Mission Is the Boss
Liu Mingyu on the Cosmos Team and NVIDIA's Organizational Philosophy

Host: The NVIDIA book and Jensen Huang's biography both say a lot about his management style. Can you describe, from a more front-line perspective, what is different about how NVIDIA is managed? You already mentioned no layoffs and Top 5, where employees state the five most important things they are working on

Liu Mingyu: Top 5 lets people see signals large and small across the company, including two teams working on the same thing. NVIDIA is a relatively transparent organization; it pursues information transparency

Liu Mingyu: We often say Mission is the boss. NVIDIA sometimes behaves like an organism, without such a rigid fixed framework. When the company needs to move in a direction, teams spread across different organizations but with the relevant skills automatically align and move the same way. Cosmos is a case like that

Source location: "On Jensen Huang: Mission is the boss", prompted by the host asking about NVIDIA's front-line management style

Host: Many people may not belong to this team, yet they can all support its work

Liu Mingyu: Mission is the boss, rather than your boss being your boss. Mission is the boss. It is not a rigid framework; it is an organism

Source location: the host's follow-up and Liu Mingyu's reply in "On Jensen Huang: Mission is the boss"

Liu Mingyu: About eighty or ninety researchers and engineers in Cosmos Lab report directly to me. Other teams in the company also work on Cosmos; they do not report to me directly, but I am the overall PIC of the whole mission and decide the direction. Those people work together with the team. The whole Cosmos project is about two to three hundred people

Source location: team structure, 00:05:25

Liu Mingyu: What I worry about most, on the way to becoming a manager, is losing touch with the ground and not understanding the core of the problem. So I force myself to review everyone's code, read papers, challenge everyone's ideas and debate with them, to stay grounded and see clearly the pain points and each person's contribution

Liu Mingyu: If only I could decide the data mix and the architecture, this model would not succeed, because my knowledge is limited and so is my time

Source location: "From researcher to NVIDIA VP" and the discussion of the large Cosmos team

Host: Cosmos is a large team of nearly three hundred people. What is hard about driving a team that size to work together

Liu Mingyu: It is a hard problem of organizational structure. You want everyone aligned on the goal, but you also want everyone able to make decisions on their own. If the goals are aligned but nobody can decide, progress is very slow, because only a small number of decision makers can decide. Suppose only I could decide the data mix and the architecture: the model would not succeed, because my knowledge and my time are both limited

Liu Mingyu: If everyone can decide but the goals are not aligned, every team only wants to do its own part well and ignores the rest, the project diverges, different architectures appear, and you even get horse-racing between teams

Source location: the discussion of organizational design for the large Cosmos 3 team

Liu Mingyu: We need everyone to be able to make decisions that matter for the organization and for the model's development, so we have to make information transparent and give everyone enough information to make the right decisions

Liu Mingyu: When a person is afraid of deciding wrongly, they will not decide. So the organization has to let people decide, and also accept that some decisions will be wrong, but once a mistake appears it must be corrected quickly

Liu Mingyu: Everyone has to be clear about the final goal of the whole project. You keep describing the future clearly so that different roles see the same picture. What finally gets built may be 90% similar to the initial description and 10% different, but the shared picture lets each person, in their own position, make the decision that is right for the team

Source location: how to let more than two hundred people make decisions at the same time

Host: How do you make information more transparent and collaboration smoother? Are there concrete management methods

Liu Mingyu: You have to put real thought into how to explain things so people understand, and pay attention to communication. There are many team meetings; I communicate regularly with the main leaders and with the different parts such as Video and Understanding, and try to make meetings more engaging so that everyone participates

Liu Mingyu: Everyone skipping meetings and doing their own thing will not necessarily succeed; meeting all the time and only discussing without acting will not succeed either. The balance has to be adjusted to the task and the stage; there is no fixed formula

Liu Mingyu: The management style is both bottom-up and top-down. The key is judging what is right for the team at that moment, rather than always sticking to one mode

Source location: the discussion of information transparency, meetings and communication, and top-down versus bottom-up management

The interview opens by describing Liu Mingyu as more of an engineering leader than a typical researcher. Jensen Huang's assessment of him is GSD. Looking back on ten years of change, he himself says that early on it was not a lack of ambition, but that he did not understand how to get the company to see the importance of a direction and to secure enough resources to make it happen

The ability he later developed is judging which problems are worth solving, translating technical judgment into an investment case the company can understand, securing GPUs, people and organizational support, then getting different teams to collaborate and keep delivering results

Source location: the opening, 00:00:00, and the look back at execution in "Surprises, epiphanies, turning points"

Host: What is the one thing Jensen Huang said to you that left the deepest impression

Liu Mingyu: Once, in a meeting comparing two solutions, I was responsible for one and someone else for the other. I said the setting the other side used was unfair, so our performance fell short of expectations and lagged the competing solution

Liu Mingyu: Jensen Huang asked me, Are you a crybaby, and told me nothing in the world is fair

Source location: the sentence from Jensen Huang that left the deepest impression on Liu Mingyu

Liu Mingyu: When writing papers we often say the settings differ, so they cannot be compared directly. But if the goal is to solve a problem, you have to create the conditions that solving it requires; if the conditions are not there, find a way to get them

Liu Mingyu: That line made me hold myself to a stricter standard. Finding reasons is useless; you have to give everything you have. I had heard no excuse before, but hearing Jensen Huang say it to me in a real decision-making moment left a deeper impression, and after that I stopped looking for reasons

Source location: the follow-up explanation of settings, fairness and no excuse

It was not the first time Liu Mingyu had heard no excuse. What really affected him was that Jensen Huang pointed it out right when a specific solution was lagging and he was explaining that the conditions were unfair, letting him see how he was using an objective constraint

Source location: Liu Mingyu's reflection on how the line changed him