Vibe Coding Isn't Dying — Here's What Karpathy Actually Said
Every few weeks a new YouTube essay shows up in my recommendations with a title like "Vibe Coding Is Dead" or "Why Karpathy Just Killed Vibe Coding" or "The End of the Vibe Coder." The thumbnails are red. The takes are confident. The comments are full of developers feeling vindicated.
I watch them between debugging sessions on the production tools I am, you know, currently vibe coding. The dissonance is funny enough that I want to write about it instead of just rolling my eyes.
Here is the short version: the takes are misreading what Karpathy actually said. He did not declare the practice dead. He said the term has been diluted past the point of usefulness — a completely different claim. And the work itself, the actual loop of describing software in language and shipping it, is not slowing down. It is growing. What is dying is the lazy version of it, the one that fits in a tweet. What is growing is the serious version, which most people writing obituaries have never tried.
This is a commentary post, not a tutorial. I want to put the record straight, then explain why I — a non-technical solo founder running the internal tools and agents for two brands through this exact workflow — am not worried at all.
What Karpathy actually said
The original "vibe coding" framing came from Karpathy in a February 2, 2025 post on X, where he described a new kind of coding in which you "fully give in to the vibes, embrace exponentials, and forget that the code even exists." That post was the spark. Collins Dictionary named "vibe coding" its Word of the Year for 2025, which is the kind of thing that simultaneously legitimizes a term and dissolves its precision. Within months, "vibe coding" was being used to describe everything from senior engineers using Copilot autocomplete to teenagers generating fart-noise apps in Lovable.
What Karpathy has been saying more recently — most notably in his Sequoia AI Ascent 2026 talk — is that the term got out from under him. He did not invent it to describe how everyone should build software. He invented it to describe a specific posture in a specific kind of one-off prototype work. When people started applying it to production systems and acting like reading the code was old-fashioned, he started clarifying. The shape of his clarification, across multiple appearances, is roughly: vibe coding is fine for throwaway projects, but if you are building something that needs to run in production, you should be doing agentic engineering with code review and tests, not vibes.
That is a sensible thing to say. It is also not the same as "vibe coding is dying." It is closer to "the word got bigger than the original meaning, and the original meaning was always narrower than how people use it."
The English-language internet does not really distinguish those two claims. So the headline becomes "Karpathy killed vibe coding," and the comment sections fill up, and meanwhile actual builders keep building.
What is actually dying
Something is dying. It is worth being precise about what.
The version of vibe coding that is genuinely on its way out is the one that treats AI as a replacement for thinking. Sit down, type a one-line prompt, accept whatever comes back, ship it without running it, get surprised when it breaks in production. That posture was always going to age badly, and it is aging badly now. The market is filling with apps built that way — fragile, generic, leaking API keys, three weeks from being deindexed by Google for thin content. Twitter feeds are full of post-mortems from people who shipped twenty AI-generated SaaS products in a weekend and watched them all crash within a month.
That version deserves to die. It was never a viable way to run a business. It was a content category, optimized for screenshots of revenue dashboards on day one, with no plan for day thirty.
The other thing that is dying is the meme aesthetic — the "I do not read the code I just vibe" posture as an identity. That was a useful provocation in early 2025 when the gatekeepers needed shaking. It is a less useful posture in 2026 because everyone who could be shaken has already been shaken. The next phase is not provocation. It is practice.
So when I see "vibe coding is dying" content, I almost always agree with the narrow version of the claim. The lazy version is dying. The meme version is dying. Good. That clears the field for the version that matters.
What is actually growing
What is growing is the serious practice underneath the meme. The work where you describe software in language, iterate with AI, ship production code, and maintain it through the same conversational loop — but with discipline. Tests where they matter. Code review (often AI-on-AI, but real). Security audits. Monitoring. Actual operational hygiene.
That practice is increasingly called agentic engineering — a term Karpathy himself proposed as the replacement for vibe coding in production work — and it is the natural evolution of the original posture. Same loop, different seriousness. The AI is not just generating one-shot code; it is operating inside a structured workflow with planning, execution, verification, and review steps. The human role is shifting from "type prompts" to "design systems and judge outputs." The deeper version of this argument is in The Future Is Agentic Engineering — Not Just Vibe Coding, so I will not relitigate it here. The point for this post is: the people declaring vibe coding dead are mostly pointing at the chrysalis and refusing to notice it is hatching.
The market data backs this up rather than contradicting it. Anthropic, OpenAI, and the platform companies (Vercel, Replit, Lovable, Bolt) all keep growing through 2026. Claude Code, Cursor, and the new generation of agentic IDEs are getting more capable every release cycle, not less. The 2025 Stack Overflow Developer Survey found 84% of respondents are using or planning to use AI tools, with 51% of professional developers using them daily — up from "experimental minority" status only two years earlier. If the practice were dying, none of that would be happening.
What is also growing — and this is the part I notice most because it is where I live — is the population of non-technical builders who are quietly shipping real products this way. Most of them are not on Twitter. They are not making YouTube takes. They are not arguing with developers about whether their workflow counts as "real" coding. They are running newsletters, courses, consultancies, agencies, software products, and small SaaS tools. They paid for a Claude subscription, they pay for Cursor or Lovable, and they are extracting more value per dollar than almost anyone else in the economy right now.
Get essays like this — plus behind-the-scenes builds — in your inbox every week. Subscribe free →
I am one of them. I know dozens more. We do not care whether the meme is dying.
My view as the person living inside this
I do not have a computer science degree. I have never worked at a tech company. My most technical skill three years ago was a nested IF formula in Google Sheets. Today I run Soulin Social (content multiplier, social.soulin.co), the KINS Sales Agent (sales.soulin.co), and a personal LifeOS, all in production, all maintained alone, all built primarily through the loop the takes claim is dying.
When I read a "vibe coding is dead" post, I check three things on my own dashboard before I take it seriously.
First, did the products break this week because the practice is no longer viable? No. They broke the way software always breaks — a Stripe webhook out of order, a Supabase policy I forgot to update, a Vercel environment variable that needed rotating. I fixed all of them the same way I always do, by describing the symptom to Claude and iterating until the fix was correct. The practice still works exactly the way it did six months ago, except the models are better.
Second, did I get less productive? No. The combination of Claude 4.6, Claude Code, and the workflow disciplines I have built around them (a CLAUDE.md file per project, a content engine, a QA checklist) means I am shipping at a faster rate than I was a year ago, with fewer bugs reaching production, on a stack that costs me roughly one hundred dollars a month to operate.
Third, are the people making this content actually building anything? Sometimes yes — those are the takes I take seriously, even when I disagree. Often no — those are the takes that read like a developer cope, dressed up as analysis, monetized on a thumbnail.
The dying claim only makes sense if you assume "vibe coding" means the worst, laziest, most parodic version of the practice. If you take it to mean the actual workflow that thousands of solo founders use to run real businesses now, the claim falls apart immediately. The numbers do not support it. My products do not support it. The other builders I talk to weekly do not support it.
What it actually means for non-technical solo founders
If you are reading this and you have been on the fence about whether to invest time in this workflow because the death takes spooked you — please do not let them.
The death of a meme is not the death of a practice. The meme was always thinner than the work. The practice — describe software clearly, iterate with AI, ship and maintain in production, build the debugging muscle, treat the AI as a collaborator and not a magic box — is more viable in 2026 than it was in 2025, which was more viable than 2024. The trajectory is monotonic. The hype cycle just makes it feel jagged.
What you should do, if anything, is upgrade your seriousness. Start treating the work like work. Use a system prompt file (a CLAUDE.md or AGENTS.md) so the AI has context every session. Use version control. Run things before you ship. Have one repeatable QA routine. Take security seriously — Veracode's 2025 study found AI introduced security vulnerabilities in 45 percent of coding tasks across 100+ models, which is fixable with discipline but not ignorable. Read the deeper post on agentic engineering and let it shape how you build, not whether you build.
The "vibe coding is dying" takes are a permission slip to stop building. Do not take it. The people writing them are not building the things you would build. They are not solving the problems you have. They are commenting on a meme that was always too small a container for the actual practice.
I am at my desk in Berlin. The products are running. The code is shipping. The meme can do whatever it wants.
I write about freedom, healing, and building alone. The full archive is at soulin.co.