News Every Day | 15:30

Vibe coding is coding, period

Hello, and welcome back to Fast Company’s Plugged In.

“Programming, as it turns out, is just typing.”

Talking at Cisco’s AI Summit in San Francisco on February 3, Nvidia CEO Jensen Huang made that pithy observation to sum up the phenomenon of people using AI coding tools to simply describe in plain language software they want to exist, with an algorithm doing the heavy lifting. The comment came during a wild, wide-ranging riff on how AI is changing the world, and Huang kept joking that his chatter might have been influenced by several glasses of wine. (Hey, he was the after-dinner speaker.) But even if alcohol-fueled poetic license was involved, the sentiment captured the present moment.

The earliest evidence that AI could transform how people program computers came even before ChatGPT’s arrival, dating to when GitHub released the first version of its Copilot in 2021. At that point, AI was autocompleting snippets of code for humans rather than generating software from scratch. The progress has been radical since then, reflected in the boom for coding agents such as Cursor, Windsurf, Replit, and the industry’s current darling, Anthropic’s Claude Code. Along the way, the act of willing software into reality through AI got a name: vibe coding.

At the Cisco event, Huang, OpenAI CEO Sam Altman, Andreessen Horowitz cofounder Marc Andreessen, and other Silicon Valley luminaries talked about the whole industry having arrived at a crucial juncture in the pivot to AI software generation. Anthropic’s chief product officer, Mike Krieger, whose boss, Dario Amodei, predicted last March that AI would be writing “essentially all of the code” within a year, suggested that’s in the neighborhood of coming true—at least at Anthropic: “Right now, for most products . . . it’s effectively 100%.”

Along with potentially upending the entire tech industry, AI’s ability to write programs could have a powerful democratizing effect on how the world uses technology. For the past few decades, most people who use computers have been wholly dependent on software written by trained professionals. What happens when that trained professional might be an algorithm, available to the masses to create whatever pops into their minds?

I’ve been exploring that question since last March, when I used Replit to bring my dream note-taking app to life. The experience was amazing enough that I put up with Replit’s many rough edges, including its iffy debugging skills, repeated introduction of security flaws, and sycophantic tendency to tell me my ideas were pure genius. Since then, I have had better luck with new and improved versions of the service. I’ve also dabbled with several other coding platforms with increasingly impressive results.

But Claude Code, which I’ve been using recently to reimagine a game I wrote back in high school, is the most uncanny of them all. As a lark, I fed it my 1980s BASIC code, expecting it would have no clue what to do with something written in such an obsolete language. Instead, it roughed out a modern, web-native version in minutes. Since then, we—Claude Code and I—have been collaborating to improve the game and dress up its graphics.

I say “we” because it truly feels like we’re working as a team. Claude builds out my ideas without me having to spell them out in excruciating detail, and sometimes comes up with ones of its own. Its ability to understand what I want the game to do, and why, can feel like it borders on the clairvoyant. When I’ve finished fooling around with the new version—soon—I’ll share it here so you can judge the results for yourself.

(Full disclosure: I had one bizarre issue with Claude Code. For a few days, it labored under the mistaken understanding that some of my requests were examples of prompt injection—a nefarious third party issuing commands meant to interfere with the project—and kept assuring me that it was ignoring them. Despite that, it continued to code up a storm. I asked Anthropic what was going on, but the company hasn’t yet provided an explanation.)

Quirks and all, I’m thoroughly enjoying making AI-generated software. But I do confess that it’s brought out my inner Edsger Dijkstra. A celebrated computer scientist and A.M. Turing Award winner, Dijkstra bristled at the notion that anyone should be able to create software. He maintained that proper programming required an especially deep understanding of mathematics. Mere mortals shouldn’t even try.

In a 1975 essay, Dijkstra ripped into BASIC, the language I used to write the original version of my game. Created at Dartmouth in 1964 and initially intended for non-techie liberal arts majors, BASIC emphasized approachability over elegance. Instead of demanding too much from these neophytes, it was simple to learn and tolerated sloppy code. He hated it.

As someone who once programmed a fair amount but allowed my skills to atrophy, I am nagged by the fear that vibe coding is a form of cheating. It feels too easy. I’m also bothered by the fact that I don’t fully understand the code Claude wrote, and in fact have barely glanced at it. In short, I haven’t been entirely comfortable with the prospect of software becoming something that anyone can make.

Dijkstra, who died in 2002, isn’t around to chime in on Claude Code or other forms of vibe coding. I can’t imagine he’d be thrilled with them, though. In many cases, their algorithms seem to settle for the most expedient approach to a job, resulting in software that may be less than optimal even if it gets the job done. I cheerfully admit to being unqualified to judge Claude’s coding proficiency, but my high school programming buddy Charles, who went on to become a professional developer, took a peek and deemed some of its techniques “cringe-worthy.”

Legitimate reasons exist to be skittish about the quality of vibe-coded software, particularly on the security front. Last week, an app called Moltbook—a social network for AI agents—made quite a splash. According to security firm Wiz, it also left its database of user information vulnerable to leaks, due to a misconfigured server. Vibe coding may have been to blame.

My reluctance to be responsible for assuring other people’s privacy is the biggest reason why I haven’t shared any of the productivity apps I’ve vibe coded for myself. Presumably, software companies with human engineers in the loop—such as Nvidia and Anthropic—have charged them with vetting the robustness of AI’s handiwork. It’s tough to imagine the day coming when that isn’t essential.

Still, I am slowly getting around to the belief that vibe coding is not an alternative to coding, but a legitimate form of it. Even the most gifted programmer typically needs help translating their work into something a computer can understand. Most of them rely on high-level programming languages that break tasks into the reduced set of low-level instructions a processor performs natively.

Until now, those high-level languages have had names such as Python, JavaScript, Swift, and C++. Thanks to remarkable tools such as Claude Code, they can now have names like “English.” I’m looking forward to seeing what happens once the floodgates break wide open.

You’ve been reading Plugged In, Fast Company’s weekly tech newsletter from me, global technology editor Harry McCracken. If a friend or colleague forwarded this edition to you—or if you’re reading it on fastcompany.com—you can check out previous issues and sign up to get it yourself every Friday morning. I love hearing from you: Ping me at hmccracken@fastcompany.com with your feedback and ideas for future newsletters. I’m also on Bluesky, Mastodon, and Threads, and you can follow Plugged In on Flipboard.

Vibe coding is coding, period

More top tech stories from Fast Company

Read also

Predicted Tottenham XI for Man United clash at Old Trafford as Van de Ven starts

‘People We Meet on Vacation’ brings charming rom-com magic to the screen

School vouchers, rooted in segregation, don't belong in Illinois

Sports today

All sports news today

Sports in Russia today

Friends of Today24