她为什么创办 World Labs

李飞飞从 2000 年读研开始做计算机视觉。她实验室做的 ImageNet，和神经网络算法、GPU 一起，被公认为开启了现代深度学习时代。之后她做过 Google Cloud AI 首席科学家，也是斯坦福 HAI 的创始院长。

生成式 AI 兴起后，前沿研究的重心从学术界转到了工业界，她判断创业是最能产生影响的路。2024 年初，她和 Ben Mildenhall（NeRF 的作者之一）、Justin Johnson 一起创办了 World Labs。

公司的核心信念是：光靠语言训练不出完整的智能。世界不是由文字组成的，是由真实的物体组成的。科学、娱乐、机器人这些领域的很多关键问题，都需要模型理解物理世界的结构和运行规律。这就是她一直说的"空间智能"。

两年多做出了什么

建了一支在图像、视频、三维重建方向很强的模型训练团队，还收购了 SceniX，目标是做机器人仿真。

最近发布了 Atlas，一个从零训练的全模态（omni）模型。它的思路很好懂：大语言模型根据前文预测下一个 token，Atlas 根据几张 2D 图像预测"下一个视角"该是什么样子。

她说 Atlas 把生成模型和多视角几何结合起来，基本解决了计算机视觉里的老难题"稀疏重建"，效果甚至超过了专用模型。

应用方向包括：给机器人强化学习搭建训练环境，为心理治疗和娱乐生成场景，以及房地产、设计、建筑行业的真实场景重建。

为什么是 AMD

她给的理由是，要把研究放大，就得离硬件更近。没有硬件，AI 的效率和规模都会受限，而且只能停留在数字世界里，这一点对机器人方向尤其重要。

AMD CEO 苏姿丰是 World Labs 的早期投资人，两家公司去年就开始深度合作，在 AMD GPU 上做模型训练和推理优化。

合作越深入，双方越觉得把软硬件、基础模型和应用整合在一起是顺理成章的事。

她还特意强调，加入 AMD 之后会继续开放地做研究，提供最好的开放模型和端到端平台。

To Seek a Newer World

World Labs is joining AMD. This is a huge moment for World Labs, our team, and for me, and I wanted to take a moment to share what this means and why I'm so excited for this next chapter.

"Come, my friends,
'Tis not too late to seek a newer world."
— Tennyson, Ulysses

Twenty-five years ago, I started my PhD in computer vision. A few years later, my Stanford lab built ImageNet, the dataset that — together with neural networks and GPUs — is widely credited with launching the modern era of deep learning. Since then, I've had many roles: Stanford professor, Google Cloud Chief Scientist of AI, co-founder and co-director of the Stanford Institute for Human-Centered AI (HAI). But through it all, the question that has driven my work has never changed: how do we make machines see the world the way we do?

This past year, that question became more urgent than ever. The frontier of AI has moved decisively out of academia and into industry. As the founder of World Labs, I believe this is where the most consequential work in AI will happen in the years ahead. And the next great leap of intelligence is spatial intelligence — the ability to understand, reason about, and generate the three-dimensional world we inhabit.

Why I Founded World Labs

In late 2023, I started World Labs with three co-founders: Ben Mildenhall, one of the creators of NeRF; Justin Johnson, a longtime Stanford colleague and a foundational researcher in vision and multimodal learning; and Christoph Lassner. Our thesis was deceptively simple: text alone cannot capture the full richness of intelligence. The world is not made of words. It is made of physical objects — with geometry, materials, lighting, and behavior. To build AI that truly serves science, robotics, entertainment, and human creativity, we need models that understand the structure and dynamics of the physical world.

We called this mission spatial intelligence, and we set out to make it real.

What We've Built in Two Years

In just over two years, we have assembled one of the strongest applied research teams in the world in image, video, and 3D reconstruction. To accelerate our work in robotics, we acquired SceniX, a leader in scalable robotics simulation. Together, we are pushing the frontier of what generative AI can do in the physical world.

Our most recent milestone is Atlas, the first fully omni-modal foundation model for spatial intelligence. Where a large language model predicts the next token from prior tokens, Atlas predicts the next view from a few 2D images. By combining generative models with multi-view geometry, Atlas solves the long-standing computer vision problem of sparse-view reconstruction — with quality that surpasses even specialized models. Atlas also enables a wide range of downstream applications: simulation environments for robot learning, immersive scenes for therapy and entertainment, and high-fidelity reconstruction for real estate, design, architecture, and beyond.

This is only the beginning. There are many more spatial intelligence products to come — and joining AMD allows us to pursue this mission at a scale and speed that would not have been possible alone.

Why AMD

To truly scale spatial intelligence, we need to be close to the hardware. Without that, AI will remain bounded in efficiency, scale, and reach — and trapped in the digital world rather than entering the physical one, especially for robotics.

That's why I'm thrilled to join AMD as Executive Vice President and Chief Scientist. Lisa Su — AMD's CEO and a longtime believer in our mission — was one of our earliest investors. Over the past year, AMD and World Labs have been working closely together to train and optimize our models on AMD GPUs. The deeper we collaborated, the more it became clear that integrating software, hardware, foundation models, and applications is not just logical — it is the future.

Joining AMD accelerates everything: our research, our products, and our ability to deliver spatial intelligence to every industry it can transform.

What Comes Next

At AMD, I will continue to do what I've always done — open, rigorous, mission-driven research — and to deliver the best open models and end-to-end platforms in the world.

Spatial intelligence is the next frontier of AI, and we are just at the beginning.

I am deeply grateful to our team, our investors, and our partners, and I can't wait to show you what we build next.

李飞飞创办的 World Labs 被 AMD 收购，她本人出任 AMD 执行副总裁兼首席科学家。这篇文章是她写的公开信，讲了三件事： 1. 为什么创办 Wor…

AI智能总结