A British startup is currently developing a new method to train artificial intelligence by leveraging data collected from video game players. The project focuses on converting the specific inputs and decision-making patterns of gamers into training sets that help AI models understand how to navigate complex, physical environments.

The core premise of the technology is that the problem-solving skills and spatial navigation required in gaming environments can be mapped onto real-world scenarios. By analyzing how humans interact with virtual spaces, the company aims to teach AI systems to better interpret and move through physical surroundings. This represents a shift from traditional AI training methods, which often rely on static datasets or simulated environments that lack the unpredictability of human-controlled gaming.

While the technology is still in its developmental stages, the startup suggests that these 'dodgy' or imperfect gaming inputs—which include human errors and reactive movements—are actually beneficial for training. These inputs provide the AI with a more diverse range of scenarios than perfectly optimized, automated data. By learning from the nuances of human play, the models are intended to become more robust and adaptable when faced with the chaotic nature of the physical world. The project highlights an ongoing trend in the tech industry to find creative, non-traditional sources of data to overcome the current limitations of machine learning models in real-world applications.