Dropping Prometheus for Physical World AI: Anandkumar and Yennik Founded Accelerated Understanding

10 September 202657 views

A Caltech professor and her colleague launched their own startup, forgoing a stake in a Bezos-backed project. The new company aims to build models that predict physical phenomena for robotics, semiconductors, and climate.

Dropping Prometheus for Physical World AI: Anandkumar and Yennik Founded Accelerated Understanding

Two AI researchers — Animashree Anandkumar and Benedict Yenik — turned down an offer to lead Project Prometheus, which is being developed with support from Jeff Bezos. Instead, they launched their own startup, Accelerated Understanding, focused on modeling the physical world at industrial scale. Their idea is not another chatbot, but systems that understand space, time, materials, and energy.

At first glance, the decision seems surprising: the terms of joining Prometheus looked lavish. However, the researchers chose independence — and now they explain why.

Why not Prometheus?

Prometheus was founded by Jeff Bezos together with investor and biotech entrepreneur Vik Bajaj. The project has been ambitious from the start: it focuses on automating the production of complex physical systems. Prometheus later raised $12 billion in a Series B round, but the arrival of that amount no longer influenced Anandkumar and Yenik's choice.

It's not about finances. The researchers were offered a 35% stake in the company, salaries of up to $2 million a year, and more than $2 billion in funding through Series B. But under the hood, Prometheus apparently followed a different technological logic. The scientists concluded that they could achieve more by going their own way and focusing on a fundamentally different AI architecture.

Technology: from language to physics

Anandkumar describes the Accelerated Understanding concept as a shift from a "language model of intelligence" to a nature-centric approach. Unlike conversational systems such as ChatGPT, which operate on text, the new platform attempts to predict physical phenomena in space and time. This is not an attempt to improve existing generative models, but rather the creation of a different class of tools.

The key technical foundation is neural operators. Anandkumar helped advance this field during her research work at Caltech and Nvidia. The system deliberately does not rely on the Transformer architecture on which almost all modern large language models are built. The developers claim that their solution can process up to five trillion data fragments in a single query — roughly five million times more than the typical capacity of flagship models from Anthropic and Google.

Such context headroom is needed not for long dialogues, but for full-fledged modeling of complex systems — for example, how silicon behaves when heated or how an atmospheric front forms.

Where it will be useful

The startup plans to work with enterprise clients rather than release a consumer product. Priority areas include the semiconductor industry, robotics, weather forecasting, and energy.

For chip manufacturers, the technology could predict how different materials and temperature regimes affect the final performance of components — before costly physical testing. In meteorology, the new model could describe complex physical events without the need to build separate simulations for each scenario. In effect, the company offers a universal "physics engine" around which industry-specific solutions can be built.

Anandkumar has solid experience for such a task: she is a professor of computing and mathematical sciences at Caltech, and previously spent five years as a director of research at Nvidia, where she worked on GPUs and advanced AI applications. According to her, Nvidia CEO Jensen Huang encouraged the development of precisely this broader idea. Now she and her co-founder have the opportunity to realize it in full, without looking back at the investors of a major project.

Frequently asked questions

Related materials

All materials
MathAdv: What the Final Accuracy of Machine-Generated Proofs Conceals
News

MathAdv: What the Final Accuracy of Machine-Generated Proofs Conceals

A diagnostic suite of tasks spanning 13 undergraduate and graduate mathematics topics evaluates machine-proving systems in Lean 4 not only by their final results, but also by their subject-matter mastery, the quality of their informal reasoning, and their robustness to rephrased conditions. A component-by-component analysis shows that translating a problem into formal form remains the main obstacle, and that success depends heavily on the area of mathematics and how the problem is framed.

9 November 2026
AffordanceWAM: Affordances as an intermediary between scene prediction and manipulator control
Research

AffordanceWAM: Affordances as an intermediary between scene prediction and manipulator control

The work describes a generative model that jointly predicts future RGB observations, scalar fields, affordance heatmaps, and continuous robot commands, using human videos to broaden interaction diversity without action annotations or retargeting. On RoboCasa, CALVIN ABC→D, and real-world tasks, the approach delivers consistent gains over RGB-only and robot-data-only baselines.

9 November 2026
Presentation Editing as Controlled Operations: The EditPPT Framework and DeckEdit-Bench Benchmark
News

Presentation Editing as Controlled Operations: The EditPPT Framework and DeckEdit-Bench Benchmark

The authors break down slide editing into specific objects, permitted actions, and checks to ensure unrelated content remains untouched, rather than relying on free-form code generation. To measure quality, they created DeckEdit-Bench (183 tasks across 28 human-made presentations), and the multi-agent EditPPT framework achieves a 66.7–74.9% task success rate across five base models while preserving over 97% of the content.

8 November 2026
Dropping Prometheus for Physical World AI: Anandkumar and Yennik Founded Accelerated Understanding