A California startup unveiled an artificial intelligence system designed to predict physical phenomena rather than language, saying it processed 5 trillion pieces of data in a single query — a volume its founders said may exceed what one computer can handle.
The company, Accelerated Understanding, says the model points toward a single AI capable of answering physics questions across industries, replacing the separate mathematical models that engineers and scientists now build for each application. Its founders told Reuters, which reported the launch on Tuesday (Aug. 25), that they are targeting business customers first, in areas including semiconductor design, robotics, extreme weather prediction and geological analysis for energy firms.
The 5 trillion data points the system handled in one prompt is roughly 5 million times what flagship models from Anthropic and Google typically take in — the equivalent, by Reuters' comparison, of reading Tolstoy's "War and Peace" 5 million times in one sitting.
The technical departure is architectural. Systems such as ChatGPT were trained on the world's text and predict the next word in a sentence. Accelerated Understanding's model was built to predict how physical systems behave across space and time, and it does not use the Transformer architecture that Google researchers introduced and that underpins most large language models.
Instead it relies on neural operators, an approach co-founder Anima Anandkumar helped develop years ago while working on scientific computing. Anandkumar is a professor of computing and mathematical sciences at Caltech. Her co-founder, Benedikt Jenik, is an AI infrastructure engineer.
"The language-centric view of intelligence is humans at the center. Putting physics at the center is a nature-centric view," Anandkumar said.
Chip design is one case the company points to. Firms are already applying text-based AI to write code and reason through semiconductor problems. Accelerated Understanding argues that a model with a working grasp of physics can optimize materials and heat management directly, reducing the trial-and-error testing that laboratory work requires.
The same system, Anandkumar said, could forecast severe weather, guide robots or sort through geological data. She described the alternative as brittle: custom models built one at a time for each narrow problem.
The approach draws on work she led at chipmaker Nvidia, where she was hired in 2018 to head a team examining how the company's graphics processors could support frontier AI. One early project showed that AI could produce weather forecasts as accurate as the intensive computations meteorologists rely on, but far faster.
The startup is not alone in looking beyond text. Companies led by AI researchers Yann LeCun and Fei-Fei Li are developing what the field calls world models — systems meant to grasp spatial reality more reliably than AI trained on language. Accelerated Understanding's wager is that neural operators, which can model phenomena invisible to the eye, offer the better route.
