According to Anima Anandkumar, Accelerated Understanding is building large scale AI models that can simulate and understand physics not just to describe the world, but to invent and discover within it.
The goal is ambitious: models that understand the physical world directly in 4D—3D space time and across multiple physical phenomena.
Going fully 4D requires an enormous amount of context. Accelerated Understanding says it has pushed context to 1 trillion during training and beyond 5 trillion during inference.
And this matters because AI generating a bigger haystack of ideas doesn’t necessarily solve the invention problem.
The bottleneck is increasingly shifting from generating ideas to testing them.
If AI can accurately simulate and understand physics, it can attack that bottleneck directly allowing researchers to test possibilities computationally rather than relying solely on expensive, slow real world experimentation.
People have been trying to solve this problem for years, but often by taking shortcuts.
Narrow surrogate models can work extremely well when you have enough of exactly the right data and your design problem stays within the distribution the model has seen.
Video models look spectacular, but they can sidestep physical accuracy. Some static world models eliminate physics almost entirely.
And importantly, a huge amount of interesting physics isn't visual.
So what does not taking those shortcuts look like?
Space remains 3D. Time remains part of the model. That's 4D.
And the model needs to understand multiple physical modalities not just what can be seen.
That is the system Accelerated Understanding says it has built.
The key ingredient is scale.
To represent the physical world, you need enough context to capture what is happening across space and time. In a 4D setting, individual samples become enormous so large that they no longer fit comfortably on a single accelerator or even an entire node.
Accelerated Understanding has developed architectural techniques to make this tractable.
The company says its models have reached 1 trillion parameters during large-scale pretraining, can train with up to 1 trillion tokens of context when needed, and can perform inference with more than 5 trillion tokens of context, without subsampling or patching.
The underlying idea builds on earlier successes in AI driven weather forecasting, fusion simulation, medical device design, drug discovery and chip design.
But the question now becomes:
What happens when you combine that physical understanding with the scale and universality of modern foundation models?
That is what makes this so interesting.
We may be moving from AI that primarily predicts information about the world toward AI that can model the world itself and use that model to discover what could exist.
"Accelerated Understanding" (https://x.com/AnimaAnandkumar/status/2092236528898675014)