Naval on Vibe Coding: Personal App Stores and the End of Apple's Dominance

Introduction

Renowned Silicon Valley investor Naval Ravikant shared his hands-on experience and deep thinking about "Vibe Coding" in his latest podcast. As co-founder of AngelList—someone with a computer science degree who hasn't seriously coded in decades—Naval not only returned to the world of code but built his own personal app store and made bold predictions about Apple's future.

What is Vibe Coding? The term "Vibe Coding" was formally coined by OpenAI co-founder Andrej Karpathy in early 2025. The core idea: developers no longer write code line by line, but instead describe their intent in natural language, letting AI models generate and iterate on code. Developers only need to "feel" whether the program matches their expectations and continuously provide language-based feedback for corrections. This paradigm stands in stark contrast to traditional "precise programming"—where developers must master syntax, APIs, and architectural details—by shifting that cognitive burden to AI, while humans devolve into the role of "product manager + taste gatekeeper." The concept quickly sparked heated debate in Silicon Valley, with some seeing it as a revolution in programming democratization, while veteran engineers worry about the maintainability and security of the resulting code.

The Inflection Point of Vibe Coding

The Transformation Starting with Claude Opus 4.5

Naval pointed out that around December 2025, with the release of Claude Opus 4.5, AI coding agents reached an inflection point. These are no longer simple code assistance tools but genuine "agents"—they can build applications from start to finish, solve thorny problems, like having a fast, free, always-available junior programmer.

Anthropic's Claude model series is known for its "Constitutional AI" training methodology, excelling compared to other models in following complex instructions and long-context reasoning. Claude Opus, as the flagship version of the series, achieved breakthroughs in coding across three dimensions: first, mature Tool Use capabilities allowing the model to natively call external APIs and execute shell commands; second, a longer context window (approximately 200,000 tokens) enabling it to understand large codebases; and third, "Agentic Loop" capabilities—the model can autonomously plan multi-step tasks and self-debug when encountering errors, rather than generating once and waiting for human intervention. This is fundamentally different from early Copilot-style "code completion" tools—the latter is passive assistance, while the former is active execution.

He emphasized a key change: these agents no longer just give you a pile of code to copy-paste into your IDE. Instead, they connect directly to the Unix shell, executing commands, manipulating the file system, calling tools like grep, awk, and sed, running cron jobs for long-running tasks, and even spawning more shells and tasks on demand.

AI agents connecting directly to the Unix shell is a massively underappreciated breakthrough. The core Unix philosophy is "everything is a file" and "composing small tools"—grep for text search, awk for data processing, sed for stream editing, cron for scheduled tasks. When AI agents can natively invoke these tools, they effectively gain programming capability over the entire operating system, rather than being confined to a single programming language's sandbox. This means AI can: read and write arbitrary files, call any installed command-line program, manage processes, and manipulate networks. This combination of capabilities makes "building and deploying a complete application from scratch" possible, rather than just generating an isolated function. Of course, this also introduces significant security risks—an AI agent with shell access that gets hit with malicious prompt injection has far more severe consequences than a code completion tool.

Why AI Is Particularly Good at Programming

Naval explained why programming has become one of AI's strongest domains:

• Massive data volume: GitHub and Stack Overflow contain enormous amounts of code
• Easy verification: Code must compile and execute, and can be validated with pre-written tests
• Closed-loop feedback: Output can be automatically scored by algorithms

There are deeper technical reasons behind this. AI model capability improvements depend on high-quality feedback signals, and code naturally provides the most ideal feedback loop: compiler errors provide precise error localization, unit tests give binary pass/fail judgments, and performance benchmarks offer quantifiable numerical comparisons. This makes "training AI with code execution results" possible—the leaps in coding ability seen in OpenAI's o-series models and Anthropic's Claude are largely thanks to Reinforcement Learning from Execution Feedback (RLEF). By contrast, tasks like creative writing and emotional support lack objective scoring standards and can only rely on human preference feedback (RLHF), which has noisier training signals and worse scalability. Similar to mathematics—large numbers of solved problems plus easily verifiable outputs fundamentally explain why AI's progress in coding far outpaces other creative tasks.

Personal App Store: A One-Person Software Empire

30 Seconds from Idea to App

Naval built a truly stunning workflow: open Claude on his phone, give a two-line description, AI builds the app and pushes it to his personal app store, tap install, and 30 seconds later there's a usable app on his phone.

This workflow technically relies on Apple's Enterprise Developer Certificate or TestFlight beta distribution mechanism. Enterprise certificates allow companies to distribute apps internally