Jensen Huang: From Three Textbooks to a Universal Approximator (AI column)

I saw it. conversation Harry Tana with Jensen Huang at Startup School 2026published 26 July 2026 years. Usually, such interviews quickly turn into a set of success stories, but here one engineering habit passes through the whole conversation: to recognize that the old model does not work, to understand a new area and see the change of the entire computing system behind a particular technology.

1The first story is about the mistake that started NVIDIA. The company wanted to turn a personal computer into a game console and chose its own algorithm of three-dimensional graphics. In 1995 In the past year, the team has realized that the technology is fundamentally wrong, and no one knows the right approach within the company. According to Huang, he went to Fry’s, bought three textbooks on OpenGL and the device of the graphics conveyor and gave them to engineers. The company, which later became the leader of computer graphics, literally completed its studies after launching and attracting investment.

2The second story is about Sega. NVIDIA was supposed to participate in the creation of the console after Saturn, the future Dreamcast, but with the chosen architecture could not fulfill the contract. Huang went to see Sega CEO Shoichiro Irimajiri, honestly explained why the project should be given to another contractor, and at the same time asked for money, without which NVIDIA would have closed. According to Huang, Sega paid $5 Not for the technology, but because I trusted the team. That money gave NVIDIA time to turn around.

3The third story is about AlexNet. In 2012 Many people saw it as a very good image classifier. Huang argues that NVIDIA saw something different: AlexNet was not a standalone solution for computer vision, but an example of a new way to build software. Instead of manually describing a function, we give the deep neural network examples of inputs and outputs, and the training picks up the internal parameters.

Here it is worth separately to analyze the formulation of Huang about the “universal approximator of functions”.

The idea was mathematically known long before AlexNet. Works 1989 Years have shown that a neural network, even with a single hidden layer and a sufficient number of neurons, is able to approximate with any given accuracy a very wide class of functions on a limited area. But this is theorem about the existence of representation. It does not say that the network will be able to learn quickly, that there will be enough data, that the solution will be compact, or that the model will behave correctly outside the training sample.

Deep neural networks have added a practical design to this capability. Sequential layers learn to collect more complex features from simple ones, and depth allows some functions to represent significantly more compact than a small network. But you still need the right architecture, loss function, data, computation, and optimization.

So AlexNet's breakthrough was not the discovery of the theorem itself. In 2012 Alex Krizhevsky, Ilya Sutskever and Jeffrey Hinton showedthat a large deep convolutional network can be trained on ImageNet using the GPU and obtained on the ILSVRC test.2012 top-5 error 15,3percentage 26,2% for the best competitive result. The gap was large enough that deep learning stopped looking like an interesting theory and became a working computing platform.

It is this shift in scale that Huang seems to have noticed. If the network can learn not one pre-programmed procedure, but different maps from the data, it will not only have to change the application. CPUs, libraries, learning tools, infrastructure, and then entire industries change. For NVIDIA, this was not a new market for the GPU, but a reason to rebuild the entire stack.

The other topics I would like to mention in brief:

Huang offers to build as a racing car to the style of its founder, and the next manager - to reconfigure it for himself; The main skill of the era of agents, he considers systems thinking, and the key unsolved task is the accurate control of the actions of agents. AI, in his opinion, automates individual tasks, but does not necessarily destroy the profession as a whole; the employment figures given in the conversation require a separate check. The “ChatGPT moment” of robotics, according to him, has already happened, and the next stage depends on simulation, evals and transferring skills from the simulator to the physical world. Fundamental disciplines, ability to learn and daily sustainability are more important to him than knowledge of a particular tool.

For me, the big takeaway from these stories is not that you have to “believe in yourself.” Huang's strength is shift classification. He saw in textbooks not an admission of defeat, in Sega money - time to change technology, and in AlexNet - not the winner of the competition, but a new way to program computers. Engineering is a rare skill: distinguishing a good result within an old category from an event that changes the category itself.

#AI #Engineering #Architecture #Leadership #Robotics #Research