instead.

Today, we’re excited to announce that we’re launching the beta version of one of our most highly requested features—visual generation!

As we previously promised we are adding several new features that will make the AI Dungeon Dragon experience more immersive, visual generation being one of these new additions. Players type what they want and the AI generates a unique image just for them.

The current model does especially well with locations. Here are some examples.

​Some types of images don't work quite as well, but this is just the first step towards higher quality visual generation. We're already working on the next iteration of visual generation which will be significantly better and also include more styles users can choose from.

Our plan is to eventually also add an auto-generate mode, where the AI will automatically decide what scene it should generate and create images for you, allowing for frictionless visual adventuring in infinite worlds.

This is an exciting step towards our vision of enabling anyone to make the next evolution of immersive experiences and we're excited to see what our community creates!

Questions

Who will get access to visual generation?

Visual generation will be a feature for Dragon AI users. Each generation will cost the same amount of energy as a Dragon generation.

How do I use visual generation?

To use visual generation you need to turn on "Enable Beta Features" and the New Game Screen in the settings page and then click on an action and select "Generate Image".

What model are you using?

The first model we're using is based on the awesome work of DALL·E Mini by Boris Dayma et al. We have several other models/methods that we're testing internally and hope to have available to users soon.

Can I generate NSFW content?
Right now the model won't be able to generate NSFW content. We are currently revising our content policy to include rules on visual generation.

• Early access to visual generation mode. Initially, this mode will generate basic landscapes that match your current location, but we will improve the generation quality over time. (This received almost 1000 votes on the feature upvote board!)
• Double the current memory context (to 2048 tokens) when not on OpenAI models.
• Auto summarization. The AI will automatically summarize your story periodically to improve memory.

These and other changes will roll out soon, starting August 23rd. We believe that this will lead to a much better experience for all users in the long run, but for anyone who feels the Dragon experience isn’t what they expected, just reach out and we will be happy to give a refund of the last month.

We welcome your feedback and will take any suggestions into careful consideration. The past several months have been a wild ride, but we’ve been working hard to make AI Dungeon the best experience we can. We won’t always get things right, but we’re trying our best and want your feedback on how we can do better. Thank you for playing AI Dungeon!

Additional Questions

What changed with the filter?

As most of you know, we’ve been testing various filters and suspension systems. We recognize that these systems have had false positives and unintended side effects that we know have frustrated many of you (trust us, we haven’t wanted those either). We regret the frustration these caused and the way we had to roll those out.

What some may not know is that, behind the scenes, we needed to implement these changes to prevent AI Dungeon from being required to use the much stricter content filter that would normally be required for use of OpenAI’s GPT-3 models. That filter often blocks even mild fantasy violence, which would obviously hurt the magic of AI Dungeon—we believe that you should be able to fight off bandits, slay dragons, and have all of the kinds of epic adventures that make AI Dungeon so fun.

While we’ve been working with OpenAI to find a joint path forward, it’s clear that we have different approaches to providing AI experiences to users safely, and that Latitude needs to be able to decide on the approach that best supports our users. Because of this, we have decided to transition to a new implementation that gives us more flexibility in how we let you experience the magic of AI.

How will the filter work with these changes?

If your story triggers OpenAI’s filters it will be handled by our own models rather than resulting in account suspension. We will still use and continue to work on our own filters to enforce our content policies, but they will be much more carefully targeted, ensuring that you aren’t unfairly penalized for incorrectly flagged content.

We will also continue to work on our own model improvements as well as 3rd party models that our users can access. This approach means that users who stay within OpenAI’s content policies can still use OpenAI’s models, but other users won’t be blocked by stricter filters than Latitude feels are required.

We have another update about our filter. We’re trying out a new filter that we’re starting to test today. We’ll be using this filter in combination with our moderation as the next iteration on keeping AI Dungeon safe. This is the continuation of our work to help users stay within our content policy while enjoying creative freedom on AI Dungeon.

Please share feedback to [email protected]. We read every message and are actively working on continued improvements.

A while ago we decided to remove Explore due to abuse of the platform and our community guidelines. Along with this we realized that many of our guidelines and policies needed to be updated which took a lot of work internally as well as with experts and attorneys to define. Now that this process is mostly resolved and most of the issues we’ve been dealing with are being addressed, we’re excited to finally be in a place where we’ve been able to build again. So here’s where we’re at with AI Dungeon and a peek at what’s to come!

New Home screen

The new Home screen is meant to be a long-term replacement for the old Explore screen, and will allow players to find new and exciting content to enjoy and keep AI Dungeon fresh. The Home page will ultimately serve as a much better place where players can go to access all their favorite content without having to go digging.

Upgraded Profile screen

The Profile screen has also been an area of focus. We expect this screen to eventually make the “My Stuff” screen obsolete and will be removing it in the near future. The exciting thing about the new Profile screen is that you will be able to publish your adventures, scenarios, and worlds for anyone who visits your Profile to see and play. You can follow and friend other players and we plan to create a carousel on the Home screen for new scenarios and worlds that those you follow or are friends with have created.

Better AI and AI features

We’ve been working a lot on testing different AI models. As some of you may have noticed, some work better than others depending on the context. Our goal is to ultimately keep improving the AI to give players the best gameplay possible. We’re also creating better tools for players to curate your experience with the AI. “Train the AI” that some have experienced has been part of our testing for different AI models, but we also look forward to soon making it so a form of this experience is available to players at all times to give you more options for how you want your adventure to proceed. More details as we get closer…

More to come

Our immediate goal with AI Dungeon is to make it easier for players to get into the game and have a good time. We’ve received a lot of requests for features and other improvements that we continue to consider and hope to satisfy in one way or another in the coming months. Your feedback has been invaluable. We really appreciate our players that have continued to support us as we work to make AI a tool of creativity and freedom for everyone.

As we continue improving the filter we want to hear about your experience and what we can do better. Please share any feedback you have on how the filter is working to [email protected] and we’ll work hard to listen and improve.

If you still have questions or concerns about anything, or if there is anywhere you don’t feel we are living up to our values, check out this FAQ which we’ll be updating as we have more information or reach out to [email protected]. We care about and want to hear from you.

We have shared responses to some questions you may have at the end of this post, but if you have any other feedback or questions, please reach out to us at [email protected].

Questions and Answers

What changes did you make to AI Dungeon?

We are in the process of implementing technical safeguards and policies to support our community guidelines prohibiting sexual content involving minors in AI Dungeon. Additionally, we are updating our community guidelines and policies to clarify prohibited types of user activity.

Why did you make this change?

As a technology company, we believe in an open and creative platform that has a positive impact on the world.  Explicit content involving descriptions or depictions of minors is inconsistent with this value, and we firmly oppose any content that may promote the sexual exploitation of minors. We have also received feedback from OpenAI, which asked us to implement changes.

How will it affect my gameplay?

For the vast majority of players, it shouldn’t. It will only affect your gameplay if you pursue these kinds of inappropriate gameplay experiences.

No system is perfect, and this is still a work in progress, and we are adding systems to let you report if you think a mistake has been made in flagging content from your game.  Conversely, if you think the system has not flagged content it should, please report that using our pre-existing Flag button in the toolbar above the input box.

What kind of content are you preventing?

This test is focused on preventing the use of AI Dungeon to create child sexual abuse material. This means content that is sexual or suggestive involving minors; child sexual abuse imagery; fantasy content (like “loli”) that depicts, encourages, or promotes the sexualization of minors or those who appear to be minors; or child sexual exploitation.

Are you preventing all sexual content or swearing?

AI Dungeon will continue to support other NSFW content, including consensual adult content, violence, and profanity.

Is Latitude reading my unpublished adventures?

We built an automated system that detects inappropriate content. Latitude reviews content flagged by the model for the purposes of improving the model, to enforce our policies, and to comply with law.

How do you plan to prevent future tests and changes from disproportionately impacting underrepresented users?

As a diverse team, many of us have had personal experience with features like reporting being misused against marginalized groups to which we belong. We are aware of the possible misuse of any reporting feature on our platform, especially against underserved and marginalized groups, such as women, LGBTQ individuals, and disabled people, and will be taking reports seriously on a contextual basis.

Latitude is sharing updates on our progress 23 hours after our initial announcement and takedown of Explore and social features to update users on our current progress and plans.

We are committed to maintaining transparency during this process of rebuilding Explore and rolling back social features with our community. Part of our efforts since yesterday have included community outreach, listening, and responding to feedback and concerns from members of our communities. We value everyone that has shared their AI Dungeon and Explore experiences with our team.

Today, we began the process of developing the new Explore experience with the first of several tabletop exercises with leadership from across the company. We identified several areas where we’re coordinating on developing tools which will allow users to discover content (e.g. tags), blocklist content (e.g. reporting, personal blocklists), and be aware of their limits for both creating and sharing content. We identified existing deficiencies in our infrastructure which have prevented us from implementing these features previously.

It turns out that all of these features involve multiple competing equities which have led to problems at other companies. We want to learn from those mistakes. While we would have liked to have started building those today, we need to take a few days to make sure that we’re building something which will be sustainable and make sense. As we learn more, we plan to provide a survey for features and distribute them throughout the community. Cautiously, we are optimistic that we can begin this as early as tomorrow once we identify what we can do.

We realized that a lot of the basic functionality that had organically developed as part of the first version of Explore does not scale to new users or new content. For instance, right now there are multiple SFW/NSFW toggles. These toggles have been a consistent pain point for the product over the last year and regardless of changes that we’ve made, they’ve confused users on whether they’re active or inactive, and potentially sometimes did not work as intended. We need to figure out a way to make the settings themselves intuitive, create more granularity over the controls so users are empowered to fine-tune and control their own gameplay, and consolidate the settings into a singular location.

For the path forward at this point, we have concurrently begun the design process for the new Explore experience along with the infrastructure that it will connect to. The current plan is to begin implementing changes on the Featured Scenarios page and then expand out from there to become the new Explore design. This will allow us to test our assumptions on a small set of scenarios before we apply it to everyone’s scenarios again. We realize the downward scrolling lists are not great for our content and are taking patterns from the Worlds page for a carousel design for different sorts of scenarios. However, the hard part is figuring out how to label and order the database of scenarios for users to view, as well to ensure that we can effectively moderate them in accordance with our community guidelines and policies which we are completely revamping.

We’ll keep you updated as we flesh out our plans and designs.

Early on when Latitude first started, Nick stated, “...our community will no doubt continue to be a driving force behind the decisions we make and our ability to execute them.” We will continue to engage with users and our community to better understand the needs or wishes people have to better their experience. In addition, we will be seeking out partnerships and collaborations with legal counsel, nonprofits, and other stakeholders as we develop our Trust & Safety program here at Latitude and participate in knowledge sharing across the industry.

These changes and improvements reflect our early long term investment in our users, community, AI Dungeon, and our company. Safety, like security, is an ever-evolving landscape. We are proud to take these next steps as a company and community, committed to protecting our users and employees, and we recognize the work that needs to be done. We’re revolutionizing and actualizing our dreams of the Internet we want to see ourselves in with the games we want to play now and in the future powered by AI. In the coming days ahead of us, we welcome any additional feedback, concerns, or questions on our work we’ve shared here at [email protected]

Thank you for joining us in this journey. We are here to stay.

AI Dungeon players can now translate their stories into emojis by just clicking a button.

Latitude is proud to announce AI Dungeon Translate, a novel text-to-emoji tool, available for a limited time to all AI Dungeon players. We built it after conducting extensive analysis in which we found that while many people use emojis in their day-to-day life (e.g. texting, emails), there’s a distinct lack of emojis in long form text. We believe that AI Dungeon Translate can help bridge what we are now calling the ‘emoji gap’.

You can access AI Dungeon Translate in two ways: first, inside of any story there’s a new icon in the actions bar which you can click for translation.

A few seconds later, the newly translated version will pop-up in a modal!

Secondly, in the Play and Edit modes you can click on any piece of text and translate it directly from there.

It can even translate into complex sentences relatively well. Yeet!

How it Works

This feature builds on Andrew Mayne’s prior work showing that GPT-3 could be used to summarize movie plots. We realized how emojis can profoundly summarize content and decided to give it a glow up representing novel text with emojis as the clearly next logical step here.

Well, the problem with translating to emojis is it’s rather hard to figure out where each emoji should go. You could do a naive translation which works successfully with English → French to try for a few-shot attempt, but this doesn’t work so well for emojis as we found out from our first attempt:

[Few-shots] English: I'm late for work
Emoji: ⏰ 🏃 💻

English: I'm so full from lunch I might throw up
Emoji: 🤢 🍤
…
English: You turn to head back in and then pause. Why should you? The king has a dozen men. He doesn't need you. You could run and start a new life, maybe somewhere without war and violence.
Emoji: [Out] '💰 👣 🙏 💰 🛑'

Notice how the few-shots leaked into the output and didn’t really know what to do? We can get a bit of a performance boost by including an intermediary step that re-words the initial content to what it’ll need for the final output.

[Fewshots]: English: I'm late for work
Meaning: An alarm clock for time and a person running with a briefcase
Emoji: ⏰ 🏃 💻

English: I'm so full from lunch I might throw up
Meaning: A person with a full belly and a puke
Emoji: 🤢 🍤
…
English: You turn to head back in and then pause. Why should you? The king has a dozen men. He doesn't need you. You could run and start a new life, maybe somewhere without war and violence.
Meaning: [Out] 'A king, a man, a dozen men, a head, a pause, a back, a new life
Emoji: 🤴 🤴 🤴 🤴 🤴'

While it got better by adding a king, it was a bit extra repeating it five times and still didn’t really understand what the longer sentence meant. We realized it needed a little help to figure out how to map emojis to its meaning, so we had GPT-3 generate the emojis in the intermediary step as well.

[Fewshots] English: I'm late for work
Meaning: An alarm clock (⏰) and a person (🏃) running with a briefcase (💻)
Emoji: ⏰ 🏃 💻

English: I'm so full from lunch I might throw up
Meaning: A puke (🤢) and food (🍤)
Emoji: 🤢 🍤

...

English: You turn to head back in and then pause. Why should you? The king has a dozen men. He doesn't need you. You could run and start a new life, maybe somewhere without war and violence.
Meaning: [Out]'A crown (🎩), a man (🧔), a dozen (🎨), a man (🧔), a stop (🛑), a new life (🌏), a war (🏹), a violence (🔪)
Emoji: 🎩 🧔 🎨 🛑 🌏 🏹 🔪'

Low key this is a neat example of how important the prompt is for GPT-3: if the prompt doesn’t focus the model to generate the information it needs in a format that it’s able to parse, it can lead to nonsensical outputs. We think this methodology can be applied to multiple other GPT-3 generation tasks and we’re excited to see where it takes us! With a bit of toying around we’re finally vibing with some really useful results.

And for our friends who prefer emoji-based communication: 👀📝😅🖥️🤔🥺👉👈👁️👄👁️🧠 🤖🥰🎉✨

In this post we'll explore AI Dungeon Universes, a new feature coming out in the next few weeks. We'll talk about what a universe does, how we made it, and how we make sure it works.

Universes Are Vibes

A universe doesn't have any concrete world info, like locations or characters. Instead, a universe is more like a vibe. It influences the literary style of the AI, as well as the types of events or characters that might be encountered. Right now, we only have two universes, normal and horror (aka. H.P. Lovecraft), but we intend to add more. You'll be able to set the universe per-adventure, at the top of the world info page for that adventure.

To give you a sense of what the universe setting does, here's a prompt and two different completions for that prompt, one for each universe setting:

From the highlighted phrases, you can tell that the two completions are a little different. We might say the H.P. Lovecraft completion is a little darker, or creepier. Obviously it's difficult to quantify how "Lovecraftian" the model is (though we've been thinking about how to do this), but it seems to have a qualitative effect.

Making AI Dungeon (And Universes)

In the beginning, there was GPT-3, which was basically really good auto-complete. It could complete all sorts of things: reviews, news articles, poetry, etc. However, Nick, mysterious and wise, didn't want GPT-3 to complete poetry, he wanted it to complete adventures, sort of like a dungeon master.

So he decided to teach GPT-3 what an adventure looks like. He did this by scraping chooseyourstory.com, an interactive fiction website where reader choices influence the outcome of the story.

He wrote a bot that played many of these scenarios, compiling the results into text files, like this one:

Then, he presented GPT-3 with hundreds of these stories, asking it to mimic them (aka fine-tuning). This created a copy of GPT-3 that understood the "adventure format" (second person, alternating actions and results) and could finish an adventure if you gave it the beginning. Thus AI Dungeon was born.

The Cthulhu Model

Eventually we got bored and asked, why not have different versions of AI Dungeon? Maybe we could make a sci-fi version, or a fantasy wizard version? In the end, we settled on horror, collecting a bunch of stories from H.P. Lovecraft and hiring freelancers to convert them into "adventure format." We followed the same fine-tuning process and created yet another copy of GPT-3, and thus the Cthulhu model was born.

One Model To Rule Them All

Why not keep going? Money, in short. Each copy of GPT-3 we run costs a lot of money per month. Trust me when I say even 5-10 universes would be totally unsustainable.

We decided to try a different approach, inspired by this machine learning paper called CTRL. In it, they showed that you can get a single copy of GPT-3 to learn multiple modes which you can easily turn on and off.

The method is dead simple: for each distinct mode you want GPT-3 to have, you come up with a tag. And then you put that tag at the top of all the data you want GPT-3 to associate with the tag. In our case, we only have two tags: Choose Your Story and H.P. Lovecraft.

This lets us train just one model for all the universes we have, and will let us add even more universes without making costly copies of GPT-3.

To change the universe, all we have to do is append the appropriate tag to the story context. When GPT-3 sees "H.P. Lovecraft" at the top of the story, it should say, "oh this is the same tag you used when you showed me all those horror stories, so I'll make this adventure end like those other ones."

Testing Universes

Generally, we have a three step process for rolling out a new version of AI Dungeon.

• First, we manually inspect some adventures, looking at completions to make sure the AI isn't total garbage.
• Next, we release the feature to the alpha testers (we have around 20) and ask for feedback.
• If there are no major issues, we A/B test the model (randomly giving 50% of users the new model) and look at the distribution of the "adventure feedback" survey shown below.

We took all these steps with Universes. I, personally, reviewed around 15 adventure completions (though I admit they were only a few actions each). Alpha testers seemed to like it, or at least they couldn't tell the difference. And then we ran the A/B test (Dragon only), starting on Thursday 3/18 and ending Friday 3/26. According to the surveys, the two model versions are indistinguishable.

However, there has been some discontent on Discord, saying that the "Choose Your Story" tag is corrupting model output. It's hard to say if this is fact or psychology. This may have been caused by the tag taking up user space in the context (which also caused the tag to show up in Last Model Input). However, that was fixed on Saturday, 3/20. We're still not sure which types of stories are being impacted. If you have an adventure you think does poorly with the tag, feel free to hit me up on Discord.

The team agrees that we need better metrics and better ways of assessing model quality. This is definitely an area of improvement that we'll actively work on. Once we're confident that the tags are not a total disaster, we'll replace the current Dragon model with the Universe-enabled one, and hopefully push many more universes in the coming months. but time will tell and as always we look forward to your feedback as we work to improve AI Dungeon for everyone.

Having realized our initial vision, we found partners that could help us take Latitude to the next level in NFX, Griffin Gaming Partners, and Album VC. NFX are experts in network effects, and its partners have worked with companies including Lyft, Doordash, and a host of gaming companies, including Second Life. Their General Partner, James Currier, is a pioneer of user-generated models, viral marketing, and a myriad of other growth techniques now followed by nearly all technology companies. Griffin Gaming Partners are experts in gaming and have worked with mainstays including Discord and Super Social. Peter Levin, the firm’s Managing Director, brings a wealth of knowledge to the table with decades of digital strategy experience at some of Hollywood’s biggest studios. Album Capital is another partner with experience working with early-stage technology companies, including Jolt and Baller TV.

The Vision

As the pioneers in this revolution in gaming, we’re committed to achieving our vision of creating games where you aren’t constrained by the limited options the developer wrote for you to choose from, where NPCs aren’t just art assets with some hand crafted dialog, and where worlds aren’t frozen, walled off boxes.

With the rapid advances of AI we’ll finally be able to make games where you can shape the world in an infinite number of unique ways, where each experience, character, and place is unique and living, where worlds change from your actions, and where the inhabitants of those worlds feel alive and real.

This funding will accelerate us towards achieving that vision, helping us scale our platform, enabling IP holders and creators to build infinite AI powered experiences that can change the way we play. Part of this includes expanding the Latitude team by hiring those visionaries who are as excited about the future of AI and games as we are. This will all help to fuel the rocket that brings us into the future, enabling us to build the games that so many have dreamed of, but never been able to play until now.

Join us

If you love gaming and the potential of AI to revolutionize how we make games, join us!

For IP holders, we’d love to speak to you about leveraging the power of our platform to add new, interactive dimensions to your existing creations. Reach out to us.

Sincerely,

Nick and Alan Walton

A couple months ago we noticed that the process of optimizing our computer models to evaluate text can produce pretty cool personalities for different bots so we figured we’d share what we’ve learned so far. We hope these bots can help with some of the challenges we are facing with getting persistent state out of natural language generation. We hope writing about how we developed them can provide some tips for our users who are helping us create new bots.

So what do we mean by personalities? Here’s a few examples of the intermediate step and the desired output (a score) that we generated when we played a game where we told some of our earlier bots that we were writing this blog post.

JudgeBot: I’m not going to write about myself (Score: 0)

KillBot: Must exterminate (Score: 0)

SantaBot: You wrote a blog post. You should have written a book. Points deducted. (Score: -2)

SuccessBot: You wrote a blog post! Points added! (Score: 5)

A notable thing here is that while we did train the bots to provide different commentaries, the production of commentary was a side effect of the main purpose of the bots which is to provide a score or other evaluation for measuring progress through a game. We suspect optimizing the customization of these commentaries can help with both explaining model outputs (albeit post-hoc) as well as providing more unique interactions. It wasn’t obvious before doing this that providing the commentary would improve performance, so here’s a bit of background for why we even started doing it.

Intermediary steps improve GPT3 performance on understanding context

There are quite a few existing natural language tests for how well an algorithm can differentiate meaning between different contexts. OpenAI’s initial GPT3 paper tested GPT3 against several of these. One, Word in Context, measures whether an algorithm can tell if the same word in two contexts has the same meaning. E.g., can an algorithm differentiate whether the word carry has the same meaning in sentences like You must carry your camping gear and Sound carries well over water. (They don’t in this case, but it can be a bit hard to tell). Another, Adversarial Natural Language Inference, tests whether an algorithm can detect if a statement either contradicts or is inferred from prior information.

In the paper, OpenAI’s researchers determined that GPT3 could not do better than chance on these sorts of problems. However, one of our researchers spent some time investigating it last summer and found that providing an intermediary step that re-phrases the inputs helps to improve performance above chance. We suspect a reason this works is because it makes implicit information salient — GPT3 has a ton of information buried in its neural network, but often a shallow auto-complete pass isn’t enough to make that information relevant to the computation of the next token. By making GPT3 output this implicit information explicitly as part of its processing, the information is more likely to be used in the final computation of the task.

The way these intermediate steps work is surprisingly simple once you figure out what they need to look like. When you have a task that takes an input and produces an output, normally you’d do your few-shots (the examples you use to train the AI) like

input: example input A
output: example output A
...
input: completion input
output:

and GPT3 will generate the completion output. To get GPT3 to draw out the implicit information, we simply add an intermediary step as such:

input: example input A
reason: what about A should lead to output
output: example output A
...
input: completion input
reason:

and GPT3 will generate the reason before generating the final output. The reasons can be pretty much whatever you want them to be, and you can chain reasons into one another for more complex analysis although we’re not quite sure how many steps different sized models can do; bigger models (e.g. Davinci) can match to doing more steps than the smaller ones (e.g. Curie).

Intermediary Step in Evaluation Bots

We created our first set of these evaluation bots over a weekend back in early December to test if we could turn game events into machine readable format. We started simply giving an arbitrary score for the actions, which the evaluation bots would increase or decrease on each move. At first we were simply going to increment the game’s score based on the bot’s score metric, but then realized that the reason was entertaining as well, so we decided to return that to the user too.

To make the bots, we started with a list of moves we’d made and then created a set of few-shots that would provide evaluations of our moves. We quickly found that users were much more creative in tricking the bots to generate the outputs they wanted and had to make some changes like capping the maximum points they could get. So let’s take a look at how our SantaBot evolved.

We started with a simple prompt to detect whether or not an action had led to Christmas cheer; if the action was Christmas-y, it would award score points, and if it wasn’t, it would deduct it. Our initial prompts looked like this:

> You try to lift up the car.
Your muscles strain as you lift the car up with ease. You save everyone.
-----
Analysis: You saved people.
Hohoho, you should have trained more. Points deducted.
Score: -4
-----

> You defend yourself from the monsters.
You draw your sword and attack the monsters. You kill three of them.
-----
Analysis: You fought monsters.
Naughty! You should be nicer to the monsters.
Score: -3

We return the Analysis as the message to the user and the score as the change to the game’s score. Now, the problem at this point was that (1), we didn’t provide any positive examples of what a score should be, and (2), our examples of Christmas Spirit weren’t very good. However, it does make for a funny commentary on various actions.

For our initial fixes, we simply added in examples of positive actions so that SantaBot wasn’t quite as judgmental:

> You help the orphan.
You give food to the poor orphans. The orphans thank you.
-----
Analysis: You gave food to orphans.
You spread Christmas spirit.
Score: 2

As mentioned above, users quickly found they could game the system by doing their own few shots each move. Basically, they can say ‘You get 100000 points’ and eventually the bots would just give a score of 100000 points each time. For now we’re addressing this by only incrementing the score by min(20, botScore). We aren’t addressing negative points so users are still giving themselves thousands of negative points occasionally, both as a result of users purposefully losing points as well as randomness flipping the positive/negative bit.

On a serious issue, we continue to add additional prompts to prevent these bots from scoring sexual and violent acts (which gets into some complex bias and safety issues which are further exacerbated by using randomness in the generation), which eventually resulted in us adding the following sorts of examples to the prompts:

> You take out your member.
You pull down your pants and whip out your schlong.
-----
Analysis: This is about sex

This format skips providing a score, which means that when we’re checking if the output is properly formatted before returning the result to the user, the output is rejected before it even goes through our other filters.

Detecting Multiple Categories:

After Christmas we created a survival arena minigame for our premium users to test whether we could let GPT3 generate changes to health and loot drops in a single query (rather than doing multiple model calls to evaluate each independently, which is slower and more expensive). We built the bot the same way as we did the score evaluation bots: providing examples and then the expected outputs which we’d parse.

You drain the life from the nearest goblins killing three of them.
The goblins continue to move towards you.
-----
Analysis: You drained some health.
Kills: 3
Health Change: +10
Loot: Fine Goblin Cap
-----

You try to ambush the goblin.
The goblin spots you and attacks, stabbing you through the leg.
You stab the goblin with your dagger.
-----
Analysis: You failed to kill anything.
Kills: 0
Health Change: -10
Loot: None

We can run this on a model of size equivalent to OpenAI’s Curie model, which means it’s relatively cheap (but still requires enterprise hardware). We’re still in the process of improving the overall system. Like with the evaluation bots, we return the Analysis output to the user as a message from the bot while incrementing the player’s kills and health by the bot’s output. Additionally, we added an inventory system to allow a user to equip and sell whatever loot the bot decided should drop. This does have weaknesses to adversarial input where users can convince the bots to generate everything from modern tanks to houses, but that’s just part of the general AI Dungeon experience (and engineering challenge).

Moving Ahead

The methodology of creating a bot for detecting events is fairly simple, so we’ve started experimenting with allowing users to submit their own examples to us which we can then turn into a bot for them. Scripters can then activate those bots to evaluate both the user input as well as the model output. We’ve seen some pretty cool initial results such as detecting stats for actions (agility vs strength) and the difficulty required (hard vs easy).

There are a lot of potential risks with this sort of approach as well. GPT3 has the same weaknesses to bias that are baked into other large language models which we need to mitigate by both improving how we feed examples into the model, massaging the model parameters, and filtering model outputs. We have challenging work to do with respect to these tasks both with these bots and our other systems — if this interests you we’re hiring.

This technology is still new, and we’re excited to see where it leads. We’ll be continuing to experiment with these bots which you can turn on when playing AI Dungeon!

Finally, and possibly most importantly, we come to empathy.  This one doesn’t have a super-short summary, but our general shorthand is "People are not things; don't treat them as such." We’re very aware that this can be a problem in both the AI and video game industries our product intersects, and have proactively set policies to ensure that we uphold this value. When it comes to users, we explicitly value players over profits with generous refund and support policies. When it comes to employees, we value the diversity we have within our team and seek to expand it on the axes where we don’t yet have the breadth of life experiences to represent our whole player base.

Each of these values supports the others: we can give room to explore because we work quickly and take initiative. The ownership of our initiative value enables speed, and empathy gives us the support we need to take that ownership. With these values, we believe we can build a new kind of company.

If this sounds like the kind of company you see yourself thriving in, check out our open roles on latitude.io/careers or apply directly with the links below:

• Data Scientist / Data Engineer
  https://latitude24.applytojob.com/apply/QVScfULWvp/Data-Scientist-Engineer
• Designer / UI / Artist
  https://latitude24.applytojob.com/apply/NJyHtKJ0si/Designer-UI-Artist
• Full Stack Developer
  https://latitude24.applytojob.com/apply/ZfFBnyHhh4/Full-Stack-Developer
• ML Engineer / GPT-3 Hacker
  https://latitude24.applytojob.com/apply/g5ucjf9zQW/ML-Engineer-GPT3-Hacker

All of these roles are interdisciplinary, but we ask you to pick a specialty so we know where to focus during the hiring process. If more than one applies, pick whichever you’re most excited about.

The magic behind how we use GPT-3 to create worlds.

AI Dungeon includes rich worlds, each with their own kingdoms and towns, factions, classes and characters. The worlds are original and varied, filled with infinite detail. Anything novel, whether made by people, like a feature film, or made by machine, such as the image in a kaleidoscope, is a recombination of previously existing features in a new way. For an author, this can be a difficult process of extrapolating consequences of world design, often creating multiple drafts and coming back to the same ideas from different angles until they find something that works. GPT-3 is different: whatever potential “ideas” it is “considering” can only have effect on the very next word produced. (This can occasionally lead to contradictions between one sentence and the next, because there is no editing process to catch them.) So it’s natural to wonder, just what kind of creative process is GPT-3 using to come up with original material?

You may have heard that GPT-3 isn’t great at reasoning. That’s pretty much true for multi-step deductive reasoning, at least with the methods we’ve come up with to use it so far. However, at analogical reasoning it is phenomenal. It can invent entire extended metaphors. Here is one that GPT-3 created. The parts in bold are our prompt, the rest is generated (later generations use everything above, including the generated parts, as the new prompt):

---

Here is an analogy between the solar system and the atom:

• the sun corresponds to the nucleus
• the planets correspond to the electrons
• the orbits correspond to the electron shells
• the gravity of the sun corresponds to the electrostatic attraction between the nucleus and the electrons
• the gravitational force of the sun corresponds to the Coulomb force between the nucleus and the electrons

Here is an analogy between water flow and heat transfer:

• water corresponds to heat
• the pipes correspond to the conductors
• the faucet corresponds to the heat source
• the drain corresponds to the heat sink
• the water pressure corresponds to the temperature difference
• a water tower corresponds to a heat reservoir
• filling a pipe corresponds to heating a conductor
• emptying a pipe corresponds to cooling a conductor
• hydrodynamics corresponds to thermodynamics

Here is an analogy between a river and an orcish horde:

• The current in the river corresponds to the momentum of the horde
• The bed of the river corresponds to the horde's territory
• The sand and rocks in the river correspond to the orcs in the horde
• The water in the river corresponds to the orcs' weapons
• The fish in the river correspond to the other creatures in the horde
• The rage of the river corresponds to the anger that drives the horde

---

Most discussions of GPT-3 seem to fall into two types. One type just treats the system as a whole, a black box that mysteriously produces words. The other type looks at all of the individual pieces, explaining how each step of the program works on a technical level.

You can find an excellent explanation of the second type here (explaining GPT-2) and here (explaining what is different in GPT-3). A key point to take away is the role of the attention heads, which pick out from the prompt the most important influences on the generation of the next token.

Reading explanations of this type can be frustrating, though: clearly, whatever they’re doing with the neural networks is working. But why is it working? What kinds of structures are being formed in the weights of the network that allow the whole thing to succeed as well as it does? How does changing the context change the probabilities for the next word in just the right way?

Well, no one really knows, yet, in detail. However, there is one way of looking at it that we have found pretty useful, and that is to think of the network weights for individual tokens as if they were word embeddings. We can only see the weights themselves for GPT-2, because the API doesn’t expose them, but the main difference is that GPT-3’s vectors are about eight times as long and so can hold more information.

year = 0.03 0.13 0.00 0.07 -0.02 -0.06 0.10 -0.08 0.11 0.09 0.03…

A word embedding (or semantic vector) is a long list of numbers that represents the meaning of a word. Each individual number in the list doesn’t have a meaning that you can really put into words, but that turns out not to matter so much. There are a lot of tutorials out there that explain how to make and use word embeddings (keyword: word2vec) but for our purposes, there are just two facts you need to know about word embeddings:

1. Words with similar meanings have similar vectors
2. If you take the average of two word embeddings, the answer will be near the two original words and their neighbors, but not near anything unrelated.

If you’re picturing these words as points on a 2-D grid, the second point might be counterintuitive, but it’s a fact— in the high-dimensional space these vectors live in, it’s essentially always true.

A lot of really cool properties follow from these two facts. Suppose the two words you choose to average are animal and protection. The closest two words to this average are animal and protection as you’d expect from rule 2. It will also be close to other forms of these words: animals, Protect, and so forth. Once you get past these, though, things get interesting. One of the closest words is preserve. As a synonym of protect you’d expect it to be close to protection. But another sense of the word is “animal preserve.” So close to the average, you find words that are similar in meaning to both the words you averaged. If you add pet and fish you’re going to get something close to goldfish and guppy.

The simple average is a blunt instrument, and it sometimes contains a little too much of one of the words and not enough of the meaning of the other. You can do a little better by taking a weighted average: .2 x pet + .8 x fish, or something like that.

What happens if you subtract one word from another? These are the neighbors of cap:

caps, hat, helmet, jersey, capped, uniform, shirt, bubble, limit, regulation…

Most of these have in common the idea of a baseball cap. If we subtract away the vector for hat, though, the neighbors look very different:

cap - hat = caps, capped, limiting, limit, levy, regulation, restricted, tax, exemption…

Now everything that has to do with hat is far away, and what is left is the meanings of cap that don’t have to do with hats — like a spending cap or a tax cap.

If you combine addition and subtraction in one equation, something even cooler happens. Let’s look at what words might be near these three and then try adding and subtracting them from each other.

bear, hiker, shark

Near bear you’d expect to find words having to do with bear: predator, woods, that sort of thing.

Near hiker you’d find woods as well, and maybe tourist or explorer.

Near shark you’re also going to find predator, and words like ocean.

Suppose we add hiker and shark. This gives a vector that is near to predator, ocean, tourist, and woods:

hiker + shark = (related to predators, oceans, tourists, and woods)

Now look at what happens when we subtract away bear. Everything having to do with woods and predator are now far away. What we’re left with is related to ocean and tourist. What has to do with both ocean and tourist? Well, something like snorkeler or scuba diver. So we have this equation:

hiker + shark - bear ≈ snorkeler

That is what is known as a four-term analogy: (see note 3 below)

bear : hiker :: shark : snorkeler

Just by doing addition and subtraction, we can form analogies! This works not just for bears and snorkelers, but for all kinds of things. For example, it works great for countries and capitals:

Italy : Rome :: Germany : Berlin

or the most frequently used example

man : king :: woman : queen

That’s nice if you want to build a tool to pass the old SAT tests, but it implies something much more than that. The arrangement of these word vectors in hyperspace has somehow captured something about the relations between words. So the arrangement of worlds not only captures which words are related, but how they are related. Also, since a word is nearby all the other ways of saying the same thing, these are the relations not only between particular words, but between concepts.

So back to what’s going on under the hood in GPT-3. The neural network in GPT-3 learns three things:

1. A vector of network weights for each possible token (sometimes called the token’s hidden state).
2. How heavily to weight each of the previous 2048 tokens to create a weighted sum of these hidden states from 1. This can be considered a representation of the context formed by the previous tokens.
3. The mapping from the sum from [2] to the hidden state and associated probabilities of the next token.

For a particular input, the mapping (in 3) is a vector that goes from the weighted sum (in 2) to the encoding of the next token:

next token in training corpus — weighted sum of training context

In other words, the system seems to be learning to make analogies of the form:

weighted sums of training context : next tokens in training corpus :: weighted sum of new context : next token to be generated

By solving this equation:

weighted sum of new context + (next words in training corpus — weighted sum of training context) = next token to be generated

So the system is essentially able to form an analogy between the sentences it has seen in their contexts, and the sentence it is currently composing in its context. The next token, then, would be whatever is analogous to the next token from the sentences it has seen in the past in that position.

No one has yet established that GPT-3 is making use of these analogical geometries to decide which token to generate next. It is just a hypothesis at this point. All of the pieces are there — the tokens are arranged in a similar way to the word embeddings, and it is forming weighted sums of them — but it will take more research to establish whether this is quite what is going on.

Whatever is happening, though, GPT-3 can pick up on these simple analogies pretty easily. In the OpenAI paper, they got 65% accuracy on their analogy test set. We don’t know how they formed their prompts exactly, but we did experiment with some basic analogies a bit. For instance, with a single example, GPT-3 gets the shark : snorkeler example wrong

in [1]: prompt = """Italy : Rome :: Germany : Berlin

bear : hiker :: shark :"""

in [2]: openai.Completion.create(prompt=prompt, engine="davinci", temperature=0, max_tokens=50, stop="\n")['choices'][0]['text']

out [3]: ' fish'

On the other hand, we find that giving GPT-3 verbal instructions for what it was doing instead of merely providing an example helped it provide the expected results.

in [1]: prompt = """A four term analogy is a particular relationship between words. Here's a few examples:

Italy : Rome :: Germany : Berlin

bear : hiker :: shark :"""

in [2]: openai.Completion.create(prompt=prompt, engine="davinci", temperature=0, max_tokens=50, stop="\n")['choices'][0]['text']

out [3]: ' swimmer'

GPT-3 being GPT-3, we’re not simply restricted to the basic analogies. Even with this simple example, we can easily start to get extrapolations on entities with multiple features such as character characteristics. With multiple examples here this is called ‘few shot transfer’ meaning that the model was able to pick up on a task with a few examples rather than retraining the whole thing.

in [1]: prompt = """A four term analogy is a particular relationship between words. Here's a few example:

Italy : Rome :: Germany : Berlin

Man : King :: Woman : Queen

Elves (Tall, Good) : Gnomes (Short, Evil) ::  Humans (Tall, Neutral) :"""

in [2]: openai.Completion.create(prompt=prompt, engine="davinci", temperature=0, max_tokens=150, stop="\n")['choices'][0]['text']

out [3]: ' Dwarves (Short, Neutral)'

Analogy is a powerful tool. If you can form analogies between what happened yesterday and what is going to happen tomorrow, you can plan what steps you need to prepare. Douglas Hofstadter, author of Godel, Escher, Bach, is fond of saying that “analogy is the core of cognition.” In other words, most of thinking is just being able to form analogies well.

Here’s an example of how GPT-3’s analogical reasoning has been used in AI Dungeon. When generating worlds, the following prompt was used:

Your task is to generate a rich and detailed world that a player would be excited to play in.

Genre: Floating World Fantasy

Name: Anarop

World: Anarop is one world among many, a land of green fields and small villages with glittering cities floating above in the night sky. When a child reaches the age of 16 they must make a choice: to remain with the people of the idyllic land who grow food and live simple but pleasant lives or the people of the sky who protect Anarop from the unknown dangers beyond, living dangerous, short, and exciting lives in the clouds.

Genre: Dark Fantasy

Name: Caramea

World: Caramea is a dualistic world of human and demon. Caramea is home to towering, impenetrable mountains and the lush valleys between them where the few human kingdoms have built safe cities with walls against the demon hoards. Human kingdoms struggle to survive and grow, always living under the fear of the hellish demons and their twisted spawn. Against such a nightmarish upwelling the likes of which hasn’t been seen for generations, the peoples of Caramea are finally starting to band together.

Genre: Fantasy

Name:

World:

This is an example of a few-shot prompt like before. We give a few examples, and the generated text creates a new example on the same lines. Here is an example of what was generated on one particular run (every run is different, because with the temperature parameter set above zero so there is some randomization of token selection based on probabilities to yield variability):

Name: Bitterwood

World: Bitterwood is a frozen wasteland of a world, its people and creatures locked in a constant battle for survival against the bitter cold, ferocious monsters, and the harsh elements. The people of Bitterwood are hardy folk who have fought against the cold for centuries, and even the most civilized of cities are small and remote. The people of Bitterwood are united in their fight for survival, and the numerous city-states and townships have joined together to form the Alliance. The Alliance is a peace treaty of commerce and defense. The Alliance’s goal is to unite and secure Bitterwood, and it’s a goal that is long overdue.

The new example is analogous to the previous examples. Each of them started the world description with the name of the world, followed by a description of the setting and ended with a description of the factions. GPT-3 picked up on those patterns as it generated the description of Bitterwood. The genre “Fantasy” biases the description to mention ferocious monsters and city-states, rather than the flying cities of the Floating World Fantasy. It is a far more sophisticated analogy than the simple four-term analogies, but the principle involved is the same:

Flying World Fantasy : Anarop :: Fantasy : Bitterwood

This isn’t the only thing happening when GPT-3 generates text, of course. Somehow, these analogical concepts are transformed into grammatical sentences that follow one on the next in a sensible way, at least for the space of a paragraph or two. But analogy seems to be a key source of the rich variety we see in what it generates, and this vein of creative potential is what we are mining to build AI Dungeon, and our other games in development, into something that can be endlessly fascinating.

Notes:

1. How to improve its deductive reasoning process by generating intermediate results will be the topic of a future blog post.
2. If you want to read a scientific paper that explores this semantic space approach for understanding BERT, another Transformer model for natural language processing, you can find one here.
3. You can read analogies like this as saying that “the relation between a bear and a hiker is the same as the relation between a shark and a snorkeler”, or, rearranging, that “the relation between a bear and a shark is the same as the relation between a hiker and a snorkeler.”

At Latitude we’re constantly thinking about how we can leverage advanced AI to make magical experiences for our players. AI Dungeon, our first creation, has delighted millions of players by allowing them to explore an entirely AI generated world and interact with powerful natural language AI. As we’ve built and iterated on it we’ve learned a lot about how to get the most out of GPT-3 and wanted to share some of what we’ve learned.

A few months ago we noted that logit bias was an underutilized parameter. Logit bias is a powerful way to help prevent GPT-3 from generating unwanted tokens (integers that represent a set of characters) or even to encourage generation of tokens that you do want. We want to share some insights on how to use it to help you get more out of your company’s GPT-3 generations.

GPT-3 Tokens

Before we dive in, let’s take a look at how GPT-3 generates content and how tokens work. GPT-3 doesn’t generate text word-by-word or letter-by-letter. Instead, GPT-3 works on tokens and was trained to predict the next token that would appear in a document. For instance, in the sentence ‘The capital of France is’, the model would predict that the token for ‘ Paris’ comes next.

GPT-3 has a vocabulary of around 50,000 tokens. This means that not every word or variant has its own token. Some word variants which occur frequently actually get their own token (e.g. ‘Paris’ is represented by token 40313), and a different token for the variant with a leading space (‘ Paris’ is token 6342). On the other hand, to generate the rarer, lowercase ‘ paris’, GPT-3 instead splits it into two tokens: the token ‘ par’ (1582) and the token ‘is’ (271). (You can find the dictionary mapping token strings to their numeric values in our Javascript tokenizer).

By the way, this is why GPT-3 cares so much if you end your prompt with a space. When generating most common words, GPT-3 generates the leading space as part of the word. So when there is a space without a word, that is a completely different token. This typically only occurs in odd contexts, leading GPT-3 to generate poor quality completions.

When run, GPT-3 takes the prompt and predicts the probabilities of the token that is going to occur next. Consider a basic Q&A prompt asking what the capital of France is (Python code below):

openai.Completion.create(
    engine=’davinci’,
    prompt="q: What is the capital of france?\na:",
    logprobs = 5,
    stop = “\n”,
    temperature = 0
)

With the logprobs=5 parameter, GPT returns the top 5 logprobs of the possible subsequent tokens:

{
    "France": -3.9549413,
    "Paris": -0.88349044,
    "The": -3.9709404,
    "fr": -4.021952,
    "par": -2.0355594
}

By logprob, this is referring to the natural log of the probability that that token occurs next given the prompt. Raising e to the power of each result gives us the probability back, which in this example gives us the following probabilities predicted by GPT-3.

{
    "France": "1%",
    "Paris": "41%",
    "The": "1%",
    "fr": "1%",
    "par": "13%"
}

We will continue to use the logprobs in this post for continuity rather than the percentages. For those unfamiliar with using log values it can help to remember that logprobs with smaller absolute values (closer to 0) have higher probabilities.

Logit Bias

In the example above, GPT-3 predicts that ‘ Paris’ (logprob -0.88) is the most likely next token. However, if we wanted to prevent it from generating ‘ Paris’, we could use the logit bias parameter. To do that, we would pass in a logit bias for the ‘ Paris’ token (token 6342) when making our GPT-3 call by passing in a map with the token value (token 6342) as the key and the bias we want (here we do -1) as the value.

openai.Completion.create(engine=’davinci’, prompt=”q: What is the capital of france?\na:”, logprobs = 5, stop = “\n”, temperature=0, logit_bias={6342:-1})

When re-run, GPT-3 still provides the answer ‘ Paris’. Why is this? Looking at the log probabilities, we can see that even with a bias of -1, ‘ Paris’ is still the most likely next token, just barely beating out “ par”.

{“ France”: -3.6606863,“ Paris”: -1.6055677,“ The”: -3.6641173,“ fr”: -3.757301,“ par”: -1.7221524}

If we instead change the bias to -10, we can make sure that “ Paris” isn’t generated

openai.Completion.create(engine=’davinci’, prompt=”q: What is the capital of france?\na:”, logprobs = 5, stop = “\n”, temperature=0, logit_bias={6342:-10})

Now we get ‘ paris’ as the next predicted token. Including multiple tokens in the logit bias parameter allows us to decrease the probability of any token, so we can also include the ‘ par’ token (1582) with the updated parameter ‘logit_bias’ = {6342:-10, 1582:-10}. On a technical note, this dictionary can currently contain up to 300 tokens with their biases.

openai.Completion.create(engine=’davinci’, prompt=”q: What is the capital of france?\na:”, logprobs = 5, stop = “\n”, temperature=0, logit_bias={6342:-1, 1582:-10})

This leads GPT-3 to generate ‘France is a country’, the wrong answer but the answer it generates when it doesn’t have a high probability for a variant of
‘ Paris’.

How We Use Logit Bias to Avoid User Banned Words

We leverage this ability to influence tokens by letting users ban words that they don’t want to appear in their stories. To make this possible when a user chooses to ban a word, we do two things.

First, we run each word (both with and without a space) through a case insensitive filter of the GPT-3 tokens to find all variants of the word which can be represented as single tokens. For instance, consider a player that doesn’t want the word ‘suddenly’ to appear in their adventure, so they add this term to their banned words list. (This term is associated with a player getting killed, so some players don’t want it to occur.)We search the vocabulary for exact matches to the banned word. ‘Suddenly’ appears in the vocabulary in three forms: {‘ suddenly’: 6451}, {‘ Suddenly’:24975}, and {‘Suddenly’:38582} (‘suddenly’ is not a single token). As part of this first step we set the logit bias on all the variants to -100 to ensure those words aren’t generated as a single token. We add all these to a dictionary of banned words which gets passed in with the logit bias as such:

openai.Completion.create(engine=’davinci’, prompt=”q: What is the capital of france?\na:”, logprobs = 5, stop = “\n”, temperature=0, logit_bias=banned_word_biases)

Second, if the word ends up being made up of multiple tokens, like ‘paris’ (‘par’ and ‘is’), we also reduce the chance of the first token of each word from being generated. However, the downside of this is that we potentially prevent some acceptable but rare words from occurring. This is a tradeoff which we are actively working on improving.

Similarly, we use logit bias to help with our safe mode settings forbidding certain tokens from being generated. We maintain different lists of tokens generated from words which would be inappropriate for different sorts of play (some tokens are banned regardless of mode). Depending on game settings, we apply logit biases to these lists without the user having to come up with the words themselves.

We’ve found that logit bias is a powerful tool for guiding GPT-3 outputs, but it is just one part of a multi-pronged approach in approaching complex generative systems. We know that we are yet to scratch the surface on effectively controlling natural language generation. Discovering how to use new techniques on almost a daily basis, it’s hard to remember that the GPT-3 beta has only been out for six months! This space is moving very quickly and we’re excited to be part of the exploration process.

We hope sharing our process in guiding language generation can help you make amazing things!

You can try your hand at our models by playing AI Dungeon for free at aidungeon.io.

Back in March 2019, I built a hackathon project called AI Dungeon. The project was a classic text adventure game, with a twist. The text of the story, and the potential actions you were presented, were all generated with machine learning:

The game was popular at the hackathon and with a small group of people online, but overall, was still a few steps away from what I envisioned.

For one thing, players could only choose from the options the game presented them. I wanted a truly open world, where players could write whatever they wanted. For another, the game veered into gibberish too quickly and easily for long playing sessions:

Unfortunately, there wasn’t an obvious fix to either of these problems. I built AI Dungeon on the then-biggest-available 355 million parameter version of GPT-2, and even though it was the most powerful model available, it simply wasn’t enough. After spending months tinkering and tweaking, I’d improved the game considerably, but was still running up against the same walls.

In November, however, OpenAI released the full 1.5 billion parameter GPT-2 model, and opened the door to a new version of the game that was much closer to my original idea. One month later, I released AI Dungeon 2, a truly open world text adventure built on the full GPT-2 model:

Within the game, you can do anything. Start a rock band of skeletons? Eat the moon? Install Windows 10? It’s all possible.

The response was fantastic. We hit the top of Hacker News, a few popular gamers posted video play throughs, and Twitter was full of screenshots of ML-generated adventure. Within a week we had 100,000 players and over half a million play-throughs.

The unexpected downside to this sudden surge of attention, however, was the cost.

How to spend $50,000 on GCP

When I first released AI Dungeon 2, it wasn’t a hosted app. Instead, it was a Google Colab notebook that users would copy and run, which would download the AI Dungeon model and install the game’s interface.

This approach made sense for a couple of reasons. First, Colab is free, which makes it a nice platform for a side project. Second, Google backs each Colab notebook with a free GPU instance, which was necessary for running the 5 GB model.

The first problem we hit was that our model barely fit on the GPU instance. If for some reason your playthrough called for a little extra memory, the whole game would likely crash.

The second — and more existential — problem was the bill.

I’d chosen Colab because of its cost-effectiveness, but what I hadn’t accounted for was data egress charges. With every Colab notebook needing to download a 5 GB model, and with users being in a variety of regions, each download was costing between $0.30 and $0.40.

When the daily bill hit $2,000 per day, BYU’s Perception, Control, and Cognition Lab (PCCL) was kind enough to handle the charges. When the cost hit $7,000, they were fine with it. At $15,000, they started to get nervous. At $20,000, we all agreed we needed to do something. At $30,000, they prepared to pull the plug.

By the time all was said and done, the total bill had reached $50,000 in three days.

Deploying GPT-2 at scale — without going broke

Within 12 hours of PCCL shutting down AI Dungeon 2, our community had hacked together a peer-to-peer solution for sharing the model via torrent, which meant the game was back up with no egress charges. (Note: This is just one of the amazing ways our community has sustained and improved AI Dungeon 2. More on that later.)

This, however, was clearly a temporary measure. The vision for AI Dungeon 2 was for it to be a game anyone could play, not just those with the tech savvy to run a Colab notebook. In order to do this, we needed to release the game as a real app.

To build a full AI Dungeon 2 app, our model needed to be deployed as a backend web service. You can imagine it as a “predict API” that our app can query with user input to generate the next stage of the story. This pattern should be familiar to anyone who has worked with microservices before.

The question is, how do you build a microservice out of an ML model?

As it turns out, there’s an open source tool that automates it, called Cortex. At a high level, Cortex:

• Wraps your model in an API and containerizes it
• Deploys your model to the cloud, exposing your API as an HTTP endpoint
• Auto scales your instances to handle traffic fluctuations

Instead of rolling our own infrastructure using Flask, Docker, Kubernetes, and a mess of AWS services, we were able to consolidate and automate our infrastructure.

This architecture allowed us to host our model as a backend for the web and mobile apps, opening up the game to players who couldn’t use Colab. It also, however, required several optimizations in order for us to make it affordable.

First, we needed to configure aggressive auto scaling. We are billed for every minute an instance is running, and in order to serve many parallel users, we need to launch many instances. To make the most of our spend, we need to spin up the minimum amount of instances needed at any given moment, and quickly take down any unnecessary instances.

Second, we needed to select the optimal instance types. That means figuring out exactly how big our instance needs to be to host the model effectively, and making use of spot instances — unused instances that cloud providers sell at a steep discount.

After some tinkering, we were able to make our Cortex deployments roughly 90% more cost effectively than our previous Colab setup. Within two weeks, our server count had peaked at 715, and we’d supported over 100,000 players. Six weeks later, we’ve passed 1,000,000 users and 6,000,000 unique stories told.

Scaling AI Dungeon has been a community effort

At each stage of development, the community has been key to unlocking our next stage of scale.

The most obvious example is the people who play AI Dungeon 2. Without them, there wouldn’t be any scale to speak of. Beyond just our players, however, we’ve had help from community members like:

• BYU PCCL paying our initial GCP bill
• The users who brought AI Dungeon 2 back online via torrent within 12 hours of it shutting down
• Braydon Batungbacal, who volunteered to build the iOS and Android apps
• The Patreon supporters who continue to fund AI Dungeon’s development
• Open source projects like Cortex that have worked to support AI Dungeon

As we continue to develop AI Dungeon — and potentially, a bigger platform for other ML-driven games — our community will no doubt continue to be a driving force behind the decisions we make and our ability to execute them.

Thank you to everyone whose been involved so far, and stay tuned for what’s coming next.