Dr. Zsolnai ran 25 ChatGPT agents in a simulated town and watched them plan days, gossip about elections, and form relationships. One agent organized a Valentine's Day party that spread through the population. Human evaluators rated the agents' behavior as more human-like than actual humans doing the same task.
Dear fellow scholars, this is Two Minute Papers with Dr. Zsolnai. Today we are going to find out what happens when several AI agents play a video game together. So in particular, what happens if we put 25 of these amazing ChatGPT AI agents, give them motivations and memory and put them in a simulated town? The answer is magic. Magic happens. What I mean by that is that interesting social behaviors emerge. They plan their days, form relationships, invite each other to parties. They love to talk and talk and talk. Hmm, I wonder what these discussions are about. Yes, of course I am going to tell you soon. Initially we find them asleep and then they wake up and start their morning routine. They brush their teeth, get in the shower and dress up. Or like good fellow scholars, they want to read papers or get this to work on their own papers. Yes, really. That's how we know that this is a good scholarly game. But all this is not surprising. The surprising part starts now. Remember, they are ChatGPT agents. So yes, they converse in full natural language. Here is an example conversation John asks Tom hey, have you heard anything new about the upcoming mayoral election? Then Tom no, not really. Do you know who is running? Yes, there really is an election going on in this little simulated town. How cool is that? For instance, the later Isabella says to I'm still weighing my options, but I've been discussing the election with Sam Moore. What are your thoughts on him? Where Tom, a separate ChatGPT agent? To be honest, I don't like Sam Moore. I think he's out of touch with the community and doesn't have our best interests in heart. Wow. They ask each other's opinions and try to be informed citizens for the elections. Really cool. So what is the result? The result is that the life these agents live resemble the interactions of real humans. Ok, but how much do they resemble real humans? That will be a huge surprise. I will tell you by the end of this video. What I also loved is that these agents also engage in self reflection. They think who am I and what is important for me. For instance, you see how one of the agents, Klaus, reflects on his observations and finds out that he is highly dedicated to doing research. Haha. A fellow scholar at heart. I see. Like it is true for us humans too. Reflection is important. Sometimes we have to look back, reflect and learn some new lessons based on our experiences. Thus these agents can also be asked what they are the most excited about. Then they check their memories and Isabella finds out that she is most excited about throwing a Valentine's Day party. And that is one of the Most interesting parts of the paper because of two reasons. Question one is, how did this party come about? And question two, did the party really happen? One, the party came about because one of the scientists specified to the character that it wants to throw a Valentine's Day party. And get this. The AI heard this as its own inner voice. Very cool. And tried to act on it. But Here comes question 2. Does the party happen? And if so, how? Organizing this is not easy at all. Not even in a small simulated game. You see, lots of things have to happen for this to take place. All of the characters have to know about the party, and they also have to show up. But she really wants this party to happen. So she tells a resident fellow scholar, Klaus, who then fortunately tells Abigail about it. And look at that. Isabella hears back about the party from someone she didn't even invite yet. Information is flowing in this little city like it would in a real city among real humans. I love it. Quoting the paper, the agents autonomously spread invitations to the party over the next two days, make new acquaintances, ask each other out on dates to the party, and coordinate to show up for the party together at the right time. Wow. And the answer is yes, the party indeed happened. Good job, Little A's. And there is one more thing that I would love to show you about this paper. So here is where I fell off the chair. Look. What? Are you kidding or is this real? Scientists at Google also asked a bunch of humans to play the game. Here they are. And then they asked a separate set of evaluators who found the AI's behavior more human like than than the humans themselves. Yes, they are more human than humans according to evaluators.
Wow.
Of course, every test like this has its limits. And note that this is just a simple game with simple interactions, not real life. But this still keeps me up at night. Some agents accepted the party invitation and some didn't. And this decision came back to them, reflecting back on who they are, what their goals are and and what their schedule is. This very much sounds like human behavior. They have their own little lives and their own preferences. Little virtual friends. Really cool experiment. What a time to be alive.
If you're looking for inexpensive cloud GPUs for AI, Lambda now offers the best prices in the world for GPU Cloud Compute. No commitments or negotiation required. Just sign up and launch an instance. And hold on to your papers. Because with Lambda GPU Cloud you can get on demand a 100 instances for $1.10 per hour versus $4.10 per hour with AWS that's 73% savings. Did I mention they also offer persistent storage? So join researchers at organizations like Apple, MIT and Caltech in using Lambda Cloud instances, workstations or servers. Make sure to go to lambdalabs.compapers to sign up for one of their amazing GPU instances today.
Thanks for watching and for your generous support, and I'll see you next time.
