AMD Is Betting on AI That Understands the Physical World

Image: Digiopedia / Illustration

Artificial intelligence has become extremely good at processing digital information.

The next challenge is understanding the physical world.

That means recognizing three-dimensional environments, reasoning about objects and movement, and eventually helping machines interact with the world around them.

AMD made a major move in that direction in September by agreeing to acquire World Labs, an AI research company co-founded by computer-vision researcher Fei-Fei Li, in an $8.2 billion all-stock transaction. Reuters+1

The deal is centered on an emerging area often described as physical AI or spatial intelligence.

What is physical AI?

Most generative AI operates in a digital environment.

A language model works with text. An image model works with pixels. A coding model works with software.

Physical AI has a different problem.

It needs to understand a world containing objects, distances, movement, physics and uncertainty.

A robot cannot simply recognize that a cup exists. It needs to understand where the cup is, how far away it is, how to reach it and what might happen if it moves another object.

That requires a richer representation of the world.

Why AMD wants this

AMD has traditionally been known primarily for processors and graphics hardware.

World Labs gives the company access to research focused on models that understand physical environments. Reuters reported that the acquisition is intended to strengthen AMD's position in hardware, software and AI systems for emerging applications such as robotics and simulation. Reuters

The significance goes beyond one acquisition.

AI hardware is increasingly being designed around the workloads that future models will require.

If physical AI becomes an important category, chip companies will need to understand those models just as closely as they understand today's language models.

Robotics is an obvious application

Robots are one of the clearest potential beneficiaries.

A warehouse robot needs to understand its surroundings. A household robot needs to navigate unpredictable rooms. Industrial machines need to interact safely with physical objects.

Better spatial models could make those systems more capable.

Simulation is another important area.

Instead of testing every scenario in the real world, developers can use simulated environments to train and evaluate AI systems.

That could reduce development time and cost.

The computer is learning to understand space

The broader significance is that AI is moving beyond language.

Understanding words is useful, but intelligence in the physical world requires understanding space and relationships.

AMD's World Labs acquisition shows that major chip companies increasingly see this as a strategic part of AI's future. TechCrunch

The next generation of AI may therefore not just answer questions about the world.

It may be able to model the world itself.