-
talk to bots to find who is the President!
-
gather information from other participants
-
we think ChatGPT is too polite in most instances
-
sample conversation with Grace, another player (who is likely the Assassin!)
Inspiration
Many social deduction games require a lot of people to participate. We think that bots can do just as well given enough prompting. Also, it is fun to embed characteristics to bots (we all know a friend who is too predictable).
What it does
Social deduction game of 8 players. There are two teams, blue and red, led by their respective leaders, the President and the Assassin. There are also two rooms, and the players are placed randomly at the start. At the end of the game, the blue team's goal is the isolate the President and the Assassin. The red team's goal is to place the President and the Assassin in the same room. In each round, within a room, the 4 players choose a person to vote out into the other room. We made use of LLMs (ChatGPT, LLaMa, IBM Watson) to generate the dialogue for characters and trained a model on InterSystems' IntegratedML to make voting decisions based on complex combination of factors.
How we built it
Svelte Flask SQLite3 OpenAI
Challenges we ran into
Our DB was locked in the middle of the night. Turns out to be a permission issue with the parent folder! https://matthewsetter.com/sqlite-attempt-to-write-to-readonly-database/
Accomplishments that we're proud of
Completing the project as a two-people group
What we learned
How to construct human-like characteristics in GPT (quietness, likelihood to lie)
What's next for 1-epsilon
Better UI, more animation.
Log in or sign up for Devpost to join the conversation.