The former AI director of NVIDIA subverts Transformer, a physical AI with 5-trillion context that can simulate and deduce the entire universe.

It's absolutely mind-blowing!

AI has begun to predict the entire universe ——

LLMs predict the linguistic descriptions of the world, video world models predict (dynamic) images, but this time it directly predicts the physical states of four-dimensional space and time.

The most shocking part is that it adopts a brand new architecture called "Neural Operators", instead of the red-hot Transformer.

It directly outputs the complete 4D trajectory in a single inference, with the time dimension fully calculated in the process.

The long context length required for this is undoubtedly extremely large. Now the startup Accelerated Understanding has boosted its training context to the trillion (1T) level, with the inference context exceeding 5 trillion!

What does 5 trillion mean? It is 5 million times larger than the context of current top-tier LLMs! It's equivalent to asking AI to read *War and Peace* 5 million times in one go, and memorize every single word of it.

Think about it, think about it carefully.

Time is not a fundamental entity, it is more like an appearance derived from a deeper underlying information structure. Isn't that fascinating? If you push this line of thinking to the extreme, you are essentially generating the entire physical universe.

What's even more surprising is that the two founders of this startup have just rejected a lucrative recruitment offer from Jeff Bezos, the former world's richest man, which comes with 35% equity, an annual salary of 2 million USD, and a financing commitment of over 2 billion USD.

The "spiritual shareholders" standing behind them, and the possible hidden driving force, may well be that man in the leather jacket —— Jensen Huang, CEO of NVIDIA.

5 Trillion Context Is Not Ilya's "Safe Superintelligence"

Yesterday, Silicon Valley investment bigwigs spoke in vague terms, but the entire internet was abuzz, with everyone speculating that Ilya's SSI might be about to drop a game-changing breakthrough this time.

Now the secret seems to be uncovered: it is not Ilya, not even a language model, but a real "Universe Generator".

This set of staggering data from Accelerated Understanding is going viral:

• Model parameter scale: The pre-training of the model with up to 1 trillion (1 Trillion) parameters has been completed.
• Extreme scalability: The scale of expansion experiments has soared all the way to 35 trillion (35 Trillion) parameters.
• Training-period context: 1 Trillion.
• Inference-period context: Exceeding 5 Trillion.

Supported by such a huge context, this model demonstrates incredibly formidable capabilities:

When conducting physical deduction, it requires no sub-sampling, no chunking processing at all.

Facing an extremely complex 4D space physical system, it can generate the complete motion trajectory of the entire space-time with "One-shot deduction"!

Not only that, thanks to the unification of underlying logic, the very same model can handle physical problems in completely different fields at the same time!

This was unthinkable in the past.

In the past, meteorologists used meteorological models, and materials scientists used materials models; but now, Accelerated Understanding tells the world that physical laws are interconnected at the underlying level, and one general physical foundation model is more than enough to cover all scenarios.

The most disruptive core function of Accelerated Understanding lies in that it has built a perfect closed loop of "Simulation → Refinement → Simulation". It is not just a "Generator", but a real "Reality Validator".

Moreover, this targeted feedback also hits the point of Ilya's training at test time.

"Is Transformer Dead?" The Awakening of Physical AI

If you ask any AI practitioner today "What is the most popular AI architecture right now?", the answer will be 100%: Transformer.

From ChatGPT to various video generation models, this architecture invented by Google has dominated the AI industry in the past few years (it is the very "T" in ChatGPT).

But from Anima's perspective, the Transformer path has gone off track.

"The language-centric view of intelligence is 'human-centric'", Anima pointed out sharply in an exclusive interview before the launch event, "Putting physics at the center, on the other hand, is the 'nature-centric' view."

How do current LLMs and video world models work? Essentially, they are playing a "probability game".

ChatGPT is predicting the next most likely word to appear.

And those stunning video generation AIs essentially take Visual Shortcuts —— they only make the generated images "look" consistent with physical laws, but if you dig into the gravity, hydrodynamics, and material tension inside, you will find that all of them are wrong.

They don't understand physics at all, they are just high-order "pixel repeaters".

Text only has 1 dimension, video only has 3 dimensions, but the real space-time is 3-dimensional space plus 1-dimensional time.

In addition, many models adopt the autoregressive approach —— predicting frame by frame, with errors accumulating step by step.

Therefore, Accelerated Understanding has completely rejected Transformer, and also rejected visual shortcuts.

Their ultimate killer feature is a revolutionary technology that Anima helped pioneer many years ago —— Neural Operators.

Different from text processing, Neural Operators are born specifically for processing complex, invisible physical data. It does not need to force the physical world to reduce its dimension to text or pixels, but directly understands multi-physical phenomena in the complete 4D dimension (3D space + time dimension).

Invisible airflows, plasma turbulence in nuclear fusion reactors, microscopic thermodynamic distribution inside chips... these physical processes that are invisible to the naked human eye and cannot be guessed by traditional AI using "visual experience" have become clearly visible and computable under this brand new architecture.

Stunned Jensen Huang, Rejected Bezos

Reading up to here, you may ask: How much computing power is needed to train such a terrifying model? Where does the money come from?

Although refusing to disclose the current financing details, the co-founder only implicitly stated that "we have reached cooperation with computing providers, who provide hardware clusters to develop and run the AI".

But all clues point to her former old employer —— NVIDIA, and that Jensen Huang who shouts "AI is the future".

Back in 2018, Anima was hired by NVIDIA as the Director of AI Research.

Over five years, she led a team of top scientists to explore how to apply NVIDIA's GPU computing power to cutting-edge AI fields.

At that time, the team completed an early stunning project: using AI to accelerate weather prediction. The results showed that the prediction accuracy of the AI was as high as that of the extremely complex traditional computational fluid dynamics models used by meteorologists, but it was several orders of magnitude faster.

This result directly "shattered" Jensen Huang.

At NVIDIA GTC 2021, Jensen Huang took the stage in person to showcase the research results of Anima's team on "Neural Operators" to the whole world. "He was so excited at that time," Jensen Huang recalled.

When Anima half-jokingly said to Jensen Huang: "AI might take away the lunch of theoretical physicists".

At that moment, Jensen Huang's eyes lit up, and he responded domineeringly: "I want it to eat all their lunches!"

According to Anima, it was Jensen Huang who initially encouraged her to pursue and realize this crazy idea of "Physical AI".

Although NVIDIA has not officially responded to Reuters' inquiry about whether it has invested in the company today, everyone in the AI circle knows very well —— without the secret support of top-tier computing power, it is impossible to build such a behemoth with 5 trillion context.

Jensen Huang's grand strategy has finally made its move.

However, to understand how powerful this "physical foundation model" is, we have to go back to Los Angeles at the end of 2024.

In a high-end restaurant, a secret dinner that is enough to change the current AI landscape is underway.

One of the main guests at the dinner was investor and biotech entrepreneur Vik Bajaj (who later co-founded Project Prometheus with a valuation of over 10 billion USD with Bezos).

Sitting across from him was a couple with top-tier AI background: Anima Anandkumar, Professor of Computing and Mathematical Sciences at Caltech and former Director of AI Research at NVIDIA, and her husband Benedikt Jenik, a top AI infrastructure engineer.

Bajaj came prepared with an offer called "Project Prometheus". This agreement strongly backed by Bezos has terms that are so generous that they are staggering:

Invite Anima to serve as the company's spokesperson, board member, and the leader of its scientific vision;

The couple will receive up to 35% of the company's equity;

The base annual salary is 1 million USD, which will be directly doubled to 2 million USD three months after joining;

The most exaggerated part is that the agreement clearly states that investors including Bezos will provide a "commitment fund" of over 2 billion USD for all financing rounds before Series B.

In this era when financing for foundation models is getting more and more difficult and computing power costs are extremely high, this is a super golden ticket leading to financial freedom and industry influence.

However, Anima and Jenik chose to reject it.

Why? Because what they hold in their hands has far more potential than a company that "automates the manufacturing of complex physical systems". They do not want to be under the control of others, nor do they want their technological vision to be constrained by capital. What they want to do is build a "god-level model" that can directly understand and simulate the physical universe.

In the end, Anima and her husband lay low in silence, until today, they shocked the world with their independently built Accelerated Understanding.

Conclusion: The Paradigm Shift From "Generating Information" to "Optimizing the World"

August 25, 2026, is destined to be a day recorded in the history of AI development.

While we are cheering for large language models that can write a good poem and help us polish work emails, Accelerated Understanding, like a wall-breaker who traveled back from the future, coldly tells us: Language and images are only the appearances for humans to understand the world, and physical laws are the real underlying code of this universe.

Rejecting Bezos, rejecting Transformer, abandoning visual shortcuts, and going straight to the essence of 4D space-time. Anima Anandkumar and Benedikt Jenik, with the crazy data of 1 trillion parameters and 5 trillion context, announced to the world the arrival of the "Physical AI" era.

References:

https://x.com/AndrewCurran_/status/2092244031002771643

https://x.com/AnimaAnandkumar/status/2092236528898675014

https://x.com/daniel_mac8/status/2092303081