Accelerated Understanding Launches 1 Trillion Parameter Physical AI With 5 Trillion Context

Professor Anima Anandkumar launched Accelerated Understanding to develop AI models that simulate physical systems in four dimensions. The startup's models are trained with up to 1 trillion parameters and support a 1 trillion context window during training, exceeding 5 trillion at inference. This Physical AI aims to predict full trajectories of physical phenomena across multiple modalities to optimize the design of chips, robotics, and medical devices.

The models utilize a non transformer architecture called neural operators to represent 3D space and time without the shortcuts used by traditional video or world models. This approach enables the system to suggest specific design improvements by iterating through a loop of simulation and optimization. The company's framework targets engineering bottlenecks in fields such as fusion simulation, weather forecasting, and materials science.