“There’s a new kind of coding I call ‘vibe coding,’ where you fully give in to the vibes, embrace exponentials, and forget that the code even exists.”
February 2 carries certain associations in our part of the world, but on that same date last year the IT world of the future was born. That was when a world-famous programmer, with the words above, created an entirely new field. “Vibe coding”—a world in which software is made not by writing computer code but by explaining things in human language.
In recent years Andrej Karpathy has been one of the few scientists who not only build artificial intelligence but also rethink it and explain it to the whole world—through lectures and blog posts that read like adventure novels. The Slovak coined a huge share of the terms we now use in the world of AI, as well as the concepts that help us make sense of it and put it to use. A co-founder of OpenAI and one of the builders of Tesla’s self-driving cars, he surprised everyone a month ago by announcing that he was joining Anthropic—a company that is undoubtedly paving the way to the future.
So what will that future look like, according to Andrej Karpathy? And what kind of man is one of the prophets of the age of advanced artificial intelligence?
One of Our Own
Andrej Karpathy was born in 1986 in Bratislava, then part of Czechoslovakia. At fifteen he moved with his family to Toronto. Technology drew him from an early age, and there he had the great luck to land in the course of the “godfather of AI,” Geoffrey Hinton. This was many years before today’s AI boom—even before neural networks turned out to be such a successful technology and before large language models became a synonym for “artificial intelligence.” Back then the approach was an eccentric idea that only a handful of researchers believed in.
Next came a master’s degree in Canada and a PhD at Stanford under Fei-Fei Li, the creator of modern computer vision technology.
It was here that Andrej first stood out as an innovator, creating CS231n—Stanford’s first course on deep learning, which became a legend. The lecture notes still circulate on the internet to this day, and they taught a whole generation of engineers at the moment the field suddenly proved so important and so successful.
And that is what makes our hero special. He is at once a deep researcher of everything newest in the world of AI and a born teacher who explains things so that anyone can understand.
Picture Perfect
One story from that time is worth telling. Neural networks were just beginning to recognize objects in photos surprisingly well, but… no one knew exactly how well they were really doing, because there was nothing to compare them with. What was the “human score” on the same task?
Back then we were still wondering—could we lose this contest, too? Just like the chess battle between man and machine, which later spread to every other field—sport, knowledge, skill.
Were algorithms really capable of recognizing what was in an image better than a human?
Instead of guessing, Karpathy sat down and spent hours personally classifying thousands of images from ImageNet—the huge database used to measure how good these systems are—to establish the error rate of a careful human eye. About 5 percent.
Then he pitted the network against his own result.
And he lost.
Shakespeare the Programmer
And so image recognition became yet another lost battle in the contest between human and machine.
In 2015 Karpathy published a blog post describing an experiment of his—how he had trained a neural network on a large and varied body of text and then made it write texts of its own. The system began churning out pseudo-Shakespeare plays, inventing nonexistent names, even generating plausible computer code that—alas!—didn’t work.
Sound… familiar? Quite a few years would pass before we all understood how far large language models had come. How easily they start to hallucinate, too, and how quickly technological solutions would be found for those problems as well.
A little later came the essay that became his calling card and made him truly famous—“Software 2.0.” Its thesis was radical for its time: a new paradigm had arrived in how software is made. In “Software 1.0” everything is in our hands, whereas now the human is no longer the one who formulates the logic—he selects the data, and the neural network builds the program itself on that basis. You program by supplying examples, not by writing commands.
This is not an abstract idea but a new paradigm, Karpathy was convinced, and to prove it he decided to try it on the largest scale possible. From 2017 to 2022 he led artificial intelligence development at Tesla—precisely the team that built the computer vision for the Autopilot. A car cannot drive on millions of lines of code spelling out that if it sees a pedestrian, it must stop. Instead, the system learns from its own experience while the engineers guide it. Those were exactly the years when self-driving cars advanced dramatically, and Karpathy’s team turned out to be the pioneer.
GPT
In February 2023 the Slovak returned to OpenAI—the company he had helped found eight years earlier as one of the first eleven people on the team. As you know, it was the epicenter of the buzz around ChatGPT, the moment when artificial intelligence suddenly burst into every computer and every conversation.
He stayed only a year before leaving to start his own education startup. Eureka Labs is built on Karpathy’s belief that the biggest obstacle facing education today is a shortage of excellent teachers. What if courses were designed by the best people in a given field, while the teaching itself, the exercises and the feedback were scaled up by artificial intelligence? To this day, it remains one of his biggest ideas and causes.
Around this time our hero created a long series of detailed videos explaining how deep neural networks work and how we can build them. Because he is convinced that this is the path to understanding technology. That if something looks like magic, it means we haven’t explained it well enough yet.
Karpathy popularized a number of ideas that became the way the entire IT industry thinks about itself. From the notion that English is “the hottest new programming language,” because we now speak to machines in human words, to the idea that large language models are something like “people spirits”—statistical reflections of everything humanity has ever written. And the model is the kernel of a new kind of operating system, in which we are simply the users.
The Vibe of the Vibe
And so we arrive at February 2, 2025, when Karpathy posted on X the simple sentence we began with—and practically gave birth to a whole revolution. True, the linguists haven’t caught up yet, but for people in the tech world “vibe coding” is unquestionably more than the word of the year. The term described our new way of working: “you fully give in to the vibes… and forget that the code even exists.” You just tell the artificial intelligence what you want in plain language, accept whatever it hands back without reading every line, and let the program “grow” on its own.
Vibe coding ignited a heated debate that still hasn’t ended. For some it is a liberation—suddenly people without years of engineering training can make software just by describing their idea in words. For others it is a dangerous abandonment of understanding—if no one reads the code anymore, who is responsible when something “breaks”?
Karpathy himself, balanced as a rule, stands somewhere in the middle: he admires the new power, but keeps insisting that anyone who wants to truly master it must also understand what is happening under the hood.
But are we capable of doing that? The Slovak doesn’t offer an answer—quite the opposite. At the end of last year he wrote, “I’ve never felt this much behind as a programmer.” Because he has the feeling he could be ten times more productive if only he managed to properly string together everything that has appeared in the past year alone. And he feels his failure to do so almost as a personal shortcoming.
Jagged Intelligence
He also wrote that commanding a whole army of AI agents writing code for you sometimes drives you into “a state of psychosis.” A new cult phrase appeared as well: “jagged intelligence”—the idea that today’s models are at once superhumanly good at one task and surprisingly dumb at another, often without our understanding the logic of where the edge runs.
Vibe coding itself, he also noted, is already getting old. The carefree “give in to the vibes” is giving way to something more disciplined, which people have started calling agentic engineering—not just letting the machine write, but orchestrating a swarm of agents, setting them checks, monitoring them. The same tools, a completely different approach.
Recently his next idea emerged, and it changed the approach to AI agents. In his view, each of us would do well to assemble a personal “wikipedia” of important data that is constantly updated, and to organize the way we work with AI around it.
His new job at Anthropic is on the pre-training team—the phase of the enormous training runs that give AI its basic abilities. His goal is stated quite clearly: to build a team that uses Claude to accelerate the research for the next Claude.
And it all seems to fall into place. Just as he once taught Tesla’s cars to drive themselves, just as in “Software 2.0” you don’t write the program but create the conditions for it to emerge on its own—well, here we reach the moment when once again the machine improves from its own experience. Only now the subject is artificial intelligence itself.
Simply the Future
So what can we expect? What are the next astonishing skills that will cross to the other side of jagged intelligence? Are there any left at all?
Before we part, let’s give the floor to Karpathy himself. Together with Claude, we picked five quotes from one of his few in-depth interviews:
-
In his view we are living in the “decade of agents,” not the year of them. For Karpathy, truly capable AI agents—ones that could do the work of an intern or an employee—are about a decade away, not a year or two. Today’s models, he says, have serious “cognitive deficits”: they lack continual learning, persistent memory, good multimodality and the ability to use a computer. He describes his own forecasts as 5 to 10 times more conservative than the typical expectations of the Silicon Valley magnates, yet still more optimistic than those of the hardened skeptics.
-
Karpathy rejects the scenario of a sudden leap in which AI progress abruptly sends the economy soaring to 20 percent annual growth. His argument is historical: computers, the internet and mobile phones were also deeply transformative, yet there is no “computer moment” in the GDP curve—they seeped in slowly and blended in with the other drivers of growth.
-
The future of work, in his view, is augmented intelligence, not replacement. Instead of robots that do everything for us, Karpathy bets on the “human in the loop” model—AI takes on the routine and the automatic, while the human remains for the new, the creative and the high-risk. Automation is selective, not total, and the next ten years at least, he believes, will be an era of amplifying humans, not displacing them.
-
He believes we have made great progress with AI agents so far, but the hard part is still to come—his experience with self-driving cars shows as much. The road from 99 to 99.9 percent reliability is as hard as the road from 0 to 90 percent or from 90 to 99 percent was before it—each additional “nine” demands at least as much work as the previous one. That is why impressive demos are misleading: deploying truly reliable agents will take a decade of hard work. “And we,” he adds, “haven’t even reached 99 percent yet.”
-
Today we learn for utilitarian reasons—to get a job. According to Karpathy, once machines take over labor, learning will remain for its own sake: for pleasure and to keep the mind in shape, the way we go to the gym today not because anyone needs our muscles. His vision for Eureka Labs is exactly that—an AI tutor delivering perfectly personalized education, accessible to everyone.
Quite a lot more calm, and quite a big difference from the forecasts of most tech leaders, you’ll agree. But that’s just the kind of man he is—exceptionally smart without seeming arrogant. Influential without being loud. Technically deep, yet approachable. Connected to the biggest companies, yet somehow still holding on to the image of a man who can simply sit down and explain. And even reassure.
So who will turn out to be right? Only time will tell. Time—and the next Claude, or perhaps the next essay in which Andrej Karpathy captures the spirit of the age…
