Researchers have managed to create a real-time version of Counter-Strike: Global Offensive in artificial neural networks. This achievement highlights not only the field of game development and simulation, but also the ability of generative AI to learn complex systems.
AI creates CS:GO simulation with a single GPU
Researchers have developed an AI model that simulates Counter-Strike: Global Offensive (CS:GO) using a single NVIDIA RTX 3090. The model, called DIAMOND, was trained using footage from CS:GO’s deathmatch mode played on the Dust 2 map.
Eloi Alonso, one of the researchers, shared a video showing the AI simulation. While it was impressive that the model was able to mimic the basic elements of CS:GO, such as player movement and weapon use, it also contained a number of errors that showed the limitations of current generative AI.
For example, players in the simulation could jump without limits because the model didn’t understand the Source engine’s gravity or collision detection. Weapons would sometimes change shape or appearance in different lighting, and players would occasionally pass through walls.
Despite the bugs, experts say the study shows how far generative models have come in inferring behavior from video data alone. While training such a model currently requires large computational resources, it also demonstrates the potential of AI to simulate real-world systems.