Liu Ming-Yu / 刘洺堉
Liu Ming-Yu / 刘洺堉 is the Nvidia research VP interviewed in 150. 对英伟达研究副总裁刘洺堉的4小时访谈：Cosmos 3、世界模型、武术、黄仁勋影响我的，和你不需要击败所有对手. The source title uses 刘洺堉, while the body renders the name as 劉明玉; the wiki keeps the source-title form for disambiguation. In the episode, Liu leads [[CosmosLab|Cosmos Lab]] and explains Cosmos 3 as a [[WorldFoundationModels|World Foundation Model]] project for Physical AI developers.
The source frames Liu’s career as a move from wireless communication and traditional vision into generative models, then into productized world-model infrastructure. He studied at [[UniversityOfMaryland|University of Maryland]], encountered AI and computer vision through an Intel internship, worked on vision and generative research, and joined Nvidia partly because deep learning needed more compute than his prior research environment could provide.
His operating identity in the interview is a researcher-manager hybrid. Liu says research depends on asking the right problem, but large model projects also require storytelling, engineering proof, customer feedback, compute responsibility, and enough AI Organization Design to keep hundreds of contributors aligned without removing local judgment.
Key Claims
- Liu treats World Models as useful only when the term is tied to a concrete physical-world role, which is why he prefers the World Foundation Models label for Cosmos.
- His shift from GAN-style work toward diffusion and then Cosmos is presented as a repeated move toward simpler, more scalable model families and clearer user value.
- He sees Cosmos 3 as part of Nvidia’s ecosystem strategy: if Physical AI developers move faster, Nvidia’s platform and compute demand can grow with them.
- He describes Jensen Huang’s influence through first-principles reasoning, long-termism, prioritization, and the ability to challenge people without turning the work into personal rivalry.
- His personal lesson from the episode is not that a model team must defeat everyone, but that it should be capable of building No. 1-level systems while choosing problems with the largest contribution.
Connections
- Nvidia, Jensen Huang, CUDA, and AI Infrastructure Full-Stack Moat — company, CEO influence, and infrastructure context.
- Cosmos Lab, Cosmos 3, World Foundation Models, and World Models — project and model branch.
- Physical AI, Robot Data Scale Up, Robot Generalization Performance Tradeoff, and Robotics Simulation Evaluation — physical-world AI problem field.
- Research Taste, Problem Definition In Research, and AI Organization Design — research and management themes in the interview.