Learning to Code, Part 3
Computer programmers are experiencing the first wave of AI disruption. What can we learn from them?
You can read Part 1 here and Part 2 here.
Learning to Code, Part 3
Don’t Call It Vibe Coding
AI researcher Andrej Karpathy is a big deal in the tech world. He was a member of the original team at OpenAI, the former head of AI at Tesla, and the founder of Eureka Labs, an AI education startup. A year ago, he coined a term that helped launch an amateur coding revolution.
“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,” he posted on X. He said he was now mostly just telling AI what to build rather than writing any code himself.
Vibe coding was a major philosophical shift. Many coders, like screenwriters or journalists, have a guild mentality, a canon of works that describe best practices, and see coding as a creative act. A bible of the profession is Robert C. Martin’s Clean Code: A Handbook of Agile Software Craftsmanship. “A programmer who writes clean code is an artist who can take a blank screen through a series of transformations until it is an elegantly coded system,” Martin writes. Exceptional examples of beautiful code are celebrated, and crappy “spaghetti code” is seen as a sign of laziness or incompetence and mocked on coding forums.
This is why Karpathy’s vibe coding tweet was an earthquake. A highly respected coder was suggesting that much of what you learned at MIT or Stanford or from reading Martin or by studying Bresenham’s 1962 line algorithm, which describes the most efficient way to draw a line on a screen, really mattered anymore. Karpathy’s pinned post on X said it all: “The hottest new programming language is English.”
Throughout 2025, the coding world fractured into squabbling camps. The gaming engineers took the most purist positions. “I’d rather quit the industry than use generative AI,” said one anonymous developer in a recent industry survey. The AI company leaders made the boldest predictions. In March, the month after Karpathy’s comment, Dario Amodei, Anthropic’s CEO, said, “In twelve months, we may be in a world where AI is writing essentially all of the code.” While some engineers revolted, Silicon Valley managers beat them back. Brian Armstrong, the CEO of Coinbase, literally fired engineers who refused to adopt AI.
Karpathy himself backtracked a bit and reminded his fellow coders facing identity crises that he was only talking about “throwaway weekend projects.” In October, he described AI-generated code as “slop.” But his original vibe coding post was actually correct about where the industry was headed.
The turning point came in late November when Anthropic released a more advanced version of its flagship model, Opus, just a few days before Thanksgiving. Claude Code had been widely available since May of last year, but Opus 4.5 turned it into a superpower. Then, OpenAI released its next version of GPT in December.
Many coders spent the 2025 holidays obsessed with these new, more powerful tools, while grappling with the realization that their profession had been transformed overnight. In the new year, the skepticism was replaced by existential dread.
Paul Ford, a programmer who runs an AI company, described how the 2025 holidays were for him and many others in tech, the moment when “you feel the earth shifting.” Major software projects that he once charged clients hundreds of thousands of dollars to build could now be created by Claude Code in hours. “I spent an entire session of therapy talking about it,” he wrote.
Gergely Orosz of the Pragmatic Engineer, which is one of the biggest tech Substacks and which I’m relying on extensively in this section because the site currently serves as a real-time chronicler of how coders are grappling with AI disruption, collected some of the post-holiday aha moments in his industry:
A Google engineer noted that AI “generated what we built last year in an hour.”
A developer seemed to have an identity crisis: “For more than 15 years, I thought I loved writing code, loved typing out code by hand … 2025 was the year in which I deeply reconsidered my relationship to programming. … What I learned over the course of the year is that typing out code by hand now frustrates me.”
A chief technology officer concluded that the “cost of software production is trending towards zero.”
In January, Anthropic released Claude Cowork, an AI agent that integrates into common apps to help you get work done. But what made Cowork another seminal moment in AI isn’t what it does; it’s how it was made. Claude Code itself wrote its partner app in a week and a half.
Whatever resistance was left has been snuffed out by management. Meta now tracks AI usage among its coders as a data point for use in performance reviews. At Microsoft, there’s an internal leaderboard of employee AI usage.
On Wall Street, this shift was translated with blunt force, as investors indiscriminately pulled out of old-school software brands and poured money into safer bets, such as makers of the new AI tools. In one week in February, software companies lost $1 trillion in value. Meanwhile, Cursor, an AI coding startup valued at about $400 million in mid-2024, is now reportedly raising money at a $50 billion valuation.
Gradually, after the shock of the 2025 holidays, a more optimistic view started to emerge from key figures in the coding world: ceding some knowledge work to automation might be liberating. For programmers, it can mean the freedom to move up a level of abstraction, away from lines of code and towards a focus on architecture and overall product design.
Jensen Huang, the CEO of Nvidia, has framed the debate as distinguishing between purpose and task:
The purpose of a software engineer is to solve known problems and to find new problems to solve. Coding is one of the tasks…. If your purpose literally is coding—someone tells you what to do, you code it—maybe you’re gonna get replaced by the AI…All of our software engineers—their goal is to solve problems….Nothing would give me more joy than if none of them are coding at all, they’re just solving problems.
You might respond, Well, of course, the head of a $4 trillion company selling the chips that power AI is going to put a happy face on the coming employment apocalypse!
But it’s not just Huang. The tech press is filled with non-doomsday scenarios in which AI coding emancipates programmers and elevates them to software engineers. Peter Steinberger, the programmer who turned a personal project into the phenomenally successful OpenClaw, has proudly pointed out that he didn’t even read the code before he shipped it.
“There are these people that write software the old way, and the old way is going to go away,” he said of his critics on a recent podcast. “They call it ‘vibe coding.’ I think vibe coding is a slur. They don’t understand that it’s a skill.”
Karpathy’s view has also evolved. He had the same epiphany over the 2025 holidays as everyone else. In February, to mark the one-year anniversary of his original vibe-coding post, he summarized the scale of change in his industry, noting that AI coding is now the “default workflow for professionals.”
Just don’t call it vibe coding. Karpathy now prefers the term “agentic engineering,” a new craft requiring an updated set of skills where the programmer is less of a typist and more of a conductor:
— “agentic” because the new default is that you are not writing the code directly 99% of the time, you are orchestrating agents who do and acting as oversight.
— “engineering” to emphasize that there is an art & science and expertise to it. It’s something you can learn and become better at, with its own depth of a different kind.
If he’s right, that is certainly a much happier version of the AI displacement story, at least for the coders.
Boris Cherny, who created Claude Code, recently compared coders to medieval scribes, who painstakingly produced books by hand before the invention of the printing press. They had specialized skills learned through years of training, but after Gutenberg, “This thing that was locked away in an ivory tower” was “accessible to everyone” and “none of the things around us would exist today without” it.
AI automation is not always going to have a clear upside. But it’s crucial to distinguish cases where AI is ruining something with a cheap, inferior replacement from cases where it’s actually empowering ordinary people with previously unavailable tools and knowledge.
We have all, at some point, experienced a sense of awe using an app, but the ability to create one ourselves has been out of reach because the coding skills necessary to conjure a premium piece of software have seemed like a kind of wizardry too difficult to learn. With that power now unlocked and in the hands of anyone with a computer, we are going to witness an explosion of creativity.
For journalists and other public-interest professionals, that means resisting knee-jerk hostility towards anything labelled as AI and carefully thinking through how these new tools might enrich our work.
So, armed with all of this context, I started building an alternative to TurboTax.
Coming tomorrow in Part 4: Falling Down the AI Rabbit Hole