AI - NEW MODEL PROCESSES 5 TRILLION UNITS OF DATA - Filenews 26/8
A new AI model that can process and generate huge amounts of information has been presented by two researchers who have previously been at the center of discussions about the leadership of Project Prometheus, the artificial intelligence project backed by Jeff Bezos.
The AI model processes and generates such huge amounts of information that a single computer may not be enough to handle it, Reuters notes.
Accelerated Understanding Inc., the company founded by the two researchers, claims to have developed a different approach to artificial intelligence, which in tests managed to process 5 trillion units of data in a single prompt.
That's about 5 million times the volume that Anthropic and Google's top models can typically handle, according to the company. To understand the magnitude, it is like reading Tolstoy's "War and Peace" not once, but 5 million times in a single session.
The key difference, however, is that the new system is not designed to understand language. It is designed to understand physics, the company's co-founders, Anima Anandkumar and Benedikt Jenik, explained in an exclusive interview shortly before the official presentation of Accelerated Understanding.
From language to physics
Models like ChatGPT were trained on huge amounts of text data so they could predict what is most likely to be the next word of a sentence when interacting with the user.
Accelerated Understanding's approach is different. Its model has been trained to predict natural phenomena in space and time.
At the same time, it is abandoning the Transformer architecture developed by Google — the "T" in the name ChatGPT — and processing physics data through a technology that Anandkumar had helped develop years ago, known as neural operators.
"The language-centered approach to intelligence places humans at the center. When you put physics at the center, you're taking a nature-centered approach," said Anandkumar, a professor of computational and mathematical sciences at Caltech.
The idea of so-called world models, i.e. models that can better understand physical and spatial reality compared to systems that are mainly trained on text, is attracting the interest of more and more companies. Among them are startups linked to leading AI researchers, such as Yann LeCun and Fei-Fei Li.
Accelerated Understanding is betting on a more generalized version of neural operators, which will be able to predict even phenomena that are not visible to the naked eye — and at the same time turn this technology into commercial activity.
From chips to extreme weather forecasting One of the areas in which the company sees significant potential is semiconductor design.
Already several companies are using AI systems that program computers and process information through natural language for applications in the chip industry. Accelerated Understanding, however, believes that a deeper understanding of the laws of physics could help optimize materials and temperatures, improving semiconductor performance and limiting trial and error in laboratories.
The same technology could, according to the company, be used in robotics, in the prediction of extreme weather events, but also in the analysis of geological data on behalf of energy companies.
Instead of different, specialized mathematical models for each application, the ambition is for a single AI system to be able to answer a wide range of physics-related questions, Anandkumar said.
For the initial stage of its development, the company will focus on corporate customers rather than consumer products, the two co-founders said.
Jeff Bezos' proposal
This approach had already attracted the interest of entrepreneurs, including Jeff Bezos, founder of Amazon and billionaire investor.
In late 2024, over dinner at a fine dining restaurant in the greater Los Angeles area, biotech investor and entrepreneur Vik Bajaj discussed with Anandkumar and Jenik the possibility of working together, according to meeting records seen by Reuters.
Bajaj would later become a co-founder of Project Prometheus alongside Bezos.
Anandkumar had previously worked as a scientist at Amazon, where she had participated in closed meetings on AI organized by Bezos, and had also served as a director at Nvidia for five years. Jenik, her husband, is an AI infrastructure engineer. The two of them had already started their own company.
Bajaj later presented the two researchers with a cooperation proposal called "Project Prometheus".
According to a copy of the proposal seen by Reuters, Anandkumar could take on the role of the company's public figure, sit on the board of directors and have responsibility for its scientific vision. She and Jenik, as an observer on the board of directors, could acquire a total of 35% of the company, as well as an annual salary of 1 million. dollars, which would double to 2 million. dollars after the first three months of work.
The proposal also provided for more than €2 billion. dollars in funds through "committed rounds" of funding up to the Series B round, with investors like Bezos participating.
Prometheus declined to comment.
Eventually, Anandkumar and Jenik chose to independently continue the growth of their company, while Bezos and Bajaj proceeded to fund Prometheus, which in June 2026 raised $12 billion. In a Series B round. Prometheus' goal is to develop AI that can automate the production of complex physical systems.
The starting point at Nvidia
Anandkumar did not want to disclose details of Accelerated Understanding's funding, but said the company has already entered into partnerships with computing infrastructure providers, who have hardware clusters to develop and operate its AI systems. He did not reveal their names.
Nvidia did not respond to a Reuters question about whether it was financially backing the venture.
Anandkumar mentioned that it was Nvidia CEO Jensen Huang who initially encouraged her to go in this direction.
She joined Nvidia in 2018 and led a team of scientists who worked on leveraging the company's graphics processing units (GPUs) for cutting-edge applications in artificial intelligence.
One of the first projects showed that AI could speed up weather forecasting, achieving a similar accuracy to that of the complex calculations used by meteorologists.
The results impressed Huang, who presented Anandkumar's work on neural operators at Nvidia's annual GTC conference in 2021.
"He was really very excited," Anandkumar said. When he explained that AI could "eat the lunch" of theoretical physicists, Huang replied, "I want him to eat all their lunches."
A new AI model that can process and generate huge amounts of information has been presented by two researchers who have previously been at the center of discussions about the leadership of Project Prometheus, the artificial intelligence project backed by Jeff Bezos.
The AI model processes and generates such huge amounts of information that a single computer may not be enough to handle it, Reuters notes.
Accelerated Understanding Inc., the company founded by the two researchers, claims to have developed a different approach to artificial intelligence, which in tests managed to process 5 trillion units of data in a single prompt.
That's about 5 million times the volume that Anthropic and Google's top models can typically handle, according to the company. To understand the magnitude, it is like reading Tolstoy's "War and Peace" not once, but 5 million times in a single session.
The key difference, however, is that the new system is not designed to understand language. It is designed to understand physics, the company's co-founders, Anima Anandkumar and Benedikt Jenik, explained in an exclusive interview shortly before the official presentation of Accelerated Understanding.
From language to physics
Models like ChatGPT were trained on huge amounts of text data so they could predict what is most likely to be the next word of a sentence when interacting with the user.
Accelerated Understanding's approach is different. Its model has been trained to predict natural phenomena in space and time.
At the same time, it is abandoning the Transformer architecture developed by Google — the "T" in the name ChatGPT — and processing physics data through a technology that Anandkumar had helped develop years ago, known as neural operators.
"The language-centered approach to intelligence places humans at the center. When you put physics at the center, you're taking a nature-centered approach," said Anandkumar, a professor of computational and mathematical sciences at Caltech.
The idea of so-called world models, i.e. models that can better understand physical and spatial reality compared to systems that are mainly trained on text, is attracting the interest of more and more companies. Among them are startups linked to leading AI researchers, such as Yann LeCun and Fei-Fei Li.
Accelerated Understanding is betting on a more generalized version of neural operators, which will be able to predict even phenomena that are not visible to the naked eye — and at the same time turn this technology into commercial activity.
From chips to extreme weather forecasting One of the areas in which the company sees significant potential is semiconductor design.
Already several companies are using AI systems that program computers and process information through natural language for applications in the chip industry. Accelerated Understanding, however, believes that a deeper understanding of the laws of physics could help optimize materials and temperatures, improving semiconductor performance and limiting trial and error in laboratories.
The same technology could, according to the company, be used in robotics, in the prediction of extreme weather events, but also in the analysis of geological data on behalf of energy companies.
Instead of different, specialized mathematical models for each application, the ambition is for a single AI system to be able to answer a wide range of physics-related questions, Anandkumar said.
For the initial stage of its development, the company will focus on corporate customers rather than consumer products, the two co-founders said.
Jeff Bezos' proposal
This approach had already attracted the interest of entrepreneurs, including Jeff Bezos, founder of Amazon and billionaire investor.
In late 2024, over dinner at a fine dining restaurant in the greater Los Angeles area, biotech investor and entrepreneur Vik Bajaj discussed with Anandkumar and Jenik the possibility of working together, according to meeting records seen by Reuters.
Bajaj would later become a co-founder of Project Prometheus alongside Bezos.
Anandkumar had previously worked as a scientist at Amazon, where she had participated in closed meetings on AI organized by Bezos, and had also served as a director at Nvidia for five years. Jenik, her husband, is an AI infrastructure engineer. The two of them had already started their own company.
Bajaj later presented the two researchers with a cooperation proposal called "Project Prometheus".
According to a copy of the proposal seen by Reuters, Anandkumar could take on the role of the company's public figure, sit on the board of directors and have responsibility for its scientific vision. She and Jenik, as an observer on the board of directors, could acquire a total of 35% of the company, as well as an annual salary of 1 million. dollars, which would double to 2 million. dollars after the first three months of work.
The proposal also provided for more than €2 billion. dollars in funds through "committed rounds" of funding up to the Series B round, with investors like Bezos participating.
Prometheus declined to comment.
Eventually, Anandkumar and Jenik chose to independently continue the growth of their company, while Bezos and Bajaj proceeded to fund Prometheus, which in June 2026 raised $12 billion. In a Series B round. Prometheus' goal is to develop AI that can automate the production of complex physical systems.
The starting point at Nvidia
Anandkumar did not want to disclose details of Accelerated Understanding's funding, but said the company has already entered into partnerships with computing infrastructure providers, who have hardware clusters to develop and operate its AI systems. He did not reveal their names.
Nvidia did not respond to a Reuters question about whether it was financially backing the venture.
Anandkumar mentioned that it was Nvidia CEO Jensen Huang who initially encouraged her to go in this direction.
She joined Nvidia in 2018 and led a team of scientists who worked on leveraging the company's graphics processing units (GPUs) for cutting-edge applications in artificial intelligence.
One of the first projects showed that AI could speed up weather forecasting, achieving a similar accuracy to that of the complex calculations used by meteorologists.
The results impressed Huang, who presented Anandkumar's work on neural operators at Nvidia's annual GTC conference in 2021.
"He was really very excited," Anandkumar said. When he explained that AI could "eat the lunch" of theoretical physicists, Huang replied, "I want him to eat all their lunches."
