Fei-Fei Li and World Labs Are Welcomed In by AMD; the Comments First Ask Whether ROCm Can Run 3D
AMD CEO Lisa Su publicly welcomed Fei-Fei Li and the World Labs team to AMD, and Fei-Fei Li responded that the entire team is incredibly excited. Collaboration details were not disclosed; the tech community is first focusing on how world models perform on real 3D workloads on AMD hardware and ROCm.
On November 16, AMD CEO Lisa Su posted on X welcoming Fei-Fei Li and the World Labs team to AMD. She wrote that she is very much looking forward to what they will accomplish together. Fei-Fei Li then replied, thanking Lisa Su and saying that the entire World Labs team is beyond excited; there is still a lot of work ahead, with the goal of creating technology that benefits everyone.
The tweet included this image:
The news itself does not contain much information. The form, scope, and timeline of the collaboration were not disclosed. What can be confirmed is that AMD has publicly placed Fei-Fei Li and World Labs into its own narrative. World Labs was founded by Fei-Fei Li and focuses on spatial intelligence and world models. AMD has long faced Nvidia in the AI chip market; beyond hardware, the ROCm software stack has always been a focus of developer discussion.
The reaction in the comments is more informative than the congratulations. One commenter said that what they most want to see is benchmarks of world models on AMD hardware, and that ROCm has to withstand real 3D workloads. Another commenter spelled out the difference in workloads: spatial models do not just do token inference; they involve a large amount of 3D generation. If the ROCm and MI400 stack can run well at scale, AMD will have a chance to open a second front beyond training. Another commenter directly pointed out that compute is the bottleneck for spatial intelligence.
These are commenters' judgments, not a roadmap disclosed by AMD. But they point to the same question: what kind of compute do world models need, and can AMD's software stack let developers spend less time accommodating the toolchain? Text-model inference already has a mature path. 3D generation, spatial understanding, and simulation have different requirements for VRAM, bandwidth, and compilers. The areas where ROCm has been criticized in the past will be exposed even more clearly under these workloads.
Half of the comments are congratulating, and half are already asking about ROCm. The tech community's idea of romance is very concrete.
So this photo of them together is a signal, not a conclusion. AMD has brought in one of the most recognizable teams in the world-model field. What to watch next is not repost counts but benchmarks, toolchain maturity, and developer migration costs. The group photo is the opening; the benchmark scores are the main story.
Source tweets: Fei-Fei Li's reply, Lisa Su's welcome tweet.
发布时间: 2026-11-17 05:33