1.

Andrej Karpathy told the world to stop writing code and "just vibe with the AI." Now a podcaster ships a clip generator before lunch, a freelancer deploys a chatbot without touching a terminal — and 75% of new apps may be built this way by 2026. The catch: the people building this software are the least equipped to know when it's dangerous.

2.

Describe what you want; AI writes every line

• Karpathy's workflow: prompt the AI, run the output, see what breaks, refine with follow-up prompts, repeat — no syntax, no spec document, no boilerplate
• Two modes exist: all-in-one platforms where you never see code (Hostinger Horizons) and AI coding assistants (ChatGPT, Copilot, DeepSeek) where you work with snippets directly
• The shared loop — prompt → run → debug → repeat — replaces weeks of wireframing and architecture with a conversation

3.

Non-technical creators ship working tools in hours

• A podcaster built a clip generator with AI, a creator launched a web form that auto-posts to Notion, a freelancer deployed a FAQ chatbot — none wrote code
• A Hostinger colleague went from blank screen to live web app in under one hour on camera
• The pitch delivers for small, bounded problems: personal tools, MVPs, community projects where "it works" matters more than "it scales"

4.

The audience that needs audits can't perform them

• AI-generated code skips input validation, authentication checks, and error handling — the article itself warns it "may be vulnerable and should not be implemented without testing"
• The prescribed fix — review code manually, run audits, use sandboxes — demands exactly the expertise vibe coding promises you don't need
• Debugging advice is circular: "ask the AI to explain the bug before fixing it" assumes the system that generated the error can reliably diagnose it — a known LLM failure mode the article never acknowledges

5.

Adoption is surging faster than quality controls

• By 2026, an estimated 75% of new applications will use no-code tools — a stat the article equates with vibe coding, though the original source is Hostinger's own content on low-code trends, not independent research
• Indie hacker and maker communities on Reddit and Hacker News have already adopted vibe coding as a default prototyping method
• The article's own hedge tells the story: "great for speed and experimentation, however, not yet for scaling or reliability" — growth is outpacing the guardrails

6.

"Accessible" and "requires audits" cannot coexist

• The article's core promise is that non-technical creators can build working software without learning syntax — then buries a caveat that this software must be manually audited for security vulnerabilities before production use
• This is not an edge case: the entire value proposition targets beginners, freelancers, and solo creators — precisely the users who cannot distinguish a SQL injection vulnerability from a typo
• If vibe coding is only safe when supervised by someone who already knows how to code, accessibility is not a feature — it is a marketing claim that collapses under its own logic

7.

Safety is a tool problem, not a method problem

• Both the article's optimism and the counter-argument's skepticism share one hidden assumption: that the AI assistant is a fixed-quality tool, frozen at 2025 capability
• Spreadsheets followed the same arc — dangerous in untrained hands in 1985, ubiquitous and self-correcting by 2005 — not because users got smarter but because the software absorbed the expertise
• The real question is not whether vibe coding is safe today but whether the tools will close the security gap faster than adoption widens it — and the answer determines whether you should build with vibe coding now or wait