What if AI almost recreates GTAV gameplay?

GTAV (Grand Theft Auto V) is a popular game that has sold over 140 million copies worldwide. However, YouTuber Harrison Kinsley is releasing a video in which he uses AI to reproduce the game by himself without writing any programming code without a game engine.

He and his co-developer made a GAN version of GTAV last year using GameGAN, a game that AI reproduces Pac-Man just by watching human play. GAN stands for Generative Adversarial Networks, and uses two networks, a generative network and an identification network, to learn from each other. For example, if the former learns from a sample data set and creates a fake image, the latter wants it and provides feedback to the former so that it outputs an image that is closer to the real thing compared to the dataset, increasing accuracy.

Game GAN is an advanced AI technology developed by Nvidia Labs based on GAN. Kinsley says every pixel in the video was created by a neural network while he was playing, stating that the neural network is the whole game and there are no game engine rules here.

GANs can significantly reduce the huge amount of data hitherto essential for machine learning, but require a lot of GPUs. Therefore, it is said that Nvidia loaned the DGX Station A100 to Kinsley and others for this project. The system features four Nvidia A100 GPUs and a 64-core AMD server CPU. Kinsley said that it used this computing power to run 12 rule-based AIs simultaneously. This led to the development of super-sampling AI to clean the display output while collecting the data needed to create the game, making pixel imperfections invisible.

This video shows how the game GAN mimics many of the systems built into GTAV. For example, depending on the movement of the car, the shadows and reflected light on the rear glass change, as does an object approaching from a distant mountain. Kinsley and his group said that when they first started learning AI, they didn’t think about it this far.

Of course, not all GTAVs are perfectly reproduced. One of these is the crash problem. It is said that an oncoming police car split into two collisions with his own.

It is unreasonable to reproduce by learning the gameplay through a personal game GAN, but the result of the generated project is published on GitHub, and it is said that the demo can be run on most computers. It is also possible to enjoy an AI-made GTAV that no human has written a single line of code.