Playing AI Dungeon
What are the different AI language models?
AI Dungeon and Voyage rely on AI language models to drive the experiences. Each of these models has been developed independently and has different characteristics.
Griffin uses a 6 billion parameter GPT-J model developed by EleutherAI. Griffin is one of the most performant models we have, which means you can generate more responses in less time. Griffin also tends to be one of our more unexpected models, which has both positive and negative impact on the experience. In some instances Griffin might require more retries or editing. In others it can add interesting, unexpected turns to your adventures and games. Because it’s a comparatively smaller model (it’s still massive!), more tokens can be sent with each request, which means that you can give the AI more details to consider when generating stories, which is often useful for longer narrative arcs with important details that need to be retained. Although Griffin is available for all players, larger memory is a feature available to paid users.
Dragon uses a 178 billion parameter Jurassic-1 Jumbo model developed by AI21. Dragon is significantly larger than Griffin which means it recognizes subjects and references that Griffin misses, and does a better job recalling important story elements. It is also capable of more complex writing styles and the AI responses don’t require as many retries or player editing. However, these advanced abilities come at a cost. Dragon is much slower to generate responses and its size requires significant computing power to run. Dragon is only available to paid users.
Wyvern leverages a 17 billion parameter Jurassic-1 Grande model developed by AI21 ai model. It uses the same finetuning as Griffin and Dragon. What’s interesting about Wyvern is that even though it's not as large of a model as Dragon, it’s proving to be equal to Dragon in response complexity and coherence. Similar to Dragon, it doesn’t require as many retries and edits compared to Griffin, but it is cheaper and faster to run. It is great at remembering important story details, doesn’t repeat itself, and understands obscure concepts. Wyvern is a “best of both worlds” model.
Hydra is a composite model that generates multiple responses and picks the best one to display to the player. The best response is chosen using data we've collected from Train the AI (TTAI). Hydra is an approach that can be applied to any of our models. Given the already high cost of Dragon it’s not feasible to apply this approach.
Each of our models are finetuned on choose-your-own-adventure web stories, as well as some of our own curated data.
© Latitude 2022