Podcast Atlas
source Episode summary Updated 2026-08-13 Tags: Podcast, Nvidia, World-Models, Physical-Ai, Robotics, Ai-Research

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

Summary

This 张小珺Jùn｜商业访谈录 episode interviews [[LiuMingyu|Liu Ming-Yu / 刘洺堉]], Nvidia research VP and [[CosmosLab|Cosmos Lab]] lead, about Cosmos 3, World Models, and Physical AI. Liu frames Cosmos as a [[WorldFoundationModels|World Foundation Model]] for physical-world AI developers, not as a narrow content-generation product or a direct attempt to compete with Nvidia customers. The interview also links his research path from vision and generative models into large-scale project leadership, while using Jensen Huang’s influence to explain Nvidia’s first-principles, long-horizon, ecosystem-first operating culture.

Key Claims

  • [[LiuMingyu|Liu Ming-Yu / 刘洺堉]] leads [[CosmosLab|Cosmos Lab]] at Nvidia, with a direct team of roughly 80-90 people and a broader Cosmos effort involving about 200-300 people.
  • Liu says Cosmos 3 grew from earlier image/video generation work, the Picasso/AI-foundry effort, and the post-Sora judgment that World Models should become a major Physical AI infrastructure direction.
  • Nvidia is presented as doing research and open model releases to understand what future developers need, inform long-cycle hardware and software decisions, and reduce pressure on the Physical AI ecosystem.
  • Liu avoids framing Cosmos as a fight to beat every model competitor; the source says Nvidia benefits if Physical AI succeeds because the whole market will need more compute and platform infrastructure.
  • The source defines the useful world-model target less as generic video generation and more as a [[WorldFoundationModels|World Foundation Model]] that gives developers better data, better starting points, and better environments.
  • Cosmos 3 consolidates earlier predict, transfer, reason, and policy directions into a single Omni-Model, combining language, video, audio, and action for physical agents.
  • Liu says action matters because a Physical AI agent changes the world, making action-conditioned prediction and post-training flexibility central to the model’s practical value.
  • The episode’s robotics bottleneck is generalization: robots need to learn new operations from limited demonstrations, manuals, or observation-action examples rather than only repeat trained scenes.
  • Liu separates world-model evaluation into benchmarks, arena-style comparisons, and customer-pain evaluation, while treating customer-pain evaluation as especially important for Physical AI.
  • The source links Cosmos data strategy to navigation/manipulation differences and to egocentric data, including human-eye viewpoints and hand-operation footage.
  • Liu’s research-management transition is framed as moving from papers and models toward explanation, customer success, large compute responsibility, and AI Organization Design.
  • Jensen Huang is portrayed as influencing Liu through first-principles reasoning, prioritization, long-term investment in CUDA and deep learning, and the demand to keep building toward “Cosmos 97.”
  • The source’s company-culture thesis is that mission, low ego, trust, and ecosystem success can matter more than personally defeating every opponent.

Key Quotes

“做到 Cosmos 97” — Liu’s account of Huang’s answer after Cosmos 1.

“Mission is the Boss” — Liu’s summary of Nvidia’s collaboration norm.

“Are you a crying baby?” — Huang’s criticism in Liu’s account of a proposal comparison.

Connections

Contradictions

  • No direct contradiction found.
  • Source-internal naming note: the file title and metadata use 刘洺堉, while the body renders the guest’s Chinese name as 劉明玉. This ingest treats them as the same Nvidia researcher and uses the title/metadata form.