节目 学英语,听播客 下一集
第 17 集

NVIDIA Cosmos3: When AI Learns the Real World

NVIDIA Cosmos3:当 AI 学会理解真实世界
approximately 7 minutes B1
音频已上线 点击收听
Mike

Hey Sarah, welcome back to "Learn English with Podcasts"! Today I want to talk about something really cool. Have you ever wished your robot could understand the real world?

Sarah

What do you mean? Like, not just follow commands, but actually know what's happening around it?

Mike

Exactly. That's what NVIDIA just built. It's called Cosmos 3, and it's a foundation model for what they call Physical AI.

Sarah

Physical AI? That sounds like science fiction. What does it actually do?

Mike

It helps machines understand the physical world. Think about a home robot cleaning a table. It needs to see the dishes, understand physics, and plan its actions. Cosmos 3 does all of that in one model.

Sarah

Wait, one model? Isn't that usually done with separate systems?

Mike

That's the breakthrough. Before, you needed a vision model to see, a world model to predict, and an action model to move. Cosmos 3 combines everything into a single architecture.

Sarah

That's impressive. How does it work under the hood?

Mike

It uses something called Mixture-of-Transformers, or MoT. There are two towers inside. The Reasoner tower thinks about what it sees, and the Generator tower creates videos or actions.

Sarah

So one part thinks, and the other part does?

Mike

Pretty much. The Reasoner is like the brain that interprets images, video, and text. The Generator then produces physically accurate simulations or robot movements.

Sarah

And this is open source?

Mike

Yes, fully open. NVIDIA released the model weights, training code, and even datasets. Anyone can download and use it.

Sarah

That's huge. What can people actually build with it?

Mike

A lot of things. Autonomous cars can use it to predict traffic. Warehouse robots can learn to avoid obstacles. It even generates realistic videos for training other AI systems.

Sarah

So it's like a training ground for robots? Learning in a virtual world before going real?

Mike

Exactly. NVIDIA calls it a world foundation model. It creates simulated environments where robots can practice safely without breaking anything in the real world.

Sarah

That makes sense. How big is this model?

Mike

There are two versions. Nano has 16 billion parameters and runs on a regular workstation GPU. Super has 64 billion and needs a data center.

Sarah

64 billion? That's enormous.

Mike

It is. But here's the thing. Both versions are ranked number one on several benchmarks. They beat much bigger closed models.

Sarah

Open source beating closed models? That's the real story here.

Mike

It really is. NVIDIA also created a new benchmark called HUE to test how well these models understand physics. Cosmos 3 leads there too.

Sarah

What about real-world applications? Is anyone actually using it?

Mike

Yes. Robotics companies are post-training it for specific tasks. Autonomous driving teams use it to generate rare edge cases that are hard to capture in real life.

Sarah

Like what kind of edge cases?

Mike

Imagine a self-driving car encountering a ball rolling into the street. You can't wait for that to happen naturally. Cosmos 3 can simulate thousands of those scenarios for training.

Sarah

That's actually really smart. Safer than testing on real roads.

Mike

Much safer. And much faster. You can generate years of driving experience in hours.

Sarah

You know what I find funny though?

Mike

What's that?

Sarah

The model is called Cosmos. Like the universe. But it's really just learning about... tables and chairs and robots.

Mike

Ha! That's a good point. But I guess understanding a table and a cup is actually understanding the universe, in a way.

Sarah

OK, that's deep. I'll give you that one.

Mike

And the best part? It's all free and open. So anyone, anywhere, can start building physical AI today.

Sarah

Here's to open source and robots that finally understand the real world.

Mike

Cheers to that, Sarah.

已复制