The Rise of AI: A Weekend Read. In the late 1990s, we met the computer. We had no idea of artificial intelligence. Here are the first two chapters of a new series.
How our generation first encountered the machine, long before it could "think".
There are years that announce themselves with the modesty of calendar pages, and there are years that arrive wearing prophecy.
The year 1999 was of the second kind. The millennium stood somewhere ahead of us, glowing and mysterious, like a coastline seen from a ship after a long night at sea. Television channels counted down to the year 2000 with a mixture of celebration and alarm. Newspapers spoke of Y2K, the strange computer bug that was expected to confuse machines, freeze banks, ground aircraft, and perhaps bring modern civilization to its knees at the stroke of midnight. Grown men discussed it with priestly seriousness. Governments prepared. Companies panicked. The world, it appeared, had discovered that it was now dependent on computers and was not entirely sure whether the computers could be trusted.
I had other matters on my mind.
After the long purgatory of entrance examinations, coaching classes, answer keys, rank lists, rumours, prayers, and parental suspense, I had finally crossed the last great hurdle expected of a reasonably ambitious Malayali child of my generation. I had secured admission to medical college-MBBS!
The four letters had the weight of a family title. In Kerala, they did not merely describe a course. They conferred status, vindicated sacrifices, raised the morale of grandparents, and gave neighbours something to report with envy. I had entered, or so it seemed, the most coveted professional path available to the middle-class imagination.
But before the white coat, the dissection hall, the anatomy atlas, and the smell of formalin, there was a pause. Entrance results had been declared, but the course had not yet begun. For the first time in years, I had time; not the stolen time between exams, nor the guilty time after tuition, but real, empty, free-floating time. So I decided to join a computer training class.
Back then, this was far from the mundane routine it has become today. Throughout much of the developing world in the late 1990s, a computer remained a symbol of distant promise rather than a commonplace piece of household furniture. It had not yet been integrated into our pockets, dashboards, watches, schoolbags, or kitchen counters. Instead, it resided in dedicated rooms, shrouded under dust covers and minded by individuals who behaved like caretakers of a futuristic time machine. Interacting with a computer was akin to stepping into tomorrow, strictly by appointment.
I was interested because I was, in those days, what could safely be called a technology enthusiast without much technology to be enthusiastic about. So was Unni, who was my college tuition mate and then my MBBS classmate. We knew almost nothing. Words like Fortran, C, C++, compiler, operating system, algorithm—if we had heard them at all—belonged to some remote world populated by engineers, Americans, and people who owned manuals. Our ambitions were simpler. We wanted to play games.
Every morning, from ten to twelve, we went to a small computer training centre at Sakranthy. The room had the solemn air of an institution that believed it was introducing young people to modernity. There were plastic chairs, bulky monitors, keyboards with a satisfying clatter, and the faint smell of warm electronics. The computers themselves were not silent machines. They hummed, clicked, sighed, and occasionally protested with noises that suggested great labour. To our dismay, what we were taught was Microsoft Word.
The instructor, with patience that now seems saintly, explained how to open a document, type a sentence, change the font, save the file, and close the program without destroying civilization. We learned the difference between Save and Save As. We learned that Times New Roman was the voice of respectability. We learned that a document could be made bold, italic, underlined, centred, justified, and otherwise decorated into ugliness.
None of this interested us very much. The most captivating thing in that computer for us was neither Word nor PowerPoint, but a small animated paperclip who lived inside Microsoft Office, Clippy!
It had large eyes, a flexible spine, and a confidence entirely disproportionate to its usefulness. It appeared suddenly on the screen, leaning forward with evangelical concern and proclaimed, “It looks like you’re writing a letter!”
Long before Siri was summoned, before Alexa entered drawing rooms, before ChatGPT began composing essays, code, poems, excuses, sermons, and resignation letters, there was Clippy: cheerful, intrusive, dim-witted, indestructible. It was, for many of us, the first “personal assistant” that pretended to understand human intention. Clippy did not understand much, but the pretense mattered. A lifeless program had acquired a face. We did not know it then, but this was one of the earliest domestic forms of a relationship that would come to define the next quarter century: humans staring into screens while machines offered help they had not asked for.
The year 2000 arrived. The world did not end. Aircraft did not fall from the sky. Banks did not vanish. The millennium, having been advertised as an apocalypse, settled down into ordinary life with something close to embarrassment.
And our MBBS batch began.
We were the millennium batch, born at the hinge between two worlds: one foot in the analog certainties of notebooks, libraries, landlines, and handwritten lecture notes; the other in the digital disorder that was about to swallow everything. The batch had its own private mythology, its own nicknames, factions, jokes, legends, and minor dynasties. Among them was DGT, one of the first in our circle to have a computer at home.
By today’s standards, that machine would appear prehistoric. But to us it was not a device. It was a portal. There was the large CRT monitor, deep and heavy enough to injure a careless man. There was the CPU cabinet, humming with hidden authority and proudly displaying its 800Mhz power of Intel Pentium III chip inside. There was the floppy drive, which accepted square plastic disks carrying 1.44 megabytes of data—an amount so small now that a single photograph may crush it without noticing. There were speakers, cables, installation CDs, mysterious drivers, and the sacred ritual of restarting.
Games were, naturally, the real curriculum.
The official world may have believed that computers would make us productive. We knew better. Computers were machines for war, fantasy, competition, humiliation, and escape. They contained worlds where men raced cars in NFS, bikes in Road Rash, ran through corridors with impossible weapons in Quake and Unreal, and fired artillery across opponents' heads. Reflexes and frag points mattered more than careful dissection of cadavers and rank lists.
Soon, after much pleading, negotiation, arguing, emotional blackmail, and theatrical despair at home, I too acquired a computer. DGT came over to install the essentials. He did so with the calm generosity of a man initiating a novice into a mystery. He also, in his benevolence, gave me a hidden folder.
Unfortunately, my father discovered it.
I defended myself with the confidence available only to the young and guilty.
“It is a virus,” I said. The virus, I explained, had already attacked Deepak’s computer.
This was technically brilliant and morally weak. But in those days the word virus carried almost supernatural force. It belonged to that category of digital explanation that could cover everything from a frozen screen to family embarrassment. Nobody fully understood what viruses were, which made them useful. They moved invisibly from floppy to floppy, from machine to machine, like electronic fevers. They were the demons of the new age.
Around the same time, another hierarchy entered our lives: the hierarchy of hardware.
Every computer had a processor, memory, hard disk, monitor, and keyboard. But among boys who wanted to play games, one component inspired special longing—the graphics card.
The graphics card was not merely an accessory. It was the difference between spectacle and disappointment. Without it, games limped. With it, they acquired shadow, speed, fire, water, blood, smoke, and the illusion of depth. It determined whether a machine could merely run a game or reveal it. We spoke of graphics cards the way earlier generations may have spoken of imported motorcycles.
“The decisive instruments of history often enter the world disguised as toys.”
Two names dominated the imagination: NVIDIA’s Ge Force and ATI’s Radeon. They were not just two graphics card brands. They were banners. To prefer one over the other was to enter a minor tribal war. Their rivalry travelled through magazine reviews, computer shops, campus gossip, and the verdicts of boys who had once seen a game run at full settings on somebody else’s machine. The language was technical but the emotion was primitive. Frame rates. Memory. Drivers. Chipsets. Benchmarks. Glory.
The first-person shooters were the crown jewels. Doom had already become legend. Quake and Unreal belonged to the age of acceleration. These were not games for slow hands. They demanded reflex, aggression, spatial instinct, and a willingness to die repeatedly before lunch.
Paul was the champion of Quake and Unreal. He moved through those digital arenas with terrifying ease, shooting opponents as if flicking stones off the ground. Heads vanished. Bodies fell. The screen flashed. The bullets from his guns obeyed him then just as the cricket balls that strike his bat today.
I, by contrast, lived in humbler territory. I played lower-graphics games like Worms, where strategy wore the costume of comedy and small animated creatures murdered one another with bazookas, grenades, and improbable dignity. Even there DGT’s worms regularly destroyed mine.
This was our digital childhood—not childhood by age, perhaps, but by innocence. We thought we were learning computers because we were learning to install games, evade parental scrutiny, adjust display settings, and understand why one machine could render explosions better than another. We did not know that the same graphics hardware we coveted for gaming would later become the furnace of artificial intelligence.
That is the strange thing about history. Its decisive instruments often enter the world disguised as toys. The gamer’s craving for smoother graphics had helped subsidize the machinery of machine learning.
At the time, none of this was visible to us. The word artificial intelligence still belonged mostly to science fiction, chess computers, and the occasional newspaper feature about robots. Computers were not expected to write essays, diagnose images, pass exams, paint portraits, translate languages, compose music, or answer questions with unsettling fluency. They were expected to crash, hang, beep, and run games if properly threatened.
And yet, somewhere inside that noisy transition—from Clippy to graphics cards, from floppy disks to CD-ROMs, from Word documents to virtual battlefields—the foundations of the coming world were being laid. We had met computers when they were still awkward. We had met software before it became social. We had met the assistant when it was still a paperclip.
The machine had not yet learned to speak. But it had entered the room. And once inside, it never left.
How NVIDIA’s gaming chip became the engine of deep learning
It began inside a graphics card.
The revolution did not begin in a laboratory built for artificial intelligence. For years, the graphics card had appeared to be a creature of indulgence. Its purpose was to make games faster, explosions brighter, shadows deeper, water more liquid and blood more theatrical. It was the coveted possession of boys who wanted Unreal Tournament to run without a stutter, and who discussed frame rates and drivers with the gravity of generals inspecting artillery.
A conventional CPU—the central processor inside a computer—could be imagined as a brilliant clerk sitting at a desk, disposing of one difficult file after another. This was, of course, a simplification. Even the processors of that period could juggle instructions and perform several operations at once. But their architecture was designed principally for versatility and rapid sequential work: solve this problem, move to the next, change course when required.
In the consumer imagination, this power was reduced to a number printed prominently on the box: first megahertz, then gigahertz. Intel and AMD fought for the honour of producing the fastest desktop processor, and in 2000 AMD crossed the symbolic one-gigahertz frontier just ahead of Intel.
For a time, the processor industry seemed governed by a reassuring principle: every new generation would be faster than the last. But the march of clock speed was beginning to encounter the ancient enemies of machinery—heat and power. A processor could not simply be driven faster forever without consuming more electricity and producing more heat than could conveniently be removed. The future would therefore require not only faster workers but more workers, operating together.
The graphics processor had already been built around that idea of parallel processing.
A GPU was less like a single brilliant clerk than a factory floor filled with hundreds—and later thousands—of more modest workers. Each performed a comparatively simple calculation, but many of them worked at the same time. For graphics, this arrangement made perfect sense. A screen was composed of millions of pixels, and every frame of a three-dimensional game demanded an immense procession of similar calculations: this pixel must darken, that surface must shine, this edge must move, that shadow must fall.
The GPU had been designed for parallel labour.
And hidden inside that talent was a destiny larger than gaming.
By the middle of the 2000s, NVIDIA and its co-founder and chief executive, Jensen Huang, were no longer content merely to contest the graphics-card market with ATI. Gaming was lucrative, but it could also be unforgiving. A mistimed product, a flawed chip, a lost console contract or a superior rival architecture could turn success into embarrassment with remarkable speed. Hardware companies lived, as NVIDIA’s own history had repeatedly demonstrated, rather close to the cliff.
Huang therefore made what appeared to be an unglamorous wager: software.
The wager was called CUDA—originally expanded as Compute Unified Device Architecture.
In 2006, NVIDIA unveiled an architecture that opened the parallel-processing power of its GPUs to scientific and general computation. The first public CUDA software kit followed in 2007. Until then, researchers wishing to use a graphics card for non-graphics work often had to disguise their mathematical problems as graphics operations. CUDA allowed them to address the GPU more directly, writing programs in familiar languages and sending suitable portions of the work to the graphics processor.
“In plain English, CUDA turned the graphics card from a painter of pixels into a programmable workforce.”
A scientist could take a large calculation, divide it into thousands of similar pieces and dispatch those pieces to the GPU’s many processing units. With considerably more effort, a sufficiently large problem could be divided among several graphics cards. The GPU would not replace the CPU; the two would work in partnership. The CPU remained the versatile supervisor. The GPU became the massed labour below.
It was a formidable idea—and, commercially, an awkward one.
Very few gamers cared. They wanted faster games, not a new model of scientific computation. Most programmers had little reason to learn the peculiar discipline of parallel hardware. Wall Street preferred revenues that could be counted in the coming quarter. CUDA demanded that NVIDIA spend years constructing compilers, mathematical libraries, debugging tools and documentation for customers who, in any meaningful number, did not yet exist. NVIDIA describes CUDA today as the foundation of GPU computing; at its birth, it was closer to a road laid through a country no one had decided to inhabit.
For several years, the wager appeared to lead nowhere.
CUDA downloads reached a peak in 2009 and then declined for three consecutive years. NVIDIA’s share price languished. Members of its board worried that the company’s lack of momentum might attract an activist investor eager to dismantle it. The business was stagnant, and Huang’s great expansion beyond graphics had yet to produce the market he had promised. The company had created an answer before anyone could identify the question.
Great technological revolutions are often described afterwards as though every step had been inevitable. At the time, they more commonly resemble two unrelated failures awaiting an introduction.
In California stood a chip company with a programming system that few people used.
In Canada stood a small school of artificial-intelligence researchers devoted to a theory that much of their discipline had dismissed.
One possessed the machinery without the decisive problem.
The other possessed the problem without adequate machinery.
Their meeting would take place not in Silicon Valley, but in Toronto.
And the man who joined them was a quiet graduate student named Alex Krizhevsky.
Comments
Post a Comment