Accelerated Understanding Inc launches new AI model that ditches transformers for neural operators

Most AI companies are racing to build better transformers. Accelerated Understanding Inc decided to skip that race entirely and build something different. The startup, co-founded by Caltech professor Anima Anandkumar and Benedikt Jenik, is launching an AI model built on neural operator architecture rather than the transformer framework that powers essentially every major language model on the market. The model operates in 4D, processing three-dimensional space plus time, and is designed to understand physical phenomena with a level of fidelity that text-focused models simply cannot achieve. The numbers are staggering During training, the model can handle up to 1 trillion tokens. At inference, it exceeds 5 trillion tokens. In testing, the company demonstrated the ability to process 5 trillion data points in a single prompt. The model has also been scaled to 1 trillion parameters in pre-training, putting it in the same weight class as the largest models ever built, but with a fundamentally different architecture under the hood. Why neural operators matter Neural operators work differently from transformers. Think of a transformer as a translator that converts one sequence into anothe...