Modeling the zombie apocalypse through language-based agents
Aipocalypse is a for-funsies personal project where I will throw the AI kitchen sink at building an interactive narrative experience centered around a group of survivors during the zombie apocalypse.
I will learn stuff. Maybe other people will learn stuff. Hopefully the survivor agents will learn stuff (as I do plan to build a system modeling congitive processes for language-based agents).
We can generate an interesting "episode" with locations/actors/items, and then task llm agents with choosing and executing actions for the actors. Running this in a round-robin fashion, we can watch the scene play out with reasonably plausible activities and reasonably accurate state management.
But...
- it's sloooow... a single actor turn takes ~14s
- there are both prompt-engineering techniques and game-loop techniques we can use to improve this.
- it's pure working context, no episodic or semantic memory yet
- there are a few rough edges around accuracy, though we're solid enough we can move off of prompt engineering and into other techniques to address this.
- Modeling simulation entities (locations, actors, items) in both pydantic models and Neo4j graph db
- Using TOON serialization at llm boundaries to reduce token usage
- Agent to generate a "landmark" location (e.g. a gas station, library, or other building you'd find on a map) with entities and save to graph db
- Agent to generate next action for an actor, given a current episode state
- Agent to evaluate the actor's actions and determine a plausible outcome
- Langgraph flow tying it together so we can process in a loop.
- Actors' goals need to change as they achieve them or circumstances change. We'll start with a simple prompt engineering solution and test.
- Then we'll add a player actor in the mix with a simple interface and an agent to convert natural language inputs into concrete actions for the system.
- Once we have a packed experience together, then we'll turn to enriching actor agent behaviors.
- I was inspired by Adam Lucek's agentic memory demonstration and plan to scaffold the basic blocks of episodic and semantic memory and learning. But, before we can model these, we need enough actor experiences to draw on.
- for semantic memory, we'll need to start with generated backstories for survivors and for the game world and initial zombie outbreak, which we can index in a knowledge graph and expose to actor agents
- and we'll need to generate multiple episodes so we have something to store in episodic memory :)
- we'll likely need to break down gameworld generation to create each element in phases (e.g. backstory -> key actors/landmarks/objectives -> individual episodes then generated on-the-fly as needed with this context)
- I was inspired by Adam Lucek's agentic memory demonstration and plan to scaffold the basic blocks of episodic and semantic memory and learning. But, before we can model these, we need enough actor experiences to draw on.