A new video from tech experimenter Nichonauta puts AI models to the test in classic video games. The video, titled "GEMINI 4 Esta IA puede JUGAR videojuegos SOLA," explores whether artificial intelligence can learn to play retro titles like Grand Theft Auto 1 by simply watching the screen.
The experiment uses an emulator to run GTA 1 and other classic games, then feeds the visual output into AI models such as Gemma 4 and Qwen. The AI is tasked with navigating the game world, interacting with objects, and completing objectives—all without human input.
"We test if models like Gemma 4 and Qwen can navigate, interact, and complete objectives in GTA 1 and other retro titles using an emulator," the description states.
Chapters in the video break down key challenges: visual recognition in GTA 1, the difficulty of mapping controls and action keys, image recognition analysis, practical applications beyond gaming, and conclusions about LLM performance. The creator also proposes testing the AI on new games in future experiments.
While the video is primarily a demonstration of AI capabilities in gaming, it highlights the potential for AI to understand and interact with complex visual environments, a skill that could have broader applications in automation and robotics.