Latitude Case Study

Latitude started as a Dungeons and Dragons plugin born at a hackathon, quickly growing to about 100,000 users in its first week and 1 million in its first month. It later established itself as AI Dungeon, creating an LLM-based game experience that generates a different narrative each time based on player choices.

Initially, it achieved results on top of GPT-3, but for richer text and more flexible interaction, a different partner was needed. However, OpenAI's non-violence policy constrained core elements like quests, swords, and combat, which conflicted with the role-playing experience Latitude wanted to provide.

There were two core challenges.

• Creative constraints: The inability to include combat and weapon elements diminished gameplay.
• Cost pressure: Serving thousands of daily users required lowering both model costs and latency at the same time.

Latitude tested a Jurassic-1 prototype together with AI21 Labs, and after fast feedback loops and iterative experimentation, brought it into production. The model was then scaled down from Jurassic-1 Jumbo (178B) to Grande (17B), resulting in a better balance among cost, response speed, and quality. The Hydra ensemble approach further improved quality by selecting the best among multiple responses.

Ultimately, Grande produced better results than smaller models like GPT-J or GPT-3 Curie, and the reduced cost allowed Latitude to more frequently experiment with new datasets and improve output quality. This case shows that AI game services require not just a model provider, but a partnership that coordinates policy, latency, and cost together.