Caltech Startup's Physics AI: Skips Transformers, 5T Values

A new AI system from a Caltech startup aims to replace traditional numerical solvers in science and engineering. Accelerated Understanding Inc. has unveiled a system that can process a spatial-temporal physical field with over 5 trillion continuous values in a single pass. This technology is architecturally different from current large language models. It operates in continuous four-dimensional space, unlike other AIs that use discrete token sequences. The company says this approach is better suited for modeling the continuous nature of the physical world. The co-founders, Anima Anandkumar and Benedikt Jenik, reportedly declined a significant offer to join Project Prometheus, choosing instead to build their system independently. This new method could change how we model and understand complex physical systems.