en.Wedoany.com Reported - General Intuition, a New York-based AI startup, has completed a $320 million funding round, achieving a post-money valuation of $2.3 billion. The company, spun out from the gaming platform Medal TV, is not focused on expanding text corpora but instead leverages gaming videos, player actions, and virtual environment interaction data to build world models with spatiotemporal understanding capabilities.
General Intuition's technical approach targets the shortcomings of large language models in understanding the physical world. Models like ChatGPT and Claude excel at text generation, coding, reasoning, and knowledge-based Q&A, but they primarily learn from language and multimodal static data. They still lack sufficiently robust spatiotemporal modeling capabilities for how objects move, characters avoid obstacles, actions produce results, and environments change over time. The unique value of gaming data lies in its inherent continuous chain of "action-feedback-environmental change": player movements, jumps, turns, aiming, climbing, dodging, and collisions all yield verifiable results on screen. Models can learn about walls, shadows, spatial boundaries, path selection, character perspectives, and action consequences from these continuous segments, rather than merely understanding the "physical world" from textual descriptions. General Intuition uses this gaming data as foundational material for training world models, aiming to equip AI systems with enhanced spatial reasoning, causal judgment, and dynamic environment prediction capabilities.
Investors in this funding round include Coatue, Eric Schmidt, and researchers from MIT and Google DeepMind. Capital associated with Jeff Bezos also participated. The funds will be used to expand computing power, train new model versions, and advance API accessibility.
World models are not simply about generating game screens. General Intuition aims to train intelligent systems capable of understanding environmental states, predicting subsequent changes, and distinguishing between "self" and "external environment." This capability is crucial for physical AI: robots, drones, autonomous vehicles, and industrial intelligent agents must not only answer questions but also understand their position in space, what actions will change, how obstacles affect paths, whether target objects are reachable, and whether continuous operations pose risks. Although game scenarios are virtual environments, they provide vast amounts of low-cost, repeatable dynamic data with action labels, suitable for training models' internal representations of space and time.
General Intuition CEO Pim de Witte previously founded Medal TV, a platform that accumulated a large repository of user-uploaded game clips. After the spin-off from Medal TV, the company's core assets shifted from a gaming content community to gaming data training infrastructure. Internally, the world model training environment is called a "gym"—a training ground where AI learns action patterns from massive amounts of gaming behaviors and environmental changes. The eventual external product may be an intelligent agent model with action capabilities.
This approach also raises application boundary issues. General Intuition's models could potentially be used in drones, robots, search and rescue, and defense scenarios. Once such capabilities enter high-risk systems, the challenges extend beyond gaming data training to include use restrictions, deployment reviews, model safety, and ethical assessments. The company's current key task is to transform the spatiotemporal patterns in gaming data into model capabilities transferable to real-world environments, and to demonstrate that such models are not only applicable to virtual scenarios but can also serve the perception, prediction, and action planning of physical AI systems.
