对话脉络

这是 Y Combinator 在 Startup School 2026 的对谈，主持人 Garry Tan，嘉宾是英伟达创始人兼 CEO 黄仁勋。话题顺序大致是：公司起步时选错技术 → 靠教科书自救 → 与 Sega 的那笔救命钱 → AlexNet 带来的转向 → 组织如何保持第一性原理 → 系统思维为何取代写代码 → 智能体与可控性 → AI 对就业的影响 → 具身智能与机器人 → 给年轻人的建议。

访谈里最反直觉的一点是开场的坦白：英伟达不是靠「一开始就看对了方向」成功的，它起步时选的算法和技术完全错误，是承认错误之后现学出来的。

时间轴

起步：选错技术，然后现学

1993 年，PC 刚刚开始普及。黄仁勋和两位合伙人成立英伟达时的设想是：把每一台个人电脑变成游戏机，用加速器扩展 CPU 的能力。

这个「把 CPU 加强」的方向后来被证明是对的；但当时选的算法和技术实现，完全错误。到 1995 年，公司意识到这一点时，市面上已有三四十家公司在做 PC 三维图形，而团队不仅选错了路，也不知道正确的做法是什么。

他回到公司讲的话是：如果不承认这件事不成立，公司就不存在了。

接下来的场景是这场访谈里最有画面感的一段：他带着口袋里仅有的几百美元，去 Fry's 买了三本关于 OpenGL、关于如何设计图形管线的教科书，带回公司交给工程师。此后英伟达重做了计算机图形，并在过去二十五年里拿出了这个领域的大部分重大突破。

他从中提炼的结论是：技术本身一直在变，所以技术选择对不对并不致命；能否面对现实、能否把新技术学会，才是分水岭。

而真正选对的那个「大想法」是：用加速器去算 CPU 无法胜任的问题。分子动力学、图像处理、逆物理、深度学习都属于这一类。公司因此不是「做一颗好芯片」，而是「加速一个算法领域」。

他对好公司的定义也随之而来：不是技术，不是市场，而是对世界的一个你深信不疑的独特判断，再加上实现它足够困难。技术、市场、时机都对时，日子会好过很多，但那些都是副产品。

Sega：承认做不到，反而拿到钱

让公司发现算法错误的那单生意来自 Sega。Sega 委托他们开发 Saturn 之后的主机（后来成了 Dreamcast）。

黄仁勋飞到日本，直接告诉当时的 CEO：这项技术根本行不通，我们无法履约，建议你们找别人做。但他同时说了一句实话：这笔钱我仍然需要。

对方的反应被他概括成一句反问：你是说，我委托你做的事你做不了，但你还想要全部的钱？他说，正是如此。

他补充了当时的姿态：礼貌、谦逊、诚实。对方判断这是可信的，最终给了 500 万美元，让公司撑到了找到正确方向的时候。他的总结是那句常见的投资格言——投的不是公司，是人。

后面的事：英伟达 1999 年上市，估值 3 亿美元。

AlexNet：从「一个模型」看成「一种方法」

很多人看到 AlexNet 看到的是一个识别准确率很高的模型。黄仁勋说他看到的是一个通用函数逼近器：只要能给出答案，它就能学会那个函数；而大部分有意思的问题本来就是不精确的，也不允许精确。

这个判断改变的不是一个产品，而是整条技术栈。他提到自己当时就在推演五层结构——处理器、中间件、算法、应用、以及最终的使用方式——并把计算机视觉、机器人、自动驾驶几乎立刻排上了日程。

这里有一个方法论值得单独拎出来：他看世界的「镜头」一直是算法，不是芯片。所以新东西出现时，他问的不是「这是什么产品」，而是：

• 这个算法是什么？
• 它为什么有效？
• 它还能解决什么？
• 如果把它规模化，能解决今天解决不了的问题吗？

组织：第一性原理与「造一辆自己能开的车」

主持人问：你读论文、直接找首席科学家聊，这种深入细节的方式如何与组织共存？

黄仁勋的回答不是管理技巧，而是一种性格：一切从好奇心开始。他想知道答案，会先自己去发现；当发现某个领域对公司重要时，他会尽可能学透，再把它讲给团队听——在他看来，CEO 是服务于公司和员工的，价值在于把洞察交到他们手里。

他还用冲浪解释为什么必须贴近一线：浪看起来是混乱的，冲浪者能读懂它，是因为他下过水。技术变化越快，越需要这种手感；只靠二手汇报，你会觉得它快得无法理解。

更著名的是 F1 比喻：创办人是在造一辆自己要去比赛的车，所以应该把车造成自己开得顺手的样子。有人问他，如果不用常规管理方法，他离开公司后怎么办？他的回答是——等我在任上去世，下一个 CEO 会按自己的样子重塑这家公司。

系统思维取代写代码

被问到最值得培养的能力时，他给的是系统思维：系统理解、系统设计、系统组织。

理由很直接：低层的工作正在被智能体接管。芯片设计里，晶体管、门电路早就自动综合了，设计师实际做的是系统设计；软件也一样，大部分编码会交给智能体。于是人要做的是抽象地思考系统——要解决什么问题、约束是什么、输入和输出在哪里、信息以什么速率进出、瓶颈在处理器还是内存还是网络。

他认为这类知识不会过时，只会越来越重要。

智能体：能力已经够，缺的是控制

他对智能体的判断有一个具体的落点：递归式自我改进其实已经在发生——每次使用，它都在更新长期记忆、改写 Markdown 文件、把记忆压缩成知识图谱。

但真正的瓶颈不是能力，而是细粒度的可控性。他举的例子很具体：在一个计划文件里改一个词，就应该只带来那一处精确的差异——一个像素、一个三角形、CAD 文件里的一个组件、一层布线、一个过孔——其余部分自动重新生成。

他也不要求智能体一次做对：80% 的水平，人补完剩下的；或者 99%，人补完最后一点，都可以接受。可控性比准确率更关键。

关于开源，他把 OpenClaw 类比为 Linux 时刻：这是将要承载大语言模型的操作系统，人人都应该能构建自己的 AI。英伟达的做法是把自己的工程师开放给这些项目。

就业：AI 消灭的是任务

主持人的问题是经济层面会发生什么。他的回答是：很多任务会被自动化，尤其是信息已经在你手上、只需给出回答的任务；每个岗位都会变，也会出现大量新岗位。

他给了三个反例来支撑「AI 消灭任务、不消灭岗位」：

• 编程这项任务被大幅自动化，但软件工程师岗位同比还在增长约 10%。
• 读片这项任务已被自动化，但放射科岗位近年在增长，约 20%——因为病人积压太多，能接诊更多人就需要更多护士和放射科医生。
• 律师助理岗位没有消失，反而增长很快，因为诉讼积压同样巨大。

他的解释是：一份工作有目的，目的是由许多任务组成的；任务可以被拿走，目的不会。任务自动化释放出的产能，会撞上本来消化不完的需求积压。

具身智能：从生成视频到机器人

让他确信机器人临近的，是生成式视频：如果能生成手指运动的视频、生成一只手拿起杯子的视频，为什么不能让机器人做同样的动作？

由此引出的核心问题是：机器人如何生成遵守物理规律的运动——怎么理解因果、摩擦、张力。这就是他所说的物理 AI（physical AI），配套的是世界基础模型，以及三件事：real-to-sim、基于物理与生成的仿真、sim-to-real。

他认为机器人的「ChatGPT 时刻」其实两年前就到了：就像 ChatGPT 刚出现时并不产出实际价值，但它打开了想象空间；两年前机器人已经能通过强化学习、微调并在物理世界中落地。剩下要做的是环境、评测与后训练。

落地顺序上，他们判断第一个既有足够大市场、技术又相对标准化的应用是自动驾驶。英伟达的芯片与软件栈进了 Waymo、Tesla、Mercedes 的车与数据中心，他们把自动驾驶软件栈开源（Alpamayo），因为农业、邮政、仓储 AMR 都需要，而这些单个市场都不足以支撑一套自研栈。他给出的规模是：机器人与自动驾驶业务接近 100 亿美元，未来会成为下一个千亿美元业务，时间在 3 到 10 年之间。

给年轻人的建议

• 简单的东西会被自动化。 他用长除法作比：他那一代要学，现在没人手算了。坐在电脑前写代码这件事，也会被自动化。
• 难的东西不会消失。 物理、化学、生物、计算机科学、系统工程，尤其是学科交叉处；理解技术与社会议题的交界、发现市场空白，这些仍然存在。
• 系统思维越强越好。 能编排成千上万个智能体自主解决问题的人，会更有优势。
• 还是要去上学。 他的原话就是「stay in school」。
• 态度是「能有多难」。 他承认事实总是比想的难得多，但不要提前把难度想成焦虑；让困难一点点到来，每天只解决今天这一天。
• 韧性最重要。 你不需要一天解决一生，只需要过完今天，然后朝明天走。

理解检验

1. 英伟达起步时「技术完全是错的」，为什么公司还能活下来？ 参考答案：他们承认了技术不可行，并立刻转向学习正确的做法——买 OpenGL 教科书交给工程师。黄仁勋的结论是技术选择本身不致命，面对现实和持续学习的能力才关键。

2. Sega 那笔 500 万美元为什么重要？ 参考答案：公司当时已发现算法错误、无法履约。黄仁勋坦白说明并建议 Sega 另找合作方，同时说明自己仍需要这笔钱。对方基于对人和诚信的判断支付了款项，让公司撑到找到正确方向。

3. 他看到 AlexNet 时，看到的是什么？ 参考答案：不是单个模型，而是通用函数逼近器——只要给出答案，它就能学会对应函数；大部分有意思的问题本来就不要求精确。由此推演出处理器、中间件、算法、应用的整栈重构。

4. 「造一辆自己能开的车」这个比喻在说什么？ 参考答案：创始人应该按自己的性格与工作方式塑造组织，而不是套用通用管理模板；因为他才是要驾驶这辆车去比赛的人，继任者会按自己的方式重塑公司。

5. 为什么他认为系统思维比写代码更重要？ 参考答案：低层工作正在被智能体接管，芯片综合与编码都在自动化。人需要抽象地定义问题、约束、输入输出与瓶颈，这类系统层面的判断不会被自动化替代。

6. 智能体目前最缺的是什么？ 参考答案：细粒度的可控性。理想状态是在计划文件里改一个词，只产生精确的局部差异，其余自动重新生成；准确率 80% 也可以接受，可控性比准确率更关键。

7. 他如何论证「AI 消灭任务而非岗位」？ 参考答案：编程自动化但软件工程师岗位增长约 10%；读片自动化但放射科岗位增长约 20%；律师助理岗位也在增长。理由是岗位的目的由多项任务组成，被释放的产能会去消化长期积压的需求。

与这本书其他内容的关系

这一章讲的是组织与技术在不确定中如何推进；第一章讲的是把想法讲清楚的方法。两者都指向同一件事：先在头脑里把问题结构化，再动手。下一章则是这套结构化的具体对象——日常的 AI 工作流。

The shape of the conversation

A conversation at Y Combinator's Startup School 2026 between Garry Tan and Jensen Huang, founder and CEO of NVIDIA. The arc runs roughly: the company started with the wrong technology, was saved by textbooks, survived on a Sega cheque, turned at AlexNet, and then — organisation, systems thinking, agents, jobs, robotics, and advice for young people.

The most counterintuitive part is the opening admission: NVIDIA did not succeed by picking the right direction at the start. Its founding algorithm and technology were completely wrong, and the company learned its way out only after admitting that.

Timeline

Starting out: wrong technology, then learning it

In 1993, with the PC only rumoured to be arriving, Huang and his two co-founders set out to turn every personal computer into a game console by augmenting the CPU with accelerators.

That direction — augmenting the CPU — turned out to be right. The algorithm and technology used to pursue it were, in his words, exactly wrong. By 1995 the company realised this, with 35 or 40 other companies already building 3D graphics for PCs, and the team not only having chosen the wrong path but not knowing the right one.

His message inside the company was that without confronting the fact that the technology did not work, there would be no company.

The scene he describes next is the most vivid in the interview: he went to Fry's with a couple of hundred dollars in his pocket, bought three textbooks on OpenGL and pipeline design, and brought them back for the engineers. NVIDIA went on to reinvent computer graphics and to produce most of the major breakthroughs in the field over the following 25 years.

His conclusion: technology keeps changing, so the technology choice is not fatal; the ability to confront reality and learn the next thing is the dividing line.

The big idea that was right: use accelerators to solve problems the CPU cannot — molecular dynamics, image processing, inverse physics, and eventually deep learning. That reframed the company from "build a great chip" to "accelerate an algorithm domain".

His definition of a great company follows from this: not technology and not market, but a unique perspective about the world that you deeply believe in, ideally paired with a pursuit that is hard to execute. Having the right technology for the right market at the right time makes life easier, but those are by-products.

Sega: admitting failure and still getting paid

The deal that exposed the wrong algorithm came from Sega, which had contracted NVIDIA to build the successor to the Saturn — what became the Dreamcast.

Huang flew to Japan and told the CEO directly that the technology did not work, that the contract could not be fulfilled, and that Sega should choose someone else. He also told him, plainly, that he still needed the money.

The reply, as he recalls it: so what you are telling me is that you cannot do what I contracted you to do, but you would like all the money anyway? Yes, exactly.

He describes the manner as polite, humble and honest, and the CEO recognised it as such. The $5 million that followed kept the company alive long enough to find the right direction. His summary is the familiar investing line: you do not invest in companies, you invest in people.

NVIDIA went public in 1999 at a $300 million valuation.

AlexNet: seeing a method rather than a model

Many people saw AlexNet as a model with high accuracy. Huang says he saw a universal function approximator: give it the answer for almost any function and it can learn what the function is — and most interesting problems are imprecise by nature, and cannot be made precise.

That judgement changed the whole stack rather than a single product. He describes reasoning almost immediately about a five-layer structure — processor, middleware, algorithms, applications, and how people would use them — and starting work on computer vision, robotics and self-driving cars at once.

One methodological habit is worth extracting: his lens has always been algorithms, not chips. When something new appears, the questions are not "what product is this" but:

• What is this algorithm?
• Why is it effective?
• What else can it do?
• If you scale it, what could it solve that is unsolvable today?

Organisation: first principles, and building a car you can drive

Asked how going deep — reading papers, talking to principal scientists — coexists with running an organisation, Huang answers with temperament rather than technique. Everything starts with curiosity. He seeks the answers himself first; when a field turns out to matter to the company, he learns as much as possible so he can be of service by sharing it. In his framing, a CEO serves the company and its people by handing them insight.

He uses surfing to explain why staying close to the ground matters: the wave looks like chaos, but a surfer can read it because they have been in the water. The faster technology changes, the more that tactile sense is required; second-hand reports make it feel too fast to understand.

The better-known image is the F1 car: founders are building the car they will race, so it should be built to fit how they drive. Asked what happens to the company when he leaves, his answer is that when he dies on the job, the next CEO will reshape the company to fit themselves.

Systems thinking replaces coding

Asked what capability matters most, he names systems thinking: systems understanding, systems design, systems organisation.

The reasoning is direct. Low-level work is being taken over by agents. In chip design, transistors, gates and functional blocks have long been synthesised automatically, so designers are really systems designers; in software, most coding will be done agentically. What remains for people is to think abstractly about the system — what problem, what constraints, where the inputs and outputs are, at what rate information flows in and out, and whether the bottleneck is processor, memory or networking.

That knowledge, he argues, will never become useless, and will only become more useful.

Agents: capability is there, control is not

His view of agents has a concrete edge: recursive self-improvement is already happening. Every time you use one, it updates its long-term memory, rewrites Markdown files, compacts memory or turns it into knowledge graphs.

The bottleneck is not capability but fine-grained controllability. His example: change one word in a plan file and it should produce a precise local difference — one pixel, one triangle, one component in a CAD file, one layer, one via, one connection — with everything else regenerated around it.

He does not require the agent to be perfect: 80% and a human finishing the rest is acceptable; 99% and a human finishing the last bit is fine too. Controllability matters more than accuracy.

On open source, he calls OpenClaw a Linux moment: the operating system that will hold a large language model, which everybody should be able to build for themselves. NVIDIA's response was to offer its engineers to those projects.

Jobs: AI eliminates tasks

Asked what happens to the economy, he answers that many tasks will be automated — especially where all the information is already at hand and the job is to produce a response — that every job will change, and that many new jobs will appear.

He offers three counterexamples to the "AI destroys jobs" narrative:

• Programming as a task has been heavily automated, yet software engineering jobs are growing about 10% year over year.
• Reading radiology scans has been automated, yet radiology jobs have grown roughly 20% in recent years — because the backlog of patients is enormous, and seeing more patients requires more nurses and radiologists.
• Paralegal jobs were predicted to disappear, yet they are growing fast, because the backlog of lawsuits is similarly large.

His explanation: a job has a purpose, and the purpose consists of many tasks. Tasks can be removed; the purpose is not removed. Capacity freed by automation runs into demand that was never being met.

Physical AI: from generated video to robots

What convinced him robotics was close was generative video: if you can generate video of a finger moving, or a hand picking up a glass, why can't a robot do the same?

That raises the central question of how a robot generates motion that obeys physics — causality, friction, tension. This is what he calls physical AI, built on world foundation models and three pieces: real-to-sim, physics-grounded and generative simulation, and sim-to-real.

He argues the ChatGPT moment for robots already happened about two years ago: just as early ChatGPT produced nothing useful but expanded what people imagined, robots two years ago could already be reinforcement-learned, fine-tuned and grounded in physics. What remains is environments, evaluation and post-training.

On sequencing, they judged that the first application with a large enough market and sufficiently standardised technology was self-driving. NVIDIA chips and software are in Waymo, Tesla and Mercedes vehicles and data centres, and they open-sourced the self-driving stack (Alpamayo) because agriculture, mail delivery and warehouse AMRs all need it, and none of those markets alone justifies building one. His figure: the robotics and autonomous vehicle business is close to $10 billion now, will become the next hundred-billion-dollar business, and will take between three and ten years.

Advice for young people

• Simple work gets automated. He compares it to long division: his generation had to learn it, nobody does it by hand now. Sitting at a computer writing code is next.
• Hard problems remain. Physics, chemistry, biology, computer science, computer engineering, systems thinking — especially where fields intersect, plus the intersection of technology and social issues, and spotting gaps in markets.
• Systems thinking compounds. The better you can orchestrate millions of agents solving problems autonomously, the better placed you are.
• Still go to school. His words: stay in school.
• Adopt "how hard can it be". He concedes it is always harder than you think, but the point is not to convert that difficulty into anxiety before you start; let the suffering arrive a little at a time.
• Resilience is the single most important trait. You do not have to overcome life in one day — only this morning, only today, then work towards tomorrow.

Comprehension check

1. NVIDIA started with technology that was "exactly wrong". Why did the company survive? Answer: They admitted the technology did not work and immediately went to learn the right approach, buying OpenGL textbooks for the engineers. Huang's conclusion is that the technology choice is not fatal; confronting reality and continuing to learn is what matters.

2. Why did the $5 million from Sega matter? Answer: The company had found its algorithm unworkable and could not fulfil the contract. Huang told Sega so, advised them to choose another partner, and said he still needed the money. Sega paid based on trust in the people, which kept NVIDIA alive until it found the right direction.

3. What did he see when he saw AlexNet? Answer: Not a single model but a universal function approximator — give it the answer and it learns the function, and most interesting problems are imprecise anyway. From that he reasoned about rebuilding the entire stack.

4. What does the "build a car you can drive" metaphor mean? Answer: Founders should shape the organisation around their own temperament and working style rather than adopt a generic management template, because they are the ones racing the car; the next CEO will reshape it for themselves.

5. Why does he consider systems thinking more important than coding? Answer: Low-level work is being taken over by agents — chip synthesis and most coding are automated. What remains is defining the problem, constraints, inputs and outputs, and locating bottlenecks, and that judgement is not automated away.

6. What do agents most lack today? Answer: Fine-grained controllability. The ideal is changing one word in a plan file and getting a precise local difference with everything else regenerated; 80% accuracy is acceptable because controllability matters more than accuracy.

7. How does he argue that AI eliminates tasks rather than jobs? Answer: Programming has been automated yet software engineering jobs grow about 10%; radiology reading is automated yet radiology jobs grow about 20%; paralegal roles are growing too. A job's purpose contains many tasks, and freed capacity goes to demand that was already backlogged.

Relation to the rest of this book

This chapter is about pushing an organisation and a technology forward under uncertainty; the previous chapter is about making an idea clear enough to say out loud. Both point the same way: structure the problem in your head before acting. The next chapter applies exactly that structuring to everyday AI workflows.