When it comes to achieving artificial general intelligence (AGI), large language models just don’t have what it takes. Models like ChatGPT and Claude are great at processing and generating text, but they’re significantly less skilled at understanding how things actually move through space and time. This spatial and temporal awareness, however, is an essential skill for producing intelligence that generalizes beyond simple text prediction. That critical gap, it turns out, might be filled by gaming data. That’s the bold bet behind General Intuition, a startup arguing that the virtual worlds of video games provide a far superior training ground for AI than the static text of the internet.
According to the company's CEO, the internet is fundamentally limited as a training resource because it is a flat, text-based repository of human knowledge. It lacks physics, continuous motion, and real-time consequences. Video games, on the other hand, are governed by strict physical engines and require continuous, interactive decision-making. They force AI models to learn causality, object permanence, and spatial reasoning in a dynamic environment. When an AI navigates a digital landscape, it must understand that a character cannot walk through a wall, or that dropping an object causes it to fall. These are basic human intuitions that text alone simply cannot teach.
General Intuition is leveraging this interactive medium to train a new generation of AI models. By feeding algorithms rich, sequential data from gameplay, the startup aims to create systems that possess a genuine understanding of the physical world. This approach mirrors the way biological intelligence evolves, learning through interaction with a 3D environment rather than merely reading about it. The CEO believes that without this foundational grasp of reality, current large language models will remain sophisticated autocomplete tools, forever trapped in a web of words without true comprehension.
As the AI industry pushes toward AGI, the quest for high-quality, dynamic training data has become the primary bottleneck. If General Intuition’s thesis proves correct, the path to truly intelligent machines won't be paved with scraped web pages, but with the simulated, interactive realities of video games. This paradigm shift could redefine how developers approach AI training, turning virtual playgrounds into the ultimate classrooms for the next leap in machine intelligence.