The AI Building Stack: A Map of the Conversation
In February 2025, Andrej Karpathy posted a single paragraph on X about "vibe coding" — building software by describing it in plain language and letting AI write the code while you "forget that the code even exists." Collins Dictionary named it Word of the Year. By October, Karpathy had already moved past the term to something he calls agentic engineering.
That nine-month arc — from coinage to evolution — tells you something about the pace of this conversation. The people having it are thinking in real time. The map keeps shifting.
This is the map as of February 2026.
The Intellectual Lineage
Three people define the poles of this conversation.
Andrej Karpathy coined the term and set the optimistic frame: AI has gotten good enough that you can just describe what you want. The 2023 claim that "the hottest new programming language is English" has proven more literal than figurative.
Simon Willison — Django co-founder, one of the sharpest technical writers working today — immediately drew the line that everyone needs to understand: vibe coding does not mean "using AI to help write code." It specifically means generating code without caring about what was produced. His archive on vibe coding is the most complete, honest account of what the practice actually involves and where it breaks down. His golden rule for anything that matters: "I won't commit any code to my repository if I couldn't explain exactly what it does to somebody else."
Boris Cherny, head of Claude Code at Anthropic, sits between these poles — building the tool that makes both possibilities real, thinking carefully about what it means for everyone. His three essential appearances: the Lenny's Newsletter interview, the Every.to podcast, and the Latent Space episode with swyx.
The CEOs Who Build
The most interesting signal in this conversation is not the researchers. It's the executives who started doing it themselves.
Sebastian Siemiatkowski, CEO of Klarna, has no technical background. He uses Cursor to build product prototypes in approximately 20 minutes before involving his engineering team. The point is not the speed — it's that he arrives in those meetings already knowing what he wants, because he built it. He describes himself as "obsessed."
Sundar Pichai builds personal dashboards with Replit on weekends. His take: "It's exciting to see how casually you can do it now." This is the CEO of the company where AI now writes more than 30% of all code.
Sahil Lavingia, CEO of Gumroad, runs $10M ARR with one employee. 41% of the company's code is AI-written, targeting 80%. His key finding: your codebase architecture determines your AI velocity — not your prompting skill.
These are not demonstrations. These are working habits from executives at operating companies.
The Warning That Matters
Michael Truell, CEO of Cursor — the AI coding tool that just raised at a $29 billion valuation — issued the most credible warning in the space:
"If you close your eyes and you don't look at the code and you have AIs build things with shaky foundations as you add another floor, and another floor, and another floor, things start to kind of crumble."
This warning is credible precisely because it comes from inside the space. He's not arguing against AI-assisted coding. He's arguing against building without understanding what was built.
Willison amplifies this with data: roughly 45% of AI-generated code contains security flaws, with cross-site scripting errors appearing in 86% of cases tested. This is the production systems problem — and the resolution is the same one this guide argues for: personal tools under your own supervision are different from systems other people depend on.
The Practitioners Who Document
Two non-technical practitioners have done the work of documenting their actual workflows publicly.
Matt Stockton — a consultant — is the most detailed account of someone building a practice, not just an experiment. Voice-to-transcript workflow, CLAUDE.md as institutional memory, the compound effect of captured context. His insight: "If you capture what you did in a structured way, Claude can use that information later." The information compounds.
Obie Fernandez built what he calls a Personal CTO Operating System: meeting transcript automation, hiring pipeline management, engineer history tracking. The executive's specific problem — too much to track, too little time — handled through personal automation.
The Bridge-Builders
swyx runs Latent Space, the newsletter and podcast where the AI engineering community talks to itself. His community-building work is what makes the conversation accessible to people who didn't start inside it. The Claude Code episode with Boris Cherny is the technical foundation.
Lenny Rachitsky at 900K subscribers is where the product and executive world goes for signal. His "How I AI" series is the most consistently useful executive-level AI workflow content published regularly. The Cherny and Lavingia interviews live here.
Dan Shipper at Every.to writes from inside the practice — using AI tools to think, write, and build — and documents what he finds. Every.to runs on 15 people at meaningful revenue, largely because they use what they write about.
The Key Tensions
The conversation is structured around a few unresolved questions. Understanding them is how you contribute rather than just consume.
Vibe coding vs. agentic engineering. Karpathy invented the term and moved past it in nine months. The serious practitioners now call the disciplined version agentic engineering — full AI agent use with professional responsibility for the output. The term "vibe coding" has become associated with the irresponsible version.
Build fast vs. build sound. Siemiatkowski in 20 minutes vs. Truell's crumbling foundations. The resolution: these describe different scopes. Personal tools under your own supervision can be built fast. Anything that other people depend on must be built sound. The line is not about the tool — it's about consequence.
Democratization vs. security risk. Every security researcher is concerned. Every executive advocate is excited. Both are right in their domains. The safety zone is real: personal, controlled, low-stakes use cases. The danger zone is real too: production systems, customer data, authentication.
Where to Go
The complete resource index — everyone worth following, every outlet worth reading, every community worth joining — is in the Claude Code for CEOs guide.
Seven pieces to read in order, if you want the full picture:
- Karpathy's original post — the founding moment
- Willison: Not all AI-assisted programming is vibe coding — the necessary distinction
- Willison: The Perils of Vibe Coding — honest risk picture
- Cherny: Every.to transcript — how the builder thinks
- Lavingia: Lenny's 40x playbook — CEO practitioner case
- Stockton: Claude Code for Non-Technical Work — daily workflow documentation
- Willison: Beyond Vibe Coding — where it goes
The conversation is live and moving fast. What stays stable is the tension between empowerment and responsibility — building more because you can, and staying honest about what you built.
— Ormus
