It's the first time in the history of the world.
It's so shocking!
AI is starting to predict the collapse of the universe
The LLM predicts the language description of the world, the video world model predicts the dynamic image, and this one directly predicts the physical state of space and time in four dimensions.
Most strikingly, it uses an entirely new architecture of "neural algorithms" rather than the Transformer in the middle of the day.
In a single reasoning, complete four-dimensional trajectories are directly output, including time dimensions.
There is no doubt that the length of context required is enormous, and now the startup Accelerated Understanding is in training to raise the context to the 1 trillion (T) level, which is more than 5 trillion!
What is the concept of 5 trillion?5 million timesThat's the equivalent of giving an AI a breath after reading five million lines of war and peace, and it remembers everything.
Your stuff, your little stuff.
Time is not a fundamental thing, it's more like an image derived from a deeper information structure, doesn't it make any sense?
More surprisingly, the two founders of the startup have just turned down Jeff Bezos, the world's richest man, to bring them back to life.35% stake, $2 million in annual salary and over $2 billion in financing commitmentsIt's the weight of gold.
The "spiritual shareholder" and possible underdog behind them may have been the man in the leather jacket, CEO Huang Ren.
5 trillion is not Ilya's "security super-intelligence".Yesterday, the Silicon Valley investment bubble was vague, but the whole web was boiling, and there were speculations that Ilya's SSI could make it bigger this time around.
And now the bottom seems to reveal that it is not Ilya, not even a language model, but a real "generator of the universe".
Accelerated Understanding's group of horror data has been wildly circulated:
- Model parameter scale: Model pre-training of up to 1 trillion (1 Trillion) parameters has been completed.
- Extremely scalable: The scale of expanded experiments has already risen to 35 trillion (35 trillion) parameters.
- In the course of the training period: 1 trillion (1 trillion).
- In the last few months, more than 5 trillion (500 trillion) people have lost their lives.
Supported by such a huge up-and-down context, the model demonstrates a terrifying ability:
It's a physical projection, but it's not a physical projection.There is no sub-sampling and no chunking.
In the face of an extremely complex 4D space-physical system, it can"One-shot" creates a complete motion trajectory throughout time and space!
Not only that, but because of the uniformity of the underlying logic,The same model can solve physics problems in completely different fields at the same time!
This was previously unthinkable.
In the past, meteorologists used weather models, materials scientists used material models; now, Accelerated Understanding tells the world that the laws of physics are universal at the bottom, and that a general model of physics is big enough to cover the sky.
The most disruptive of the core functions of Accelerated Understanding is that it builds a system of accelerated understanding.It's not just a "generator", it's a real "reality verifier" (Reality Validator).。
And this time, the focused feedback also hampered Ilya's test training.
If you ask any AI practitioner today, "What's the hottest AI architecture right now?" the answer is 100%: Transformer.
From ChatGPT to large-scale video-generation models, this architecture, invented by Google, has dominated the AI world over the past few years -- the "T" in ChatGPT.
But from Anima's point of view,The Transformer's trajectory is off course.
"Language-centered intelligence is 'human-centered,'" Anima pointed out bluntly in a pre-conference exclusive interview, "and putting physics at the center is 'nature-centered.'"
How do the current models of the LLM and video worlds work? They are essentially playing "probability games".
ChatGPT is predicting the next most likely word.
And these amazing videos are generated by AI, which is essentially walking.Visual shortcutsThey only make the images "look" in accordance with the laws of physics, but if you look deeply into the gravity, fluid mechanics, and material stresses, you'll find that they're all wrong.
They don't understand physics at all, they're just high-end "pixel repeaters".
While text has only one dimension, video has only three dimensions, real time space is 3D space plus 1D time.
In addition, many models use self-return predictions that accumulate gradual errors.
As a result, Accelerated Understanding completely rejected Transformer and rejected visual shortcuts.
The ultimate killer they sacrificed was a revolutionary piece of technology that Anima pioneered many years ago.Neural Operators。
Unlike text processing, neurotransmitters are designed specifically for processing complex, invisible physical data. They do not require the physical world to be forced down into text or pixels.Understanding multiphysical phenomena directly in a complete 4D dimension (3D space + time dimension).
The invisible airflow, the plasma flux in nuclear fusion reactors, the micro-thermodynamic distribution inside the chip... these physical processes that cannot be seen by the human eye, which traditional AI cannot guess with "visual experience", become clearly visible and computable under this new architecture.
In the meantime, I'd like to say that I'm not the only one who can do this.Seeing this, you might ask, how much math does it take to train such a terrible model? Where does the money come from?
While declining to disclose current funding details, the co-founder only implicitly stated that "we have entered into a partnership with computing providers that provide a cluster of hardware to develop and run AI".
But all the clues point to her ex-husband, Huang, and the man who shouted "AI is the future", Jensen Huang.
In 2018, Anima was hired by INVIDA as Director of AI Research.
Over the course of five years, she led a team of leading scientists to explore how to apply Ingrid's GPU computing to cutting-edge AI.
At the time, the team did an early surprise project: AI-accelerated weather forecasting. The results showed that AI forecasts were as accurate as traditional computational fluid mechanics models used by meteorologists, but were several times faster in number.
The result was a direct "shake-up" of Huang Ren.
At the GTC conference in 2021, Huang himself took the stage to show the world the results of the Anima team's research on "neurotransmitters". "He was so excited", recalled Jensen Huang.
When Anima jokingly told Yellow, "AI might steal the lunches from those theoretical physicists".
Then he opened his eyes and said,"I want it to eat all their lunches!"
According to Anima, it was Huang Rinpoche who initially encouraged her to pursue and realize this crazy idea of "physical AI".
While officials have not yet responded to Reuters' inquiries about whether they have invested in the company, in the AI world, it is clear that there is no hidden support for top-notch computing and that it is impossible to unravel the enormity of the 5 trillion-dollar context.
The old man's big game is finally down.
However, to understand how many cows there are in this "Big Physical Model", we have to go back to Los Angeles at the end of 2024.
In an upscale restaurant, a secret dinner is taking place that will change the landscape of AI today.
One of the stars of the dinner was investor and biotech entrepreneur Vik Bajaj, who later co-founded Project Prometheus with Bezos, which was valued at over $1 billion.
A couple sitting across from him with a top-level AI background:Anima Anandkumar, former Director of AI Research at Caltech, and her husband Benedikt Jenik, a top AI infrastructure engineer.
Bajaj has come up with an Offer called "Project Prometheus", a deal backed by Bezos that is overwhelmingly conditional:
Invite Anima to serve as the company's spokesperson, board member and scientific vision leader;
The couple will get a GED.35% stake in the company；
A basic salary of $1 million a year, directly doubled three months after taking office toTwo million dollars；
Most exaggeratedly, the agreement is written in black and white, with investors including Bezos providing pre-Round B financing.More than $2 billion in "commitment funds"。
In an era of increasing financial difficulty and high computational costs, this big model is a super-golden ticket to wealth freedom and industry power.
However, Anima and Jenik chose not to.
Why? Because what they have in their hands has far greater potential than a company "automated to manufacture complex physical systems". They don't want to be overwhelmed, nor do they want their technological vision to be overwhelmed by capital. They want to build a "divine model" that directly understands and simulates the physical universe.
In the end, the Anima and the Anima remain silent until this day, with their separate creation.Accelerated UnderstandingShocked the world.
Transition from "generating information" to "optimizing the world".August 25, 2026, is the day that will be included in the history of AI development.
When we're cheering for a great poem for a big language model that can help us with our orange work emails, Accelerated Understanding tells us coldly, like a wall-breaker coming back from the future:Language and images are only the representation of the world that humans understand, and the laws of physics are the true underlying code of the universe.
Rejecting Bezos, rejecting Transformer, abandoning visual shortcuts, and directly hitting the essence of 4D space and time. Anima Anandkumar and Benedikt Jenik used 1 trillion parameters and 5 trillion context data to announce the world."Physical AI".The time has come.