It’s been a couple years since we first reviewed AI Dungeon for GiN. Back then, ChatGPT had not yet made the scene, and artificial intelligence, or AI, was kind of a niche technology that was just getting started, having upgraded from the expert systems of yesteryear. AI Dungeon was impressive, especially considering the state of the technology at the time. It was far from perfect, and prone to repeating itself over and over again, plus occasionally making up nonsensical scenarios that didn’t advance the story you were creating or playing with it. But when it was on track, the AI was pretty amazing, adding insights to stories and also offering a few good surprises along the way.
Eventually AI Dungeon divided its AIs up into three different models called Griffin, Wyvern and Dragon. Griffin was the free model and never really improved from those early days, so it was prone to all of those aforementioned problems – though it occasionally did okay. Griffin was a better model, offering much better accuracy and also speed, plus it was also pretty good for gaming because it could keep track of characters, inventions and things like that. Dragon was the best model, and purposely designed for literary functions. But it was slow, sometimes taking 15 or 20 seconds to return an answer. And it was not great at game type applications because it seemed much more suited for say, helping to write a book than helping an adventurer explore a dungeon.
Over the years, Dragon showed the most improvements. Most of the flaws of the past didn’t happen anymore with Dragon. When they did, a simple tapping of the Retry button normally got things back on track. If not, players could jump in and right the ship themselves. Sometimes I did feel like I was doing much more work than Dragon was, but it did get better over time.
One way that players could ensure a better experience was to program the backend by adding characters, factions, locations and other miscellaneous items for a particular story which the AI could draw from. That way, if you were referring to someone named, for example, King Methuselah, the AI didn’t have to remember when you first referenced them in previous adventures. Instead, it could just look over the entry for them on the backend and have that character act accordingly. It honestly took me several months to master the backend of AI Dungeon. It almost felt like learning a programming language, but if used correctly, it could multiply the effectiveness and consistency of the stories tenfold.
Because I spent so much time learning the backend, I was a little bit miffed when developer Latitude announced that they were scrapping it in favor of a more user-friendly interface, which is the first major change we are covering in this review. The new interface involves the creation of story cards which are, I admit, much easier to create and add to stories.
Kind of like a dungeon master getting all the monsters, NPCs, towns, weapons and treasure ready before an adventure, you simply tell the program that you want to create a story card and then configure it as you go, adding it to the pile of other cards and telling the AI important things like the key words that will trigger it to look at a specific card before taking a turn in the story. Also, all of the old elements that people like me programmed using the old interface can be converted to story cards. Sometimes the change in format requires a little bit of editing of the newly converted cards, but for the most part no data is lost when converting. And those starting from scratch can learn the new interface in no time at all.
The other big change that happened recently was the addition of ChatGPT as one of the AIs that players could choose as the engine for their stories or games. Given how amazing ChatGPT is, this would seem like a good thing, and in terms of the stories that ChatGPT is able to create, it is. Even compared to Dragon, ChatGPT was a little bit more refined, and better able to hold the line when it comes to story consistency. However, that consistency comes at a steep price. ChatGPT was programmed with a lot of morals written into the code because OpenAI was afraid that people would accuse them of fostering abuse or hatred in real life. As such, ChatGPT is really militant about certain topics like sex and violence as part of its core code. However, if you think about it, what elements tend to make the movies that you watch more exciting, or the games that you play more fun? For many movies, books and games, the answer is either going to be violence or romance. And that runs smack into ChatGPT’s core morality filters.
Trying to play or craft a violent story with AI Dungeon while using ChatGPT as the AI is tough. The AI will actually stop gameplay and kind of lecture players about not doing violent acts, the sanctity of life, how we have to respect all peoples and a bunch of other odd things that are really out of context in a medieval fantasy setting. Why yes, I guess I should respect the goblin village’s right to exist and self-govern. Let me just put out this torch that I was planning to use to burn the place down with. ChatGPT is even worse when it comes to anything even remotely resembling an adult romantic theme. Even kissing might trigger the ChatGPT moralist police subroutine to activate. So, it’s great as long as you are trying to simulate either a well supervised tea party among consenting adults or a boxing match where both fighters deeply respect one another and only follow Marquess of Queensberry Rules.
At the same time, AI Dungeon retired the Wyvern AI, which was the best at traditional Dungeons and Dragons types of games. So, you only had three choices, Griffin which still is just too random and unstable, Dragon which is more tuned for literary endeavors and the puritanical ChatGPT.
But not all is lost because just before the New Year, Latitude introduced two new AIs for players to experiment with, and they are both pretty amazing. Called Mixtral and Mythomax, they offer the same or better story consistency as ChatGPT without all the moralizing. So that goblin village and their utopian autonomous collective might not be so safe after all.
Mixtral will eventually become the premium, paid version used with Dungeon AI while Mythomax will be available to both free and premium players. Both also work with the new story card interface, so its quite easy to provide a lot of depth for stories or games. For those of you who study AI, Mythomax is a Llama 2 AI with 13 billion parameters while Mixtral is an 8x7B Model from Mistral AI with 46.7 billion parameters. Mixtral is rated higher than ChatGPT 3.5 on most industry benchmarks too, so it’s quite good.
Working and playing with both Mixtral and Mythomax, I was really impressed. I created several stories designed to be played like single player adventure games and then played them using Dragon, Mixtral and Mythomax. Mixtral was the clear winner with consistent storytelling and a previously unheard of ability to craft realistic conversations and dialogue. There were times when it really felt like I was playing with a real game master.
Mythomax was nearly as good. Although it sometimes stumbled into some of the old Griffin or Wyvern problems like repeating paragraphs, it never seemed to jump the tracks too much in terms of the story, and only needed occasional nudges from me when it started to go out of bounds. Dragon performed solidly as it always does but still struggles with game-like elements, and it is beaten in terms of dialogue by Mixtral quite handily. All three are great for story-based gaming, and with Mixtral and Mythomax still in beta, it will be exciting to see how they perform and improve when officially released.
Latitude released a video recently talking about their future plans. This does not cover everything but gives a nice overview of their future roadmap, and how they want the platform to be more user-friendly.
All in all, I was highly impressed with the changes that AI Dungeon is bringing to the table in 2024. The new backend interface with its story cards is easier to use when creating new adventures, and you don’t need to study it for months to get a lot out of it. When combined with the powerful new Mixtral and Mythomax AIs, that makes for an amazing gaming experience that most people have probably never experienced before. Like many other technical experts and columnists, I predict that 2024 is going to be a huge year for AI technology, and gaming with the likes of AI Dungeon is one of those activities that will help all of us better enjoy the wonders of what AI can accomplish, even if all we want to do is have a little fun.