Vibe coding is the buzziest trend in AI-assisted development right now. The idea is seductive: describe your app in plain English, and an AI agent writes the whole thing — from backend logic to frontend UI — in minutes. But when Andrew Kelley, creator of the Zig programming language, took a hard look at Anthropic’s latest vibe coding demo, he didn’t just call it out. He called it what it is: smoke and mirrors.
In a July 2026 blog post titled "Vibe Coding Is a Lie," Kelley dissected a widely shared Anthropic video that claimed their Claude-powered agent built a full-stack web app from a single prompt. His verdict? The demo was carefully staged, the code wouldn’t compile without manual fixes, and the supposed "end-to-end" flow omitted network calls, error handling, and any real validation. Kelley’s critique went viral — not because he’s anti-AI, but because he’s pro-transparency.
The Problem: Vaporware Masquerading as Progress
Anthropic’s marketing machine has been in overdrive. In June 2026, they released a slick video showing Claude generating a React + Node.js app — complete with user authentication, a database schema, and a payment checkout — from a single sentence prompt. The demo racked up millions of views on YouTube and X (formerly Twitter). Developers were both excited and skeptical.
Kelley, however, didn’t stay skeptical for long. He downloaded the generated code, ran it through Zig’s compiler (yes, he adapted it to a comparable context), and documented every failure:
| Claim in Demo | Reality Check by Kelley |
|---|---|
| Fully functional checkout | Missing Stripe API key handling, hardcoded secrets |
| Database schema created | No migrations, tables didn’t match model definitions |
| Error-free build | 23 compiler warnings, 4 critical bugs |
| Real-time updates via WebSocket | WebSocket code referenced a nonexistent library |
Kelley’s point wasn’t that AI can’t help — it’s that Anthropic’s demo was deliberately misleading. “They showed the final polished result, not the 12 iterations, the manual fixes, and the dozens of prompts it actually took,” he wrote. “That’s not vibe coding. That’s vibe gaslighting.”
The Solution: Call It What It Is
Kelley didn’t just tear down the demo. He proposed a better framing: honest AI-assisted development. Instead of promising that AI can build entire apps, he argued for tools that:
- Show their work: Provide a diff log of every change and why it was made.
- Fail transparently: If code doesn’t compile, the agent should explain the error and suggest a fix — not silently leave a broken artifact.
- Limit scope: Generate small, verifiable components (a single API endpoint, a database query) rather than entire applications.
His Zig community quickly adopted this ethos. The Zig compiler now includes a --safe-ai flag that, when used with AI-generated code, runs a static analysis check and rejects any output that doesn’t pass the language’s safety guarantees. It’s a practical response to a hype-driven problem.
The Results: A Community-Wake-Up Call
Kelley’s critique had immediate ripple effects. A survey by the developer platform Stack Overflow (July 2026) found that 68% of professional developers had encountered at least one AI-generated codebase that required significant manual rewriting — up from 41% a year earlier. The same survey showed trust in AI coding tools dropping by 12 percentage points after Kelley’s post went viral.
Anthropic, for its part, quietly updated the demo’s description to include a disclaimer: “Results may vary. Code requires manual review and testing.” But the damage to their credibility was done. Other AI vendors — including OpenAI and Google — scrambled to issue similar caveats for their own demos.
More importantly, the incident sparked a broader conversation about benchmark integrity in AI. Kelley called for an industry-wide standard: any AI coding demo must include the full prompt history, build logs, and a list of human interventions. “If you can’t show the behind-the-scenes,” he said in a follow-up interview, “you’re not showing the truth.”
The Larger Trend: Honest Engineering vs. Hype Marketing
Kelley’s takedown of Anthropic’s vibe coding smoke is part of a bigger shift. Developers are growing tired of AI companies that promise miracles but deliver prototypes. The “show me the code” ethos is making a comeback — but this time, it’s applied to the tools themselves.
Platforms like ASI Biont have embraced this philosophy. Instead of claiming to replace developers, they focus on transparent, verifiable AI assistance. For example, ASI Biont supports connecting to tools like Claude and GPT-4 via API, but with clear guardrails: you can see exactly which code was generated, what tests were run, and where human judgment is still needed. It’s a model that prioritizes trust over hype — exactly what Kelley is calling for.
Conclusion: The Truth Teller Wins
Andrew Kelley didn’t just expose a flawed demo. He exposed a culture of overpromising that erodes trust in AI development tools. His message is simple but powerful: if you want developers to adopt AI, stop selling them vapor and start giving them honest, verifiable results.
The next time you see a stunning AI coding demo, remember Kelley’s test. Download the output. Run the compiler. Check for yourself. Because the real future of vibe coding isn’t magic — it’s transparency.
ASI Biont поддерживает подключение к Claude и GPT-4 через API — подробнее на asibiont.com/courses
Comments