I hadn’t heard of this game before, but it sounds interesting. They’re likely using an LLM for the vampire to convince others to let them in their house? Anyone have more info on it?
My name is Gabriel, and I wanted to share a game I’ve been developing with a small team of 4 called “Suck Up!”. It’s a bit of a unique project: a comedic deception game where players are vampires trying to convince people to let them into their houses. But with a twist – each character is powered by generative AI. Think Mafia meets Whose Line is it Anyway.
I haven’t seen anything built like this before, so wanted to share some of our learnings. One interesting part was figuring out how to create a balanced experience while still taking advantage of the emergent behavior of Large Language Models (LLMs). We ended up creating a “trust system”, which allows for some difficulty balancing, but doesn’t restrict the LLM too much, allowing for unexpected gameplay moments like the AI swearing or sudden door shuts. (Granda Maude never swears though!).
The main challenge was balancing accuracy with speed. We wanted to make the model responsive, while still adhering to gameplay rules. LLMs hallucinate frequently, and solving this usually requires multiple shots which can take processing times to 5-10 seconds sometimes. Our model currently takes just 1-2 seconds to generate a response, and it’s usually pretty accurate in keeping the role-playing up and following gameplay rules. We also structured the prompts in a way that allows for streaming. So in future versions we will be able to start sending the response almost immediately, as it’s being generated.
We had a bunch of internal debates about the ethics of AI in games when building this. My stance is we should only use AI to create game experience for players that we simply couldn’t do before, and NOT take away from the creative labors. Every art asset in the game is hand-produced or licensed. I think this is very important.