They turned down Bezos' Prometheus project to build a physics AI

Anima Anandkumar and Benedikt Jenik launched Accelerated Understanding, whose model abandons the Transformer in favor of neural operators and ingested 5 trillion data points in a single query.

The duo of researchers expected to lead Project Prometheus, the AI company backed by Jeff Bezos, finally unveiled on Tuesday, August 25, what they have built on their own. Accelerated Understanding Inc claims to have developed a different kind of model, capable of absorbing 5,000 billion pieces of data in a single query in tests — about five million times what Anthropic's and Google's flagship models usually swallow.

The difference lies less in scale than in purpose. Where ChatGPT learned to predict the next word from the world's texts, this system learns to predict physical phenomena in space and time. It completely dispenses with the Transformer architecture invented at Google in favor of neural operators, a technique that Anima Anandkumar, professor of computer science and mathematics at Caltech, helped to pioneer several years ago. "The vision of intelligence centered on language places humans at the center. Putting physics at the center is a vision centered on nature," she summarizes.

Several teams, including those led by Yann LeCun and Fei-Fei Li, are pursuing "models of the world" that are supposed to better understand spatial reality than an AI trained on text. Accelerated Understanding's bet is that a sufficiently generalist neural operator, capable of predicting phenomena invisible to the eye, is the best way. Chip design is one of the first targeted outlets, as the company believes that an intrinsic understanding of physics makes it possible to optimize materials and temperatures with fewer laboratory tests. Robotics, prediction of extreme weather events and analysis of geological data for energy complete the list. The founders are primarily aiming for company contracts, not a consumer offer.