刘洺堉 on 150. 对英伟达研究副总裁的4小时访谈：Cosmos

1. Key Themes

The Physical AI Opportunity Is the Next CUDA Moment

Liu Mingyu frames physical AI not as a competitive battle but as a market-creation exercise, explicitly comparing it to CUDA's role in building the AI market. NVIDIA's bet is that every robot will need compute, multiplying global demand for GPUs by orders of magnitude.

"现在家家户户都没有机器人，但如果有机器人的话，我家可能会买两三个，那每个机器人都需要算力，那等于整个世界对需要对算力的需求会大幅的成长。这个事情如果发生的话，对NVIDIA是巨大好处。" ("Right now households don't have robots, but if they did, I might buy two or three myself. Every robot needs compute, which means global demand for compute will grow enormously. If this happens, it's a huge advantage for NVIDIA.") 03:33:41

World Models Are Infrastructure, Not Products — The Foundation Model Strategy

Cosmos is explicitly positioned as a "World Foundation Model" — a base layer others build on, not an end product. Liu's open-source strategy is designed to make NVIDIA indispensable to the entire physical AI ecosystem rather than compete directly with any single application builder.

"我们目标是帮助这些solution的公司成功。我们的目标是建立一个Foundation让大家都可以建筑设他们要用的世界模型，所以我们把自己叫Foundation Model。" ("Our goal is to help solution companies succeed. Our goal is to build a Foundation so everyone can build the world model they need on top of it — that's why we call ourselves a Foundation Model.") 01:49:04

Model Capabilities Will Converge — Differentiation Comes from Ecosystem Integration

Liu makes a contrarian but well-reasoned claim: no single AI company will maintain a permanent lead. He predicts model capabilities will commoditize and the real competition will shift to ecosystem integration — analogous to how all software eventually commoditized but companies differentiated through ecosystem lock-in.

"历史上，从来没有这样发生过，就是一家公司就是掌握所有的客气，其他的公司都没有办法去追上，这从来没有在世界上发生过，然后我不觉得AI会是一个例外。" ("Historically, it has never happened that one company monopolizes everything and others can't catch up — this has never happened in the world, and I don't think AI will be an exception.") 01:17:57

Scalability Is the Only Metric That Matters in Research Direction

A recurring theme across Liu's 20-year career is his ruthless prioritization of scalable approaches over elegant mathematics. He abandoned GANs despite years of investment precisely because experiments kept confirming they weren't scalable — and pivoted to diffusion. The same logic now governs Cosmos architecture decisions.

"我一直重视就是，不能只靠模型，要跟公司的其他的ecosystem结合在一起。" ("I always believed you can't rely on the model alone — you have to integrate with the broader ecosystem.") 01:15:20

"不scalable的东西最后做不远。" ("Things that don't scale ultimately can't go far.") 00:59:29

Research in Industry Requires Selling Internally — Communication Is as Important as the Science

Liu identifies three career-defining lessons: (1) proximity to real-world applications over mathematical elegance, (2) communicating research value to non-technical leadership, and (3) collaborative teamwork at scale. He explicitly says he "scored 100 on execution but 10 on explanation" in his early career — and learned that the 10 was what determined resource allocation.

"我常常讲做我自己做了100分，但是我的解释只有10分，所以别人看到只有10分。" ("I often say I did the work at 100 points, but my explanation was only 10 points, so others only saw 10 points.") 01:45:07

Jensen Huang's "30 Days to Bankruptcy" Mindset as a Management Philosophy

Liu reveals a striking and consistent internal practice: Jensen tells his team every single day that the company only has 30 days of cash and will go bankrupt after that. This is a deliberate psychological tool to force prioritization.

"他每天都觉得公司只能下30天的现金流，30天后就要破产。" ("He believes every day that the company only has 30 days of cash flow left, and after 30 days it will go bankrupt.") 02:16:09 / 02:50:57

Open Source as a Strategic Moat, Not Altruism

NVIDIA's decision to open-source Cosmos is strategic: by publishing training frameworks, synthetic data, and model weights, they accelerate the entire physical AI ecosystem, which drives more GPU demand. Liu explicitly notes they're not worried about competitors replicating Cosmos because the open ecosystem growing is better for NVIDIA than protecting proprietary advantage.

"我不担心别人可以说出一模一样的东西。" ("I'm not worried that others can produce something identical.") 01:59:59

Physical AI Requires a Different Architecture Than Digital AI — Not Just a Bigger LLM

Liu pushes back on the assumption that scaling LLMs is the right path for physical AI. He argues that robots need action as a "first-class citizen" in model architecture, and that the intelligence requirements for physical tasks (picking up objects, factory assembly) are qualitatively different from coding or chat.

"你需要你的机器人会决凹数问题吗？你需要你的机器人会解这个Coding的问题吗？我觉得你可能更关注他能不能把你的这个工厂要组装的东西做好。" ("Do you need your robot to solve math olympiad problems? Do you need it to write code? I think you care more about whether it can properly assemble what your factory needs assembled.") 02:19:02

The "No Layoffs" Culture as a Compounding Strategic Advantage

NVIDIA has never had mass layoffs in Liu's 10-year tenure. He explains this as a deliberate choice to build trust, enable honest communication, and promote collaboration — trading short-term optimization for long-term organizational health. He notes this creates a slower transition speed when technologies shift, but the trust dividend is worth it.

"当一个员工对这个公司有认同感有安全感，他愿意讲出一些，就是，可能在一般担心被淘汰的地方不愿意讲的事情。" ("When an employee has a sense of belonging and security, they're willing to say things they wouldn't say in a place where they fear being eliminated.") 02:53:03

Generalization (泛化) Is the Core Unsolved Problem of Physical AI

Liu repeatedly returns to generalization as the fundamental bottleneck — specifically, learning from limited demonstrations and applying reliably to novel scenarios. This is the definition of the "ChatGPT Moment for Physical AI" he is targeting.

"AI的核心问题就是解泛化的问题。" ("The core problem of AI is solving the generalization problem.") 03:23:47

"我觉得，假设我今天教机器人说，我的衣服我喜欢这样折，然后喜欢这样放，我就给这个机器人看一两次我怎么做，他是不是就可以快速学会。" ("Imagine I teach a robot how I like my clothes folded — I show it once or twice, and it learns immediately. That's the moment I'm waiting for.") 02:15:42

2. Contrarian Perspectives

The ChatGPT Moment for Physical AI Will Come From Generalization, Not Hardware Demos

While the industry focuses on impressive robot hardware demos, Liu argues the real ChatGPT moment will be defined by a software/AI capability: one-shot or few-shot generalization — showing a robot once how to do a task, and it transfers reliably to new scenarios. This reframes where investors should look for the breakthrough signal.

"我个人觉得这种，就是简单的人示范一下怎么要的操作，这个机械就可以就地学会这东西，我觉得这是一个如果这东西出现，我就觉得说，yeah, That's a ChatGPT Moment for Physical AI。" ("I personally think: a person simply demonstrates an operation once, and the machine immediately learns it on the spot — if that appears, I'd say, yeah, that's a ChatGPT Moment for Physical AI.") 03:23:45

GANs Were Always a Dead End — And Everyone Building on Them Was Wasting Years

Liu invested heavily in GANs throughout his career and built NVIDIA's early AI reputation on them. But his own experiments, including with interns like Yu-Ji Hui and Jeremy Burstyn, consistently showed GANs don't scale. He eventually told his entire team to abandon the approach mid-stream — a genuinely costly and unpopular decision. Most of the field ignored this signal for years.

"到某一年，很努力啊，然后跟着这么多聪明人在研究这个，怎么让GAN更加scalable，但每个结果都走向都是得到一个结论，就是GAN不scalable。" ("After all those years, working with so many brilliant people on how to make GANs more scalable, every experiment led to the same conclusion: GANs don't scale.") 00:58:39

Sora Failed Commercially Not Because Video AI Is Wrong, But Because of Capital Allocation Rationality

Against the common narrative that Sora was a technical failure or that video generation is unimportant, Liu argues OpenAI deliberately de-prioritized it under competitive pressure from Anthropic's coding agent revenue — which generates vastly more tokens (and thus revenue) per session than conversational chat.

"你在跟ChatGPT聊天的时候，你产生的token数就是，可能最多一分钟就生成完了，那你做这个Coding Agent就是可以工作24小时48小时，这个token量非常的大。他们就是觉得这是一个很好的方向，那就集中公司的精力，往这个Coding的方向走。" ("When you chat with ChatGPT, tokens are generated in maybe a minute. But a Coding Agent can work for 24 to 48 hours straight — the token volume is enormous. They saw that as a great direction and concentrated resources there.") 02:20:53

China's Robotics Advantage Is Manufacturing Integration, Not AI — And This Is Structural

Liu observes that American robotics companies are very few (Tesla, Figure AI, a handful more) while China has a large ecosystem. The edge isn't model quality — it's that China's manufacturing infrastructure allows rapid hardware-software iteration that is structurally difficult to replicate in the US.

"因为中国强大的制造业啊，让产生一个很优势的条件，可以让这些公司做一些在美国不容易做到的事情。" ("Because of China's powerful manufacturing industry, it creates an advantageous condition allowing these companies to do things that are difficult to do in America.") 03:22:27

American Frontier Labs Have Stopped Training Interns — And This Will Cost Them Knowledge Diffusion

Liu makes an observation that most people miss: top US frontier labs (OpenAI, Anthropic) have largely stopped accepting research interns, concentrating know-how internally. Meanwhile, Chinese AI companies continue to run large internship programs. This creates an asymmetric diffusion of knowledge that will compound over time in China's favor.

"在美国啊，美国啊，尤其是些frontier lab，现在的状态就是，他们很少去遭受实习生，很多那个know-how，啊，很多这些科研的结果都只有在他们自己本身内部的实验室有。那在中国这边我看到就是，他们像这些公司都遭受很多实验室，然后帮助把这些就是know-how传递出去。" ("In America, especially at frontier labs, the current situation is they rarely take interns — a lot of know-how and research results stay only inside their own internal labs. In China I see the opposite: these companies take many interns, helping spread that know-how outward.") 03:15:17

3. Companies Identified

NVIDIA Cosmos Lab

NVIDIA's world foundation model project for physical AI, spanning ~200-300 people. Progressed from Cosmos 1 through Cosmos 3 in under 18 months, now an omni-modal model combining video, audio, language, and action signals in a dual-tower architecture. Open-sourced including training frameworks and synthetic data. Jensen Huang told Liu to build it to "Cosmos 97."

"我们把这个叫做World Foundation Model，嗯，我们目标是建立一个Foundation让大家都可以建筑设他们要用的世界模型。" ("We call it the World Foundation Model. Our goal is to build a Foundation so everyone can construct the world model they need.") 01:49:04

DeepSeek

Chinese AI company singled out by Liu as an extraordinary benchmark for iteration speed — three major model versions in one year (culminating in DeepSeek V3). Liu says DeepSeek directly influenced his own push to make Cosmos iterate faster.

"那时候深受DeepSeek的影响，因为DeepSeek是很惊人的公司，它可以一年内迭代三次，产生DeepSeek V3。" ("At the time I was deeply influenced by DeepSeek — it's a remarkable company, iterating three times in one year to produce DeepSeek V3.") 01:52:20

Anthropic

Cited as the canonical example of focused differentiation beating a larger, better-resourced rival (OpenAI). Liu specifically credits Anthropic's early, concentrated bet on coding agents — while OpenAI diversified — as the key strategic move that created their breakthrough.

"Anthropic是做击中式投资，然后最早去做出这个Coding这个东西的。所以他那时候就看到了一些，那个就是，机会，然后他又坚信，就决心，把它做出来。" ("Anthropic made concentrated investments, and was the first to develop the Coding product. They saw the opportunity, believed in it, and had the determination to execute.") 01:20:51

Getty Images

One of NVIDIA's earliest large model collaboration partners — they provided data while NVIDIA provided compute, jointly training text-to-image models under the "Picasso" project (NVIDIA's AI Foundry for Media Generation). This was Liu's entry point into large model development.

"我们就坦诚的这个合作方案，就是他们提供数据，然后我们有算力，然后我们去view这个大模型来做这个Test to Image。这就是我踏入大模型的时间点。" ("We agreed on a partnership: they provided data, we had compute, and together we trained a large model for text-to-image. This was my entry point into large model development.") 01:12:39

Mitsubishi Electric Research Lab (MERL)

Liu's first US employer, based in Cambridge next to MIT. Was once a top-tier research institution rivaling Microsoft Research in the 1990s; spawned many MIT professors (including Bill Freeman, Ramesh Raskar). Liu witnessed its decline firsthand — when the company questioned "blue sky research" ROI, they began firing top researchers, many of whom joined academia. Liu explicitly says watching this shaped his understanding of how to keep research aligned with business value.

"三菱电机在90年代非常的重要，因为它是早期这种就是没有任何限制的这个地方来做research，包含领域里面很多著名的人，像有Bill Freeman那时候在三菱电机研究院。" ("MERL was extremely important in the 1990s — an unconstrained research environment. Many famous people in the field, like Bill Freeman, were there.") 01:10:33

ByteDance (Jianying / CapCut)

Named as currently the most profitable video generation company — cited as evidence that the video generation market has viable commercial models and that first-mover advantage in model release doesn't determine commercial success.

"现在做video生成最赚钱的应该是字节了，对，Jianying，对吧。" ("The most profitable video generation company right now should be ByteDance — Jianying, right?") 01:42:17

World Labs (Fei-Fei Li's company)

Invested in by NVIDIA. Working on world models with a different approach from Cosmos. Liu notes they are well-capitalized and pursuing their own path; NVIDIA supports them as part of enabling the broader ecosystem.

"菲菲的World Lab啊，杨雷坤的这个MI，都是不一样，走不一样的路线。我们都帮助他们。" ("Fei-Fei's World Labs and Yann LeCun's AMI are walking different paths. We support all of them.") 01:50:11

AMI Labs (Yann LeCun's company)

Also invested in by NVIDIA. Pursuing a different architecture philosophy for world models. Liu mentions it as part of NVIDIA's ecosystem-enabling approach rather than picking winners.

"投资他们，我们跟他们也有合作。" ("We invested in them, we also collaborate with them.") 01:50:12

Kimi (Moonshot AI)

Cited by Liu as an example of a "late entrant" to the large model race that still found success — countering the view that the window for new model companies has closed.

"像Kimi，我是，Memo这些affer都是算比较晚期出来的嘛，还是有机会的嘛。" ("Companies like Kimi and MiniMax came out relatively late, and they still have opportunities.") 01:19:54

MiniMax

Similarly cited alongside Kimi as a later-stage entrant proving that new model companies can still compete. Liu rates Chinese model companies — including DeepSeek, Qwen, Doubao, Kimi, MiniMax, and others — as "very impressive."

"DeepSeek开始啊，签问啊，豆包啊，MiniMax，Kimi，这些都做得很好。" ("Starting with DeepSeek, then Qwen, Doubao, MiniMax, Kimi — they're all doing very well.") 03:14:31

4. People Identified

Jensen Huang (黄仁勋)

NVIDIA CEO. Liu describes him as the single biggest influence on his career. Key traits: reads enormous volumes of email daily to find weak signals; makes decisions purely on first principles without personal sentiment; has operated with a "30 days to bankruptcy" daily mindset for decades; personally names major products (named Cosmos and GauGAN); committed to never conducting mass layoffs; told Liu to build Cosmos all the way to "version 97."

"他是一个非常热爱学习一直去学习，一直去理解，他当初也不懂computer graphics，他敢踏入做GPU，他去把computer graphics学起来，他也不懂deep learning，也踏进AI，就懂deep learning。" ("He has an enormous passion for learning. He initially didn't understand computer graphics, but dared to enter GPU, and learned it. He didn't understand deep learning either, but entered AI and learned it.") 01:28:25

"Are you a quiet baby?" 03:09:12 — What Jensen told Liu when he tried to excuse underperformance by calling a benchmark "unfair."

Tero Karras

NVIDIA Research scientist, inventor of StyleGAN. Liu identifies him as one of his deepest influences for his extreme focus — publishing one paper per year, spending all his intellectual energy on a single idea rather than optimizing for paper count. Described as the model of "staying focused."

"Tero就是每年之相写一篇paper，那就是把他所有的，他也工作也非常努力，就是把所有的就是人生的经历啊，就投注在同一个idea，越转越细，找到问题的本质，然后产生重大的贡献。" ("Tero would write only one paper per year, pouring all of his life's effort into a single idea, drilling deeper and deeper, finding the essence of the problem, and making a major contribution.") 01:28:02

Yu-Ji Hui (虞佳辉)

Trained under Liu as an intern at NVIDIA, now at Meta. He was one of the early contributors to the foundational research on Sora at OpenAI. Liu recalls Hui proposing the idea of placing a separate discriminator on each GPU during GAN training — an ahead-of-its-time distributed training insight.

"包含现在在Meta的瑜伽灰，做出这个Sora的Tim Brooks，哦，都在我这里实习。" ("Including Yu-Ji Hui, now at Meta, and Tim Brooks who made Sora — both interned with me.") 01:44:53

Tim Brooks

Former Liu intern at NVIDIA, went on to OpenAI where he was a first author on Sora. Liu recalls Brooks arriving as an intern already laser-focused on a specific generative model direction. Liu describes watching a former intern produce Sora as bittersweet — proud, but a moment of self-reflection on his own missed timing.

"我还记得Tim Brooks刚加入我团队做实习的时候，他就一直想做这个..." ("I still remember when Tim Brooks first joined my team as an intern — he was always focused on doing this...") 01:54:54

Jeremy Burstyn

Former Liu intern, later invented the Muon optimizer (now widely used in training). Liu worked with Burstyn on methods to stabilize GAN training; Burstyn developed the early algorithmic ideas that eventually became Muon from those experiments.

"Jeremy就是从那时候开始，我们写了一系列的paper，就是如何用Muon早期的算法来去稳定，来去就是让这个训练更稳定，他就一路一路往下走，后来开发这个Muon的算法。" ("Jeremy started from that time — we wrote a series of papers on how to use early Muon-style algorithms to stabilize training. He kept going down that path and eventually developed the Muon optimizer.") 01:57:53

Rama Chellappa

Liu's PhD advisor at the University of Maryland. A pioneer of early computer vision and AI. Liu credits him with teaching him the generative modeling philosophy and the communication skills to make any meeting "full of laughter" while still being rigorous.

"第一个就是我的制造教授哦，Rama Chellappa，他把我带进这个地本能力谁行，他非常厉害，然后他也是非常善于沟通。" ("First is my PhD advisor Rama Chellappa — he brought me into this field. He's extremely capable and also exceptionally good at communication.") 01:26:46

Ian Goodfellow

Inventor of GANs. Liu saw Goodfellow's GAN paper at NeurIPS and credits it as the inflection point that pulled him toward generative models. They became close collaborators and friends through overlapping work on GAN research.

"也因为这样子跟Ian Goodfellow就game的那个发明的人成为好朋友，然后我们就那个就很多讨论。" ("This also led to becoming good friends with Ian Goodfellow, the inventor of GANs — we had many discussions.") 01:09:09

Fei-Fei Li (李飞飞)

Founder of World Labs (invested in by NVIDIA). Mentioned as one of the well-capitalized world model builders working on a different approach from Cosmos.

"菲菲的World Lab。" ("Fei-Fei's World Labs.") 01:50:11

Yann LeCun (杨雷坤)

Chief AI Scientist at Meta, founder of AMI Labs (invested in by NVIDIA). Cited alongside Fei-Fei Li as a well-resourced world model researcher pursuing a distinct path from Cosmos.

"杨雷坤的这个MI。" ("Yann LeCun's AMI.") 01:50:11

Bill Daly

NVIDIA's Chief Scientist and former head of NVIDIA Research. Liu reported to him during his Deep Imagination Research period. Liu credits Daly with being the person who heard his original job talk pitch about deep learning replacing computer graphics for visual generation.

"之前是汇报给Bill Dally。" ("Previously I reported to Bill Daly.") 01:28:31

Xu Ji (徐济)

Liu's martial arts teacher in the US. A well-known martial arts instructor whose own teacher was Liu Yunqiao, who traces lineage to the famous Cangzhou martial arts tradition (八极拳/劈挂掌). Liu has practiced Chinese martial arts since age 18 under Xu Ji's instruction.

"我的武术老师是徐济，他是一个很有名的武术老师，他的老师是刘云桥，刘云桥老师是李书文。" ("My martial arts teacher is Xu Ji, a very well-known instructor. His teacher was Liu Yunqiao, and Liu Yunqiao's teacher was Li Shuwen.") 01:33:35

5. Operating Insights

Mission Is the Boss — Replace Hierarchical Reporting with Mission Alignment

Liu describes NVIDIA's organizational principle as "Mission is the boss" — when NVIDIA pivots to a new direction, people across unrelated teams self-organize and contribute without being told. This removes organizational drag and allows a 200-300 person project like Cosmos to absorb contributors from across the company (e.g., Nemotron team members co-authoring Cosmos 3 papers) without formal reporting changes.

"Mission is the boss，而不是你的boss是你的boss。当我们公司今天要走向这个方向的时候，在各种不同组织的团队，但有相对的技能就会自动align，就往那个框架架进，就是像一个活的组织。" ("Mission is the boss, not your boss is your boss. When the company moves in a direction, teams across different organizations with relevant skills automatically align and contribute — like a living organism.") 03:00:43

Pre-Visualize the Outcome Before Execution — Tell the Team What the Model Will Look Like Before Building It

Liu's method for aligning a 200+ person team: before Cosmos 3 shipped, he had already described to key leaders and most of the team exactly what the finished model would look like. Even if reality differed 10%, having a shared vivid target dramatically improved individual decision-making without requiring Liu to be in every decision.

"在Cosmo 3 paper出来之前或做完之前，我就跟主要的几个leader还有大部分的团队都已经讲过，说我们做出来模型会是这样子。当大家都看到同一个愿景的时候，就比较容易集中。就像你在罚球之前，你都会想象这个球空心落网。" ("Before Cosmos 3 was finished, I had already told the key leaders and most of the team what the model would look like. When everyone sees the same vision, it's easier to align. It's like imagining the ball going cleanly through the net before you shoot a free throw.") 02:34:17

The LV Principle: Fewer Products Sell More Than More Products

When Liu told Jensen he needed 22 separate models to serve all use cases, Jensen responded with the Louis Vuitton example: LV increased sales by reducing the number of products on display, because too many options confuse buyers and prevent decisions. This led directly to the Cosmos 3 single omni-model architecture.

"Louis Vuitton LV这家店呢，当他减少他的货品，成列的这个货品数字之后反而Sale是提升的。你不希望太多的东西困惑你的那个消费者，让他没办法做出决定。" ("LV: when they reduced the number of products on display, sales actually increased. You don't want too many things confusing your consumer and preventing them from making a decision.") 01:55:30

Lazy Computation as a Life Principle — Don't Compute What You Don't Need To

Liu draws a direct analogy from computer science's core optimization principle (lazy evaluation) to personal productivity and management: ruthlessly eliminate tasks that don't need to be done, defer what can be deferred, ignore what can be ignored. The scarce resource is your cognitive budget, not your GPU budget.

"Computer science最重要的一件事情就是懒惰，就是不用算的东西不要算，可以晚点算的东西晚点算，可以直接就是ignore就全部ignore。那你人生就是这么多时间，你经历就这么多，如果你能知道什么东西不要算，就把它益出的话，你的人生运用就更有效率。" ("The most important thing in computer science is laziness: don't compute what you don't need to, defer what can wait, and ignore everything else. You only have so much time and energy in life — if you know what not to compute and can eliminate it, your life becomes far more efficient.") 01:31:49

Build Fault Tolerance Into Team Culture to Enable Decentralized Decision-Making

Liu designed Cosmos's operating model so individual contributors can make consequential decisions — but only by also building explicit psychological safety for being wrong. Without the safety valve, people refuse to decide at all, which negates the decentralization benefit.

"当一个人害怕他决定错误的时候，他是不会做决定的。所以你要让他们可以做决定，但也要接受，有些决定会错误，但是错误的时候要能快速修正。" ("When a person fears making a wrong decision, they won't make decisions at all. So you need to let them decide — but also accept that some decisions will be wrong, and be able to correct them quickly.") 02:33:21

6. Overlooked Insights

The Muon Optimizer's Origin Story — Born From Failed GAN Stabilization Research

Jeremy Burstyn came to Liu's team to research how to stabilize GAN training using novel optimization methods. The specific experiments didn't succeed at their stated goal, but the algorithmic work Burstyn did in Liu's lab became the seed for the Muon optimizer — now one of the most widely used training optimizers in the field. This is a remarkable example of how "failed" research in industry labs produces disproportionate downstream impact. No one in the podcast paused to note the significance: one of the field's most important training tools came out of a failed side project at NVIDIA Research under Liu's supervision.

"Jeremy Burstyn，他就是后来发明这个Muon Optimizer，现在大家都在用这个Muon Optimizer，然后他那时候来我这边实验，那我跟大家研究的方向是如何在在这种out weighting上面更好的算法来去把这个GAN给做的稳定。这两个最后就是方向最后都没有成功，但他们各自都开发出了非常惊人的，很有成就的工作。" ("Jeremy Burstyn later invented the Muon Optimizer, which everyone uses now. He came to my team, and we researched better optimization algorithms to stabilize GAN training. Neither direction ultimately succeeded — but each of them individually went on to develop astonishing, highly impactful work.") 01:57:23

NVIDIA's Top-5 Email System Is a Real-Time Organizational Radar for Duplicate Work

Liu briefly mentions — and immediately moves past — a specific NVIDIA management practice: every employee sends a "Top 5" email to Jensen summarizing their five most important current activities. The non-obvious insight Liu drops without elaboration: this system automatically surfaces when two teams are working on the same thing, because their Top 5 lists overlap when aggregated. This functions as a distributed, low-overhead organizational intelligence system that would normally require expensive management layers or costly coordination failures to detect. For any company dealing with siloed duplicate efforts, this is an immediately deployable mechanism.

"Top 5员工就是要跟很多人就是讲，这个他做的事情最重要的五件事情，嗯，这件事情让你看清楚，就是各种公司的大小讯号，可以看到很多两个团队在做一模一样的东西啊，那个Top 5 Email就会自然显现出来。" ("The Top 5 — employees tell many people the five most important things they're doing. This lets you see all kinds of large and small signals in the company. You can see when two teams are doing exactly the same thing — the Top 5 emails naturally reveal it.") 02:59:01