Jensen Huang and Joel Hellermark on AI, Computing and Leadership

Joel Hellermark asks when Jensen Huang became convinced of deep learning. Huang recalls researchers preparing for the 2012 ImageNet competition seeking Nvidia GPUs and help programming neural networks; the effectiveness of their work prompted him to ask how far it could scale.

Huang saw deep learning as both an algorithm for difficult-to-specify problems and a different way of developing software. He says that insight led Nvidia to rethink computing, an idea later popularly described as “Software 2.0.”

Discussing transformers, Huang says researchers have adapted the architecture to text, vision and audio. He argues that combining modalities can improve perception: words describing stripes might help a model relate a zebra to a horse.

He points to camera, radar and lidar sensor fusion, as well as the way words and gestures clarify each other. He expects multimodal models to become more capable and robust, though these are predictions rather than guarantees.

Huang suggests a model can learn descriptions of physical effects from language, while direct sensory grounding adds nuance. For robots, he proposes feedback from simulated physics as a way to connect learned behavior with physical constraints.

He describes Nvidia’s Omniverse as a simulation environment for a robot’s digital twin, and compares simulated physical feedback with human feedback used for alignment. The comparison describes a proposed training approach, not proof that simulation perfectly reproduces reality.

Asked whether current architectures can keep scaling, Huang says he does not know the science well enough to predict the limit. He speculates that explicit rules and symbolic reasoning could complement learning from data and experience.

Hellermark asks what role people will have alongside much more capable AI. Huang says he already works with people who greatly surpass him in particular skills, and hopes wider access to intelligence will amplify human expertise and passion.

Huang argues that natural-language interfaces may make computing more accessible to people who do not know programming languages. He hopes this could narrow the digital divide, while describing examples of children using prompts to create programs and images.

Nvidia uses AI to explore chip-design alternatives, Huang says. Engineers choose the trade-offs—such as speed on a critical path or energy efficiency for repeated components—and the system searches many possible designs under those goals.

Asked about defensibility for AI applications, Huang emphasizes deep understanding of a customer’s domain and situation. A generally capable model may still need context about an industry, a particular problem and the people involved.

Huang identifies three uses of AI at Nvidia: improving how products are designed, changing what products do, and augmenting the company’s internal work. He cites AI-assisted chip design and software development as examples of the first.

For products, he describes using learned image reconstruction to infer missing pixels in graphics. He presents efficiency improvements as a result of combining hardware with models trained on large computing systems; the specific ratios are his 2023 description.

For internal operations, Huang imagines an assistant connecting information across employees, market demand and supply chains. He presents this as a longer-term vision for augmenting work and identifying signals people might miss.

On risks and opportunity, Huang says transformative technologies create disruption as well as new industries. He expects retraining to matter for displaced workers and predicts new roles such as prompt engineering, a forecast made in 2023.

He calls for investment in AI safety alongside capability, citing autonomous driving as an analogy. For language models he lists guardrails, human-feedback alignment and retrieval of supporting information as possible components; none guarantees that outputs are correct.

Joel Hellermark asks Jensen Huang how he stays technically current while running Nvidia. Huang credits knowledgeable colleagues willing to teach him and says he deliberately tries to be a good student.

Nvidia works across areas from autonomous driving to climate and digital biology, Huang says. Leaders in technology businesses need to understand the foundations well enough to distinguish a fundamental shift from a side turn.

He points to relationships among generative adversarial models, variational autoencoders and diffusion models as an example of how insight in one area can inform another. Technical intuition matters when a solution takes years to build on a rapidly changing foundation.

Asked about management practices, Huang says he starts from first principles: the organization’s purpose, inputs and outputs, and the conditions of its industry. Nvidia’s stated mission, he says, is to solve computing problems that are barely possible with conventional approaches.

Huang wants a company large enough to do its work well but as small as practical, attracting people who want difficult problems. To empower them and move information quickly, he favors a flatter structure over command-and-control hierarchy.

He says he has many senior direct reports and does not hold regular one-on-one career coaching with them. His rationale is that these experienced leaders know their fields; he communicates strategic direction to groups and refines it through feedback.

Huang argues that shared information and a flat structure make the company more agile. He says Nvidia is organized to work as one rather than through separate business divisions, with its structure shaped around accelerated computing.

Hellermark asks how Huang’s leadership has evolved. Huang says he can be direct when he disagrees, and spends time explaining the reasoning behind decisions so that colleagues learn how he weighs alternatives.

Meetings are not restricted to people of a particular rank, Huang says: recent graduates and specialists from different teams can participate. Both speakers value having the person closest to the problem or most expert in the room.

Rather than conventional status reports, Huang invites employees to send a “top five things” email about what they observed, did or learned. He says he reads roughly a hundred on a typical morning; the items need not all be formal work updates.

He says the flow is from employees toward leadership, not a central list that might bias what people report. Sampling these messages helps him see whether what people actually do matches the strategy the company says it has.

Huang says Nvidia plans continuously rather than relying on periodic one-year or five-year plans. The point, in his account, is to adapt as conditions change while keeping the organization aligned on what it is doing.

When Hellermark asks how to trust a contrarian intuition, Huang says to reason from explicit assumptions. He offers his vision of future data centers as factories that process data to produce intelligence as an example of extrapolating from a technical premise.

Conviction must be accompanied by a willingness to change one’s mind, Huang says. He describes openly acknowledging mistakes and repeatedly adapting plans, rather than defending an incorrect long-term commitment.

Hellermark asks how leaders enable people to do their life’s work. Huang says the leader’s mission is to create the conditions for others to do so, starting with choosing consequential problems rather than assigning commodity work.

Huang says his company does not center conversations on market share; in his framing, competing over an interchangeable product means talented people are repeating work already done. This is his stated management philosophy and does not establish that Nvidia faces no market competition.

He says walking away from commoditized businesses reinforces the message that employees should work on new problems. Choosing the right work and leaving the wrong work are, for him, both parts of creating the right environment.

Huang closes by stressing access to information and low hierarchy. Transparency across teams and the ability of even a new graduate to contribute in a meeting, he argues, help make empowerment real.