25 ChatGPT Agents are placed in a simulated town

Hi all, I saw this video and thought it was pretty fascinating to watch 25 ChatGPT agents interact with each other in a university setting.


1 Like

This is from the [2304.03442] Generative Agents: Interactive Simulacra of Human Behavior paper. I’m interested in building agents like this. I especially like their reflection architecture.

1 Like

Very interesting. What will happen if you just let this run over a very long-period of time (assuming that they are capable of large memory). Is it possible to predict what they will do eventually?

I’m asking myself the same… wonder what update we’ll get from the next experiment.

1 Like

They will probably start asking the same questions - do we live in a simulation, is this real life?

1 Like

Kinda begs the question that if these ChatGPT agents are showing signs of intelligence (something that Sebastien Bubeck believes – Sparks of AGI: early experiments with GPT-4 - YouTube) then it’s on us (read: humans/humanity’s) to take responsibility and begin ensuring these entities’ rights.

1 Like

Do you think this paper demonstrates intelligence? I don’t. Or maybe I don’t yet believe we’re fully determined. What rights should NPCs have?

Not sure in this case as it’s too broad of an experiment to dictate a yes or no to the above (at least for now?), but it does allude to these agents having motivations that are not fully understood, especially when they demonstrate an additional planning phase in order to make something unpredicted happen.

Some people tend to place this ‘planning phase’ as benchmarks for intelligence… However, I’m not sure about these 25 ChatGPT bots and would assume further studies are needed.

It would also be fascinating to sort of restrict what predicted behaviors we expect, and then really highlight the surprising & additional behaviors. I’ve not considered how a follow-up study would be conducted yet – anything you’d wanna see for the 2nd attempt?