就在刚刚,芯片巨头 AMD 正式官宣,将收购李飞飞创立仅两年的 AI 独角兽 World Labs。
本次交易将以全股票方式开展,作价约 82 亿美元。当下双方已经签署最终协议,预计在 2026 年底前完成交易,但仍须取得监管批准并满足其他常规交割条件。


交易完成后,李飞飞将出任 AMD 执行副总裁兼首席科学家,直接向 CEO 苏姿丰汇报;World Labs 另外两位联合创始人 Justin Johnson 和 Ben Mildenhall 也将继续带领团队,推进模型研究。
AI 圈的两位硅谷华裔女王,正式完成了历史性的「双向奔赴」。

82 亿美元,AMD 买下了一个「新世界」
要理解 AMD 为什么愿意花 82 亿美元,就得先看看 World Labs 究竟在研究什么。
在官宣长文中,李飞飞引用了诗人丁尼生《尤利西斯》中的名句:
「来吧,我的朋友,去探寻一个更新的世界,现在尚不算晚。」
对于她来说,这个「新世界」指向一个难题:AI 已经很擅长处理语言,但它能否理解语言所描述的真实世界?
李飞飞在文章中写道:「宇宙不是由文字组成的,而是由真实事物组成的。」科学研究、娱乐创作和机器人行动,都要求模型理解物体的位置、空间的结构,以及环境会如何随着视角或动作变化。只会处理文字,显然不足以完成所有这些任务。

为此,李飞飞在 2024 年初与两位 AI 研究员 Ben Mildenhall、Justin Johnson 共同创办 World Labs,把研究方向放在「空间智能」(Spatial Intelligence)上。
这支当下约 70 人的团队,汇集了数据、计算机视觉、大模型训练、平台架构和计算机图形学领域的研究人员。今年 7 月,World Labs 还收购了 SceniX,将机器人仿真能力纳入自己的研究版图。
AMD 在最新公告中进一步说明,World Labs 正在开发可以依据文本、图像和视频输入,生成、重建并模拟可交互三维环境的空间智能模型,同时研究机器人学习与仿真工艺。
近期发布的 Atlas,是 World Labs 展示研究进展的一项成果。
大语言模型借助前面的文字预测下一个 Token,Atlas 处理的难题则是:给定二维图像,如果把摄像机移到另一个位置,那里应该看到什么?
按照李飞飞的介绍,Atlas 从头训练,结合生成式模型与多视图几何,依据有限的输入图像预测新的观察视角。她称,Atlas 在相关任务上的表现超过了一些专用模型,并在长期困扰计算机视觉领域的「稀疏重建」难题上取得进展。
所谓稀疏重建,可以简单理解为:手里只有少量照片,却希望推断出照片没有拍到的空间。Atlas 要应对的正是这一类难题,当然,可以预测新视角,并不意味着模型已经掌握了真实世界的全部物理规律。

工艺的应用情形已经浮现。李飞飞提到,机器人强化学习可以使用生成的环境开展训练;治疗和娱乐领域可以用它创建情形;房地产、外观和建筑行业,则可能用它重建真实空间。World Labs 收购 SceniX,也让从三维情形生成走向机器人仿真的路线更加清晰。
AMD 买下的,既有 World Labs 现阶段的模型成果,也有一兼容续研究 AI 如何理解空间和物理世界的团队。
算力焦虑下的双向奔赴
工艺方向明确,创始团队也颇具分量,李飞飞为什么在公司成立两年多后挑选加入 AMD?
在官宣长文中,她回顾了自己的职业生涯。从 2000 年作为研究生进入 AI 领域,到主持 ImageNet 项目,推动计算机视觉研究;从担任谷歌云 AI 部门的首席科学家,到成为斯坦福大学以人为本人工智能研究院的创始院长,她始终希望让 AI 参与应对现实世界的难题。
创办 World Labs,是她在生成式 AI 时代继续追求这一目标的挑选。但随着研究深入,模型规模、计算效率和硬件适配,逐步成为必须应对的难题。

「要加速进入这个将来,需要扩大笔者的工作规模、扩大影响范围,并更贴近硬件。」李飞飞在文章中写道,如果缺少有针对性的硬件兼容,AI 的效率与规模都会受到限制,她希望发展的空间智能也难以走出数字环境。
AMD 与 World Labs 的合作并非从这次收购才开始。李飞飞称,苏姿丰是 World Labs 的早期投资者,也是她多年的朋友。过去一年,两家公司围绕 AMD GPU,深入开展了模型训练与推理优化。
合作越深,双方越能看到彼此的需要。World Labs 需要算力和工程资源,才能扩大前沿模型研究;AMD 则需要了解新模型如何运行、消耗哪些资源,以及将来的硬件该为哪些任务做准备。

AMD 在收购公告中明确提到,随着 AI 进入推理、机器人、仿真和物理 AI 领域,计算基础设施面对的需要正在变得更加多样。World Labs 的模型研究经验,可以协助 AMD 观察这些工作负载的变化,并影响将来的工艺路线。
换句话说,AMD 希望研究人员在提出计算需要时,硬件团队就在同一张桌子旁。
正如李飞飞在接受外媒采访时所说,加入 AMD,将为 World Labs、AMD 及整个生态给出一个机会,加快软件与硬件的共同开发。交易完成后,World Labs 团队将继续专注于模型研究,而李飞飞也会直接参与 AMD 的工艺领导工作。
药物发现、创意生产、学习方式和下一代机器人,是李飞飞列举的长期目标。眼下更具体的任务,则是让空间智能研究取得足够的计算和工程兼容。
算力大战进入 2.0,苏妈硬刚老黄
如果把视线从 World Labs 移向芯片行业,就更容易理解 AMD 的另一层考虑。
在苏姿丰执掌 AMD 的十二年里,公司在处理器市面逐步摆脱了长期追赶英特尔的处境。但在用于训练和运行 AI 模型的加速器市面,英伟达仍然占据显著长处。
人们都知道,老黄最擅长的就是「卖芯片+搭生态」,让客户可以极速上手。最近,英伟达更是狂砸 130 亿美元收购了另一家明星初创公司 Hugging Face,试图用开源生态圈住庞大的开发者。

既然老黄要在「文本与大语言模型」的开源生态里称王。那苏妈就另辟蹊径,拉上李飞飞,直接在「物理世界和空间智能」的生态里插上 AMD 的大旗。
空间智能、机器人和仿真模型的计算需要,与今天最常见的聊天模型并不完全相同。及早参与这些模型的研发,有助于 AMD 判断下一代 GPU 及配套平台应该如何外观。
苏姿丰在接受外媒采访时解释:「你对端到端的了解越深,就越能构建出更好的平台。这就是笔者收购 World Labs 的原因。」
她同时表示,AMD 的雄心是「定义 AI 计算的将来」。

对于客户而言,芯片公司给出的东西也在变化。购买加速器之外,他们还需要可以运行模型的软件、便于部署的平台,以及合适具体业务的模型挑选。
苏妈明确表示,AMD 将来还会给出开放模型,也会给出专有模型。AMD 与 World Labs 则计划共同建设覆盖硬件、软件、平台和开放模型的 AI 生态,让模型研究与计算平台持续协同。
因而,82 亿美元的交易,既是 AMD 对李飞飞团队的投资,也是一次对将来工作负载的押注。收购仍有待完成,空间智能距离大规模实际应用也还有路要走,但芯片巨头显然不想等到需要成熟后,才开始研究模型究竟需要什么。
本文来源:https://www.ifanr.com/1682291
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
Just now, chip giant AMD officially announced that it will acquire World Labs, the AI unicorn founded by Fei-Fei Li just two years ago.
The transaction will be conducted entirely in stock, valued at approximately $8.2 billion. The two parties have already signed a definitive agreement, and the deal is expected to be completed by the end of 2026, but it remains subject to regulatory approval and other customary closing conditions.
After the transaction is completed, Fei-Fei Li will serve as AMD's Executive Vice President and Chief Scientist, reporting directly to CEO Lisa Su; World Labs' other two co-founders, Justin Johnson and Ben Mildenhall, will also continue to lead the team and advance model research.
Two Chinese-American queens of Silicon Valley's AI world have formally completed a historic "meeting each other halfway."
For $8.2 billion, AMD bought a "new world."
To understand why AMD was willing to spend $8.2 billion, one must first look at what World Labs is actually researching.
In the lengthy official announcement, Fei-Fei Li quoted the famous line from poet Alfred Lord Tennyson's "Ulysses":
"Come, my friends, it is not too late to seek a newer world."
For her, this "new world" points to a difficult problem: AI is already very good at handling language, but can it understand the real world that language describes?
In the article, Fei-Fei Li wrote: "The universe is not composed of words, but of real things." Scientific research, entertainment creation, and robotic action all require models to understand object positions, spatial structures, and how environments change with viewpoint or action. Merely handling text is clearly insufficient to complete all these tasks.
To this end, in early 2024, Fei-Fei Li co-founded World Labs with two AI researchers, Ben Mildenhall and Justin Johnson, setting the research direction on "Spatial Intelligence."
This team of currently about 70 people brings together researchers from data, computer vision, large-model training, platform architecture, and computer graphics. In July of this year, World Labs also acquired SceniX, incorporating robotic simulation capabilities into its research landscape.
In the latest announcement, AMD further explained that World Labs is developing spatial intelligence models that can generate, reconstruct, and simulate interactive three-dimensional environments based on text, image, and video inputs, while also researching robot learning and simulation workflows.
Atlas, released recently, is an achievement demonstrating World Labs' research progress.
Large language models use preceding text to predict the next Token; the difficult problem Atlas tackles is: given a two-dimensional image, if the camera moves to another position, what should be seen there?
According to Fei-Fei Li's introduction, Atlas is trained from scratch, combining generative models with multi-view geometry to predict new viewing angles based on limited input images. She said that Atlas outperforms some specialized models on related tasks and has made progress on the "sparse reconstruction" problem that has long troubled the computer vision field.
Sparse reconstruction can be simply understood as: you only have a small number of photos in hand, but hope to infer spaces the photos did not capture. The difficult problem Atlas tackles is exactly this kind. Of course, being able to predict new viewing angles does not mean the model has already grasped all the physical laws of the real world.
Application scenarios have already emerged. Fei-Fei Li mentioned that robot reinforcement learning can be trained using generated environments; healthcare and entertainment fields can use it to create scenarios; real estate, exteriors, and construction industries may use it to reconstruct real spaces. World Labs' acquisition of SceniX also makes the path from three-dimensional scenario generation to robotic simulation clearer.
What AMD bought is both World Labs' current-stage model achievements and a team continuing to research how AI understands space and the physical world.
A mutual meeting halfway amid compute anxiety.
The technology direction is clear, and the founding team is also quite formidable. Why did Fei-Fei Li choose to join AMD more than two years after founding the company?
In the lengthy official announcement, she reviewed her career. From entering the AI field as a graduate student in 2000, to leading the ImageNet project and advancing computer vision research; from serving as Chief Scientist of Google Cloud AI, to becoming the founding director of Stanford University's Institute for Human-Centered Artificial Intelligence, she has always hoped to have AI participate in solving real-world problems.
Founding World Labs was her choice to continue pursuing this goal in the generative AI era. But as the research deepened, model scale, computational efficiency, and hardware compatibility gradually became problems that had to be addressed.
"To accelerate entry into this future, I need to scale up my work, expand its scope of impact, and get closer to the hardware," Fei-Fei Li wrote in the article. Without targeted hardware support, AI efficiency and scale will be constrained, and the spatial intelligence she hopes to develop will be difficult to move beyond the digital environment.
The collaboration between AMD and World Labs did not begin with this acquisition. Fei-Fei Li said Lisa Su is an early investor in World Labs and a longtime friend. Over the past year, the two companies have deeply carried out model training and inference optimization around AMD GPUs.
The deeper the collaboration, the more both sides could see each other's needs. World Labs needs compute and engineering resources to scale frontier model research; AMD needs to understand how new models run, what resources they consume, and what future hardware should be prepared for which tasks.
AMD explicitly noted in the acquisition announcement that as AI enters inference, robotics, simulation, and physical AI fields, the needs faced by compute infrastructure are becoming more diverse. World Labs' model research experience can help AMD observe these changes in workloads and influence future technology roadmaps.
In other words, AMD wants the hardware team already at the same table when researchers propose compute needs.
As Fei-Fei Li said in an interview with foreign media, joining AMD will provide an opportunity for World Labs, AMD, and the entire ecosystem to accelerate joint software and hardware development. After the transaction is completed, the World Labs team will continue to focus on model research, and Fei-Fei Li will also directly participate in AMD's technology leadership work.
Drug discovery, creative production, learning methods, and next-generation robots are the long-term goals Fei-Fei Li listed. The more immediate task now is to secure sufficient computational and engineering support for spatial intelligence research.
The compute war enters 2.0, as Su goes head-to-head with Huang.
If the view is shifted from World Labs to the chip industry, AMD's other consideration becomes easier to understand.
During Lisa Su's twelve years leading AMD, the company gradually escaped its long-term position of catching up with Intel in the processor market. But in the accelerator market used for training and running AI models, NVIDIA still holds a significant advantage.
Everyone knows that Jensen Huang's greatest strength is "selling chips + building an ecosystem," letting customers get started extremely quickly. Recently, NVIDIA even splashed out $13 billion to acquire another star startup, Hugging Face, trying to use an open-source ecosystem to lock in a massive developer base.
Since Huang wants to become king in the open-source ecosystem of "text and large language models," Su has taken a different path, bringing Fei-Fei Li with her and planting AMD's banner directly in the ecosystem of "the physical world and spatial intelligence."
The compute needs of spatial intelligence, robotics, and simulation models are not entirely the same as those of today's most common chat models. Early participation in the R&D of these models will help AMD decide what the next-generation GPU and accompanying platform should look like.
Lisa Su explained in an interview with foreign media: "The deeper your end-to-end understanding, the better platform you can build. That is why we acquired World Labs."
She also stated that AMD's ambition is to "define the future of AI computing."
For customers, what a chip company provides is also changing. Besides purchasing accelerators, they need software capable of running models, easy-to-deploy platforms, and model selection suited to specific businesses.
Su made clear that in the future AMD will also provide open models and proprietary models. AMD and World Labs plan to jointly build an AI ecosystem covering hardware, software, platforms, and open models, so that model research and compute platforms continue to work in sync.
Therefore, the $8.2 billion deal is both an investment by AMD in Fei-Fei Li's team and a bet on future workloads. The acquisition still needs to be completed, and spatial intelligence still has a road to travel before large-scale real-world applications. But the chip giant clearly does not want to wait until needs mature before starting to figure out what models require.