A US-based artificial intelligence startup, General Intuition, has successfully secured $320 million (approximately £250 million) in its latest funding round, valuing the company at $2.3 billion (around £1.8 billion). This significant investment underscores a growing belief in the company's unconventional approach to AI development: training sophisticated AI agents using extensive datasets derived from video game interactions.
General Intuition's core innovation lies in its ability to leverage millions of hours of recorded gameplay, particularly the precise action labels detailing player button presses and timings, to teach AI models spatial-temporal reasoning. This means the AI learns how to move through and understand virtual environments. The company's co-founder and CEO, Pim de Witte, highlights that this method provides a richer understanding of causality and the distinction between 'self' and 'environment' compared to approaches that infer actions solely from video footage.
The efficacy of this game-trained AI extends beyond the virtual realm. General Intuition has demonstrated that the same 'brain' powering an AI agent in a game like Fortnite can also control physical robots. In a recent demonstration, a quadrupedal robot, equipped with a single camera, navigated an office environment after just eight minutes of real-world data collection. This rapid transition from simulation to physical embodiment suggests a potential breakthrough in reducing the time and cost typically associated with training robots.
The startup was spun out of de Witte's previous venture, Medal, a platform where gamers upload and share video clips. This provided the initial, vast dataset crucial for General Intuition's model. Rather than selling the simulated 'world model' itself, which acts as an internal training environment, the company intends to market the agentic model capable of generalising from gameplay to physical reality. This scalable shortcut, utilising readily available gameplay, is General Intuition's bet against the slow and expensive process of gathering real-world robotics data.
While General Intuition's technology shows promise, the challenge of scaling such models reliably in the physical world remains. However, the substantial investment indicates investor confidence in their unique methodology to bridge the gap between virtual training and tangible robotic applications, potentially accelerating the development of more intuitive and adaptable AI.