This AI-populated matrix feels like Stardew Valley
Evolving AI from reactive to proactive agents
Suppose you’re building an artificial intelligence, and you want it to feel human. As you’re developing it, how can you give it years of life-like experiences without it taking years?
In the sci-fi realm of Bladerunner, replicants are built with artificial memories. In some sense, this is a crutch that might help us duplicate a single persona to multiple instances, and have each instance equipped with a kind of history. But that’s not the same thing as enabling an AI agent to have a truly lived life experience. I suspect this kind of experience will be necessary both for a progenitor of such agents, as well as for each different AI personality or truly individual identity.
What I’m leading up to is the idea behind this week’s paper: simulating a world for AI (not for human-shaped batteries). A world where AI agents can talk to each other, have goals, and take steps to do things together.
This idea marks a transition beyond the react-only dynamics of LLMs (like GPT) into a work of dreams and initiatives. It directly addresses one of the differences that David Chalmers recently pointed out between GPT and how we think about consciousness: the agency of the AI.
Beyond that, running AI agents in a simulated world may incidentally bring us video game NPCs with new levels of individualism and creativity. Especially when they can have unpredictable and socially complex interactions!
— Tyler & Team
Paper: Generative Agents: Interactive Simulacra of Human Behavior
Summary by Adrian Wilkins-Caruana
When stakes are high, humans like to practice things before doing them for real. For example, you’d probably practice baking a cake before actually making one for your friend’s wedding. But what if you want to practice social interaction, such as a job interview? One option could be to role-play interview scenarios with a friend, or you could use generative agents instead.
Generative agents are computer programs that are designed to simulate human behavior in a believable way. The agents don’t just respond to prompts like LLMs, but rather they partake in a simulation where they interact with other agents, acquire new information, retrieve and reflect on their memories, and act on them.
Here’s a snapshot of Smallville (no relation to the TV series of the same name), a sandboxed, simulated world where a community of 25 agents are going about their day — going places, doing things, and interacting with other agents.
In the simulation, agents are initialized with a “seed” memory to describe their job and relationships with other agents. Here’s an example:
John Lin is a pharmacy shopkeeper at the Willow Market and Pharmacy who loves to help people; John Lin is living with his wife, Mei Lin, who is a college professor, and son, Eddy Lin, who is a student studying music theory;
Each semicolon-delimited phrase is an individual memory (more on these in a moment). In the simulation, John interacts with other agents based on his identity. Here’s a morning interaction with his son, Eddy:
John: Good morning Eddy. Did you sleep well?
Eddy: Good morning Dad. Yeah, I slept great.
John: That’s good. What are you working on today?
Eddy: I’m working on a new music composition for my class. It’s due this week, so I’m trying to get it finished. But I’m having so much fun with it!
John: That sounds great!
All of these interactions are captured in an agent’s memory stream, which is essentially a text-based log of an agent’s observations, such as their behaviors or the behaviors of others (like “Isabella Rodriguez is setting out the pastries”), or observations about things (like “The refrigerator is empty”).
Agents use their memory streams to synthesize new actions. For example, when Isabella is asked about what she’s looking forward to, she must recall memories that help her answer the question. Her memories are scored along three axes: 1) how recent is the memory? 2) how important is this memory to her? and 3) how relevant is the memory to the question? The image below shows Isabella’s three highest scoring memories across all of these axes for a particular question. She uses these memories to synthesize her response.
Isabella’s memory was seeded with the intention to plan a Valentine’s Day party, a party which did eventually take place (her plans succeeded!). Even though some agents might not remember to tell others about the party or to attend it, the party plans were still spread around Smallville, and the agents cooperated to plan it, ask each other on dates, and play music at it.
Generative agents have a wide range of potential applications, including creating simulations of human behavior for testing and prototyping social systems, developing more personalized and effective technological experiences, and populating online forums, virtual reality metaverses, and physical spaces as social robots. They can also act as proxies for users, providing researchers a deeper understanding of their needs and preferences.