Video commentary by Károly Zsolnai-Fehér describing a paper/project where 25 ChatGPT-based agents were assigned roles (CEO, CTO, programmers, testers, documenters) and collaborated to design, code, test, and deliver playable games (examples: Five-in-a-row, Pong) using a waterfall-like process. The presenter highlights the agents' ability to plan, critique (reduce hallucinations), iterate, and produce multiple programs rapidly and cheaply (claimed: <7 minutes, <$1), and notes the project source code is publicly available.
Chapters
My goodness, this is insanity. Earlier, when running an experiment with 25 ChatGPT AI agents playing a video game together in a simulated town, absolutely magical things happened. What happened? Well, they planned their days, formed relationships, invited each other to parties which they also organized and more. I absolutely love this paper. I. And my goodness, here is an even crazier one. It has an insane idea. If these AI assistants can have their own identities, motivations, and they can converse in full natural language, why not ask them to form a company to develop video games? But how does this even work? Well, each of them get assigned an identity. For instance, we can say that you are going to be the CEO, little AI to and you are going to be the cto. Then they can make decisions together and perhaps even more importantly, they reflect upon these decisions later and evaluate whether they have done well. So what did they do? Look, they have indeed taken on roles. We now have a CEO and teams designing, coding, testing and documenting. Hmm. There is already a crazy bombshell in this image. I wonder if you fellow scholars have noticed it. I'll tell you about it in a moment. Now, they also established a chat chain that describes who gets to talk to whom. This will encourage teamwork. I love how the programmers are already taking instructions from the cto. But it doesn't stop there. Here comes my favorite part. Oh yes, cross examination. Another little tester AI is looking at the programmer's solution and says, look, I have found a bug. And when it comes to problems, they are quite picky. Then they can make suggestions as to how to improve the program some more. This is incredibly important. Why? Because of one word, hallucinations. ChatGPT is known to make up answers from thin air every now and then. And here we have another one looking at these hallucinations and saying, wait a minute, this is not quite what we are looking for. So does it work? Yes, it works spectacularly. Finally, we have someone writing the manual and of course at the end of the chain, the user installing the game, running it and evaluating it. Here you see how they go from getting comfortable with their roles to brainstorming to writing code. Loving it. Now here is a five in a row game that they wrote first without a graphical interface, and then the designers iterated on it. All of which were AIs. So good. Now what was the crazy thing in this image? Oh my, look at that. They are even using the waterfall software development model. The authors of the paper argue that it is a well established model for software development. However, it is also a rigid model where identifying errors late into the development process may not allow us to fix them. And they still managed to create something usable together. And hold on to your papers, fellow scholars, because all this happened in less than seven minutes and it costs less than $1. Wow. Bravo little AIs. Now, five in a row is not the only game they can write together. Here is the process for Pong as well. Great. But wait a minute. Is five in a row and Pong the only games they can write? Because these are games that are out there in the wild. Many people have written them before. Is it possible that it is just copying the code from somewhere? Well, fellow scholars, I have two good news for you. Good news number one is that the source code of this project is available so you can create your own AI game development company. And good news number two, a other fellow scholars have already done it, so I am itching to know the results. And now. Oh my goodness, look at that. They can write many different programs. Todo tracking apps, password generators, maze generators, even a video player, and so on. Oh, and note that many of These use GPT 3.5, which is not even the smartest version of GPT yet. Wow, what a time to be alive. So these little AIs can not only play together in a video game, they can even write the video game themselves. And what if we would ask them to write a video game and put themselves in it and perhaps do it recursively so their video game versions would also develop a video game and so on. Now that would make one super fun experiment. Hopefully we will see that in the next paper. Also, I find it very surprising that if we think about GPT as a virtual brain, which is quite a bit of a stretch, but bear with me for a moment. And perhaps the best use of it is to split it up into different little brains with different identities and have them work together. That is a crazy thought, but it works. I love it. If you're looking for inexpensive cloud GPUs for AI, Lambda now offers the best price 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 the Lambda GPU cloud, you can now get on demand H100 instances for just 1,99 per hour. Yes, 1 99. And they are one of the first cloud providers to offer publicly available on demand H100 access. 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.
