Physics-first AI model debuts, handles 5 trillion data

A pair of researchers formerly considered to lead Jeff Bezos’s Project Prometheus unveiled a physics-first AI model developed by Accelerated Understanding Inc that can ingest and generate far more data than current language models, handling around 5 trillion data points in a single prompt. They emphasize the system is not language-focused but designed to predict physical phenomena in space and time, built on neural operators rather than the traditional Transformer architecture. They argue this physics-centric approach offers a “nature-centric” view of intelligence and could advance engineering fields such as chip design, with potential applications beyond standard AI tasks. The duo described the model as vastly larger in data capacity—about five million times bigger than leading models from Anthropic and Google—intending to apply it to understanding the universe and physical reality. The development follows their earlier involvement with Prometheus, signaling a pivot from a Bezos-backed project to independent work aimed at broader scientific understanding. Their public remarks come ahead of a corporate launch, positioning Accelerated Understanding’s work as a benchmark for physics-first AI research and its potential to outpace language-oriented AI in certain domains.