Caltech Scientists Found AI Startup Using Neural Operators to Solve Physics Problems

Caltech researchers Anima Anandkumar and Benedikt Jenik have launched Accelerated Understanding Inc, an enterprise AI company built on neural operators rather than the Transformer architecture dominant in mainstream AI. Unlike Transformer-based models that process data as discrete tokens, neural operators work in continuous space and learn function-to-function mappings, making them better suited for physics and engineering problems. The company claims its system can process 5 trillion data points in a single prompt, compared to roughly one millionth of that for leading models like Claude or Gemini. The startup targets industries such as oil and gas, materials science, and industrial optimization, where solving differential equations and fluid dynamics problems at scale is critical. Anandkumar, a former head of NVIDIA's AI research, and Jenik, a mathematician, are betting that neural operators can handle a class of physics-heavy problems that Transformers fundamentally cannot.