What Jensen Huang discusses in this interview is less the success story of NVIDIA as a company and more the process of fundamentally redefining one’s perspective.It forms a single, continuous thread: from confronting a false start and relearning from the textbooks, to gaining a perspective on accelerated computing, and from the reinterpretation of the “universal function approximator” to the vision that led to the reinvention of the five-layer architecture.As I trace this single thread, I aim to dissect how a genius’s vision frames the world and where it calls for reinvention. Rather than consuming this as a success story, I want to interpret it as a procedure for reshaping one’s perspective.
Accelerated Computing as a Blueprint
Jensen Huang’s interview begins by openly acknowledging the mistakes made at the company’s founding. The idea of bringing 3D graphics to the PC was a natural fit given the trends of the time.Given that the company was founded in 1993—a time when PCs were just beginning to gain widespread adoption—the desire to bring the experience of video game consoles to computers is understandable. The algorithm they chose was a choice that had been thoroughly thought through internally. However, it eventually became clear that this choice was not the path they should have taken.
“The choice of technology we started the company with was absolutely wrong.”
Translation: The technology we chose when we founded the company was completely wrong.
I did not interpret this statement as mere self-reflection. Rather, I believe that acknowledging a wrong starting point and determining whether one can relearn in line with reality is what determines the time one has left moving forward.As Jensen Huang reflected in the interview, by 1995 the number of competitors had grown from 35 to 40, and it was clear even from the outside that the path they had chosen wasn’t working. If they had pressed on without acknowledging their mistake at that point, the company would not have survived. Recognizing a mistake is not a step backward; it serves as a starting point for reexamining the world from a different perspective.
That different perspective is the concept of accelerated computing. It is the idea that by extending the capabilities of the CPU for specific types of computations, problems that were previously unsolvable can be solved.We now view 3D as merely the first application and have reevaluated the approach, recognizing that the same framework can be applied to many other fields, such as molecular dynamics, image processing, inverse physics, and deep learning. I believe that what truly sustains a great company is not technology or the market itself, but rather a firm conviction in its unique perspective on the world—and the more challenging that perspective is to pursue, the longer its impact will last.
The reaction to AlexNet was a clear illustration of this perspective. We reinterpreted it not merely as a single technique that dramatically improved accuracy, but as an inherent property of deep learning itself.
“We just learned the universal function approximator.”
Translation: We have acquired a universal function approximator.
The Universal Function Approximation Theorem is known as a mathematical result demonstrated by Cybenko in 1989 and by Hornik in the same year.The theory already contained a guarantee that any function could be approximated. When Krizhevsky et al. used AlexNet to halve the error rate on ImageNet in 2012, the theorem transformed into a practical tool that works in real-world computing.We view this event not as the outcome of a single contest, but as the acquisition of a universal operation: function learning. As a concept discussed in this dialogue, this reinterpretation serves as the starting point for a vision that has been developed over the past fifteen years.
Three Textbooks and Million
Facing Reality and Learning Anew
Admitting a mistake is distinct from the concrete act of relearning. What left a lasting impression during the interview was the candor with which they acknowledged that no one knew the correct approach.They explained that even the founding members did not have a precise understanding of which algorithms to choose. What was needed here was not relying on external authority or protecting internal pride, but rather the extremely unassuming act of returning to the textbooks at hand.
Fry’s Electronics was known in 1990s Silicon Valley as a store that stocked technical books and components.There, Jensen Huang picked up three books on OpenGL and pipeline design. OpenGL is a 3D standard API established by SGI in 1992, which defined the grammar of modern computer graphics.Textbooks aren’t flashy. Yet those three volumes were packed with step-by-step instructions on how to build a proper pipeline. In this scene, I see the true meaning of the word “reinvention.” It wasn’t a genius idea born from scratch, but rather the process of humbly relearning publicly available knowledge and adapting it to his company’s designs.
The narrative continues, describing how, in subsequent developments, modern computer graphics were “reinvented,” driving many of the major advancements of the past 25 years. It is inevitable that technologies themselves will be replaced. The mindset that “if we keep learning, we can start over” builds resilience in the face of such changes.
Dialogue in Japan and Funding for Survival
Another pivotal point was the dialogue with Sega. Sega had released the Saturn in 1994 and was on track to release what would later become the Dreamcast in 1998 as its next-generation console.According to the interview, NVIDIA had initially signed a contract worth approximately $12 million for that next-generation console. When a flawed algorithm was discovered, Jensen Huang reportedly spoke frankly with Shoichiro Irimajiri, who was the president of NVIDIA Japan at the time, explaining that the technology would not work and advising him to switch to another company.
I interpret this situation as being less about the skill of the business negotiation and more about how trust is placed in people. The attitude of honestly stating that the request could not be fulfilled while still requesting funding—because it was still needed—is something that would be difficult to achieve in a typical negotiation.As discussed in the interview, Mr. Irimajiri nevertheless decided to set aside $5 million as an investment in people.
“That $5 million kept us alive and, you know, gave me enough time to figure out what to do.”
Translation: That $5 million kept us alive and gave me enough time to figure out what to do.
More significant than the amount itself is the fact that it provided time. The funding wasn’t merely a figure to keep the company afloat; it served as a reprieve to face mistakes head-on, relearn, and move toward the next vision.As mentioned in the interview, NVIDIA was founded in 1993 and went public on NASDAQ in January 1999, with a market capitalization of approximately $300 million at the time of its IPO. When you consider the journey from founding to the IPO, you can see just how significant a turning point the time bought by that $5 million truly was.
How hard can it be
The mindset behind this willingness to relearn is succinctly summarized in the interview:
“If it’s important to do, we’re going to go learn it, and how hard can it be?”
Translation: If it’s important to do, we’ll go learn it. How hard can it be?
I interpreted these words not as a slogan of optimism, but as a way of dealing with anxiety. If what needs to be done is important, we go out and learn it. Rather than imagining the difficulty all at once and turning it into anxiety, we break it down into the next step right in front of us. The mindset is simple, but putting it into practice is not easy. In reality, challenges far more difficult than imagined keep cropping up one after another.Even so, by changing the question at the starting point, you can take that first step. The action of running out to buy three textbooks is an extension of this question. The sequence—learn if it’s important, and deal with the difficulty little by little as it comes—creates the conditions for starting over.
Rebuilding the Five Layers Over 15 Years
3D Is the First Application
The strength of the “accelerated computing” perspective lies in its ability to treat 3D graphics not as an end in itself, but as an example.According to Jensen Huang’s assessment, the area that needs to be expanded is not a single product, but the realm of algorithms. Fields such as molecular dynamics, image processing, inverse physics, and deep learning may appear to be separate markets, but from the perspective of accelerating computation, they belong to the same stratum.I see here a masterful choice of which layer to use to frame the world. Rather than chasing individual applications, grouping them based on the underlying nature of computation ensures that the perspective itself remains even as applications change.
The concept of heterogeneous computing became widespread around 2006, coinciding with the proliferation of GPGPUs. Since then, configurations that combine CPUs with other computational resources have become a practical option. The perspective that 3D was merely the first application sounds natural in hindsight, but maintaining that perspective at the time of the company’s founding was no easy task.The tenacity of this perspective lies in the fact that, even after a specific failure—such as an incorrect algorithm—we were able to reaffirm that the fundamental approach itself was correct.
If this, then what
When considering what is required based on the reinterpretation as a “universal function approximator,” first-principles thinking comes to the fore. The approach of breaking down premises into their constituent elements and then reconstructing them from there leads not to immediate improvements, but to a complete redesign of the system.The question Jensen Huang repeatedly poses takes the form: “If this holds true, then what happens?” The “If this, then what” question functions as a mechanism for transforming a list of facts into a blueprint for the future.
If what AlexNet demonstrated was not merely the learning of a single function, but a method capable of learning any function, then its impact will not be limited to processors alone. Further questions naturally arise from there: How will the computing stack change? How will software be rewritten? Which industries will be affected?I feel that this chain of questions serves not merely as a prediction, but as a clue for design. It is a process of estimating the scope of reinvention required based on a single observation. The fact that work had already begun on computer vision, robotics, and autonomous driving as early as 2012 demonstrates the speed at which these questions unfold.
Simultaneous Reinvention Across Five Layers
This led to the vision of a complete overhaul, referred to as the “five-layer cake.”It is a vision to reinvent layers—such as the processor, middleware, algorithms, and applications—not in isolation but simultaneously. As discussed in the interview, the fact that this vision has been held for about 15 years makes me realize just how long-term this endeavor is. This is not a short-term race; building a design that spans multiple layers requires a long-term vision and incremental learning.
This led me to consider why the single phrase “universal function approximation” necessitated a complete overhaul of all layers. If one were merely approximating functions, it would seem sufficient to create a single model.However, to use an approximation in the real world, we simultaneously need the computational resources to support it, an intermediary to invoke it, the procedures to construct it, and even the design of scenarios where the approximation generates value. If even one of these layers is missing, the approximation cannot function as a complete tool.That is precisely why there is such a disparity: while the concept can be summed up in a single word, its implementation spans five layers. I see this disparity as the very point where insight—rather than remaining a mere hypothesis—is transformed into a set of tasks.
An F1 Car Tailored to You and the Granularity of Control
On-site Curiosity and the Ability to Read the Momentum
A recurring theme in Jensen Huang’s remarks is the mindset of descending from curiosity to the front lines. If listening closely doesn’t provide a sufficient answer, he goes to verify it himself.This aligns with the idea that a CEO serves the company and plays a role in sharing insights to empower everyone. In a rapidly changing technological landscape, actions can appear chaotic without a sense of the front lines.However, if you grasp first principles, the dynamics begin to reveal themselves—much like reading the waves in surfing. Even if the waves appear irregular from the outside, the surfer can anticipate their rhythm. It’s not just about organizing information in your head; it’s about experiencing the waves firsthand and learning their patterns—this is what creates stability amid rapid change.
Control Where a Single Word Changes Just One Pixel
This on-the-ground intuition takes concrete form in discussions about agents. One skill cited as valuable for the future is systems thinking.As Meadows and others have shown, frameworks for capturing inputs, outputs, constraints, and the flow and stagnation of information become increasingly valuable as low-level tasks are automated by agents. Just as the synthesis of transistors and logic gates was once automated, designers are transitioning into system designers.
It is said that agents already possess a rudimentary form of recursive self-improvement. This refers to behaviors that make them smarter with each use—such as refining Markdown or compressing long-term memory to convert it into a knowledge graph.The next leap required is not rough improvement but finer control granularity. The goal is a level where changing a single word in a plan file results in only a single pixel, a single triangle, or a single CAD component changing as a differential update, while the rest is regenerated. It is believed that even with 80 percent or 99 percent accuracy, the system remains usable as long as humans fill in the gaps.I see the same philosophy here: rather than making machines all-purpose, we’re adapting them to a form that’s easier for humans to operate.
This direction is also reflected in how tools are used. Within NVIDIA, agents such as Claude Code, Cursor, and Codex—which gained traction between 2024 and 2025—are highlighted as examples of systems that learn by running autonomously in a sandbox environment.I feel that this approach—letting a thousand flowers bloom and learning from them—resembles the cultivation of an ecosystem rather than central control. Open-source initiatives like OpenClaw and Hermes are discussed in the context of the emergence of Linux, and the idea of supporting everyone in building their own AI resonates as a distributed concept that avoids concentrating control in a single location.
“You’re going to build an F1 racer, but you’re going to build it in a way that you can drive it. You should adapt the car to yourself.”
Translation: You’re going to build an F1 car, but you’re going to build it to suit your driving style. You should adapt the car to yourself.
“Founder Mode” is known as a concept articulated by Chesky and Graham in a 2024 essay. In Jensen Huang’s account, he reveals a self-understanding that this mode has been cultivated over the course of 34 years.The metaphor of tailoring an organization to oneself resonates as a practical necessity: speed cannot be fully leveraged unless the system is in a form you can fully control. Whether it’s a car or an agent, the condition for continued use is the ability to make adjustments at the granularity of a single word or a single pixel, rather than relying on fully automated operation.
The Distinction Between Tasks and Jobs
The distinction regarding agents and employment is clear.The distinction is that AI automates cognitive tasks, not the jobs themselves. A job is a bundle of numerous tasks with a common purpose, and it is believed that even if some parts are automated, the job as a whole will not disappear. In the discussion, examples such as coding and software engineers, image interpretation and radiologists, and legal documents and paralegals were discussed in terms of this same distinction.The view is that in fields with significant backlogs, increased productivity creates room to handle more cases, and the demand for human workers will actually increase. I interpreted this perspective of viewing jobs as a “bundle” not as either pessimistic or optimistic, but as a way to identify room for growth within the context of objectives and constraints.
Attempts to Grounded in Physics and the Conditions for Continuous Learning
From Video to Joints
The movement toward physical AI began with an intuition sparked the moment I saw video generation. The question was: if we can generate footage of a finger moving and a hand grasping a glass, surely we can generate joints as well. Within the company, we were already working with progressive GANs and conditional GANs.The former was introduced by Karras et al. in 2017, and the latter by Mirza et al. in 2014. It is said that, based on the experience of using neural networks to drive simulators that they generated entirely on their own, the research shifted toward enabling the networks to understand physical laws such as friction, tension, and causality.I interpreted this shift as a transition in which the focus of generation moved from the screen to the body. The level of precision required differs between generating content on a screen and generating joint movements subject to physical constraints. Precisely because flaws become apparent the moment one encounters physics, the concept of “grounding in physics” is changing the criteria for evaluation.
From Autonomous Driving to Warehouses
A three-stage process for learning and evaluation in the real world is outlined: “real-to-sim,” simulations grounded in physics and generation; “sim-to-real”; and reinforcement learning. Platforms that incorporate physics, such as Isaac Sim and Cosmos, are positioned as training grounds for this process. Autonomous driving was initially envisioned as the first economically viable application.It was chosen as a domain where the learning cycle could be easily driven by market size and standardization; chips and software stacks are used by companies like Waymo, Tesla, and Mercedes, and Alpamayo has been open-sourced, making it adaptable for use in agriculture, delivery, and warehouse AMRs.I see this as an extension of the “five-layer” concept—a movement to apply cross-layer design to other domains, rather than simply creating a single specific product.
Just as Intel’s 4004 was built with approximately 2,300 transistors in 1971 and Blackwell’s chip is discussed in terms of over 200 billion transistors, the scale of computation has grown by orders of magnitude. The question is not the scale itself, but how to choose which problems to apply it to.Regarding the physical AI business as well, it is discussed within a timeframe ranging from several years to within a decade.
“Learning is the single greatest superpower.”
“Resilience is probably the single most important thing.”
Translation: Learning is the greatest superpower. / Resilience is probably the most important thing.
I interpreted these two statements as overlapping with a story that began with three textbooks. Treating what you don’t know not as a source of shame, but as a fact that can be addressed by seeking to learn it. Repeating this process updates your perspective. The statement that resilience is the most important thing is spoken at the granular level of getting through this morning and this day—not as a way to overcome life all at once.I view the question “How hard can it be?” not as a mantra that downplays difficulty, but as an attitude of continuing to learn while gradually accepting the hardships. Whether it’s accelerated computing, universal function approximators, or attempts to ground our work in physics—all are efforts to expand the interface where computation touches the world.It is precisely because we can admit our mistakes that our vision can endure over the long term. I want to remember this conversation as a process where, by taking that first step—seeking out what is important—on a day-by-day basis, the “five-layer reconstruction” gradually becomes a reality.
Comment