The Machine Learns to SeePart A2 — From Perspective to Pixels to AI · ~1420 CE–today

📖 The Story

The Renaissance turned painting into a branch of optics. Brunelleschi's demonstration of linear perspective (~1415–1420), codified in Leon Battista Alberti's 1435 treatise, let artists project a three-dimensional world onto a flat surface with mathematical precision. For the first time, a picture obeyed rules a machine could one day follow.

Artists also reached for actual machines. The camera obscura — a darkened room or box that projects the scene outside through a tiny hole — let painters trace reality directly, and many Old Masters almost certainly used it. The image was right there on the wall; the only thing missing was a way to make it stay.

Chemistry supplied the answer. In the 1820s Nicéphore Niépce fixed the first permanent photograph; in 1839 Louis Daguerre's daguerreotype made the process practical and stunned the world. Suddenly an image could be captured with no hand drawing it at all. The machine had begun, for the very first time, to make pictures by itself.

📖 The Story

Show a series of still photographs quickly enough and the eye fuses them into smooth motion. That simple trick, plus Muybridge's stop-motion experiments, opened the door to moving pictures. Edison's peephole Kinetoscope came first; then in December 1895 the Lumière brothers held the first public film screening in Paris, and a new art form was loose in the world.

Film grew fast: sound, then color, then the cinematic language of cuts and close-ups. And alongside it rose animation — the audacious idea of drawing motion that never existed in front of any camera at all. Frame by frame, artists like Disney's studio conjured living worlds purely from imagination and ink.

That ambition contained the seed of everything to come. Film could record the world; animation could invent one — but only as fast as a human hand could draw each frame. The obvious dream was a machine that could generate the moving image itself. To build it, the picture would first have to become something a computer could understand: pure number.

📖 The Story

With Sketchpad (MIT, 1963), the picture stopped being pigment and became mathematics: points, lines, and polygons stored as numbers, painted onto a grid of pixels. Over the following decades that idea matured into flight simulators, then dazzling CGI in films, then real-time 3D rendered on the fly.

But drawing a 3D scene by math is brutally expensive. For every one of thousands of polygons, and then every one of millions of pixels, the computer must run the same small calculations — transform this point, shade this dot, again and again and again. It is not hard math. It is just a staggering amount of it, all at once.

And here lay a problem. The ordinary CPU — the computer's main brain — is a brilliant generalist, built to do one complicated thing at a time, very fast, in sequence. Faced with millions of identical tiny sums, it choked. Drawing rich 3D worlds in real time would require a completely different kind of chip: one that could do thousands of simple calculations simultaneously.

📖 The Story

Three-dimensional games lit the fuse. Wolfenstein 3D (1992) and Doom (1993) faked 3D cleverly; Quake (1996) rendered a true polygonal world in real time. Players wanted more, and a market exploded for dedicated 3D boards — 3dfx's Voodoo (1996) was an early sensation. Then NVIDIA's GeForce 256 (1999) arrived, marketed as the world's first graphics processing unit.

The GPU's secret is exactly the answer Chapter 03 demanded: massive parallelism. Where a CPU has a handful of powerful cores, a GPU packs thousands of tiny ones, all doing the same simple arithmetic at the same instant — a perfect match for the millions-of-identical-sums problem of drawing pixels.

Then came the move that changed history. In 2006–2007 NVIDIA released CUDA, a software layer that let programmers use the GPU for any kind of math, not just graphics. The most powerful parallel calculator ever built for play had just been thrown open to the world. The question was: what else needs millions of identical sums, all at once?

📖 The Story

It turns out a neural network is mostly one operation repeated trillions of times: multiplying huge grids of numbers together, in parallel. That is the GPU's native language. When Krizhevsky, Sutskever, and Hinton trained AlexNet on gaming GPUs in 2012 and crushed the ImageNet contest, they lit the fuse on the entire deep-learning era.

From there it accelerated. The 2017 Transformer architecture, run on ever-larger clusters of GPUs, gave us modern AI — the systems that now write, reason, and, fittingly, make art. Tools like DALL·E, Midjourney, and Stable Diffusion generate images from words on the same silicon descended from the gamer's graphics card. The wheel comes full circle: the machine built to paint imaginary worlds for art has learned to make art itself.

And the scale of it is staggering. NVIDIA, the company that once just made chips so kids could shoot faster monsters, sits at the tollbooth of the AI economy — and in 2026 became the most valuable company on Earth, worth roughly $5 trillion, more than the entire annual output of all but a couple of nations. The thread that began with a hand pressed to a cave wall, fifty thousand years ago, runs unbroken to here: the oldest human impulse, to make images and stories, built the very tools that built the first minds that are not human. What they make of that inheritance is the next chapter — and it hasn't been written yet.