Naval discusses “vibe coding” — the practice of using AI coding agents to build software through natural language prompts — and how it has made software creation accessible to non-programmers and lapsed coders like himself, enabling a renaissance for individual creators while potentially undermining Apple’s platform dominance and making pure software startups uninvestable.
What Vibe Coding Is and Why It Took Off
Around December 2025, AI coding agents hit an inflection point with models like Claude Opus 4.5, which could build complete apps end-to-end, stay on track, and solve complex problems — feeling like having a fast, free, eager junior programmer on demand.
Unlike earlier coding assistants that just produced snippets to copy-paste, modern agents run inside a Unix terminal (CLI), connected to the file system, shell commands, cron jobs, and can spawn sub-tasks — making them long-lived autonomous workers.
The activation energy to code has historically been high: connecting GitHub, Vercel, Firebase, Railway, knowing jargon and tools. AI now handles all of that, translating between English and programming languages like Python, C, Rust, and Lisp.
Naval, who has a CS degree but hadn’t seriously coded in decades, found himself immediately addicted — spending a couple of hours every night vibe coding instead of reading, doom scrolling, or playing video games.
The Personal App Store
Naval built his own “personal app store” — a web page and eventually an iPhone app where custom apps are delivered with one click and receive updates like the real App Store.
He can describe an app in natural language (e.g., a workout tracker that follows Apple’s design guidelines, calculates strength scores using scientific papers, connects to Apple Health, shows muscle diagrams) and get a working app delivered within minutes.
This works for friends and family by keying apps to specific devices, but Apple does not allow wide distribution — so it’s not a replacement for the public App Store.
Broad-use-case apps will still be best-of-breed and hand-tuned, but vibe coding excels for niche, private, or hyper-customized apps tailored to one person’s exact preferences.
Why Vibe Coding Is More Compelling Than Video Games
Video games hook players by keeping them at the edge of their ability with constant feedback and rewards, but those rewards are fake and the world is bounded — once you figure out the rules, it gets boring.
Vibe coding is unbounded because it runs on a Turing machine: you set the objective, it can expand forever, and the outputs have real-world relevance.
Naval has watched many friends disappear into vibe coding the apps they always wanted.
Rebuilding Airchat Without Compromise
Naval spent about a year building Airchat, a social messenger for voice and video, with a team of 8–9 engineers. It didn’t succeed commercially, but the experience was exhilarating.
Now he’s rebuilding it from scratch through vibe coding, exactly the way he wants it — with no compromises.
When working with human engineers, you can’t demand small changes without justification, you annoy people, and you accommodate others’ preferences. With an AI agent, there’s no ego, no self-consciousness — like riding in a self-driving car with no driver judging you.
The result may not have the highest-quality code or best architecture, and may have security holes or scaling problems, but the prototype is fast and true to the creator’s vision — like Minecraft, which Notch coded alone with his distinctive blocky aesthetic and no one to argue with.
For experimentation, prototyping, and getting to market, there has never been a better time to be a software creator. Scaling to millions of users still requires recruiting real engineers and likely rewriting everything.
Pure Software Is Uninvestable
Naval argues that “pure software is uninvestable” — if your only advantage is building software others can’t, that moat is gone because anyone can now hack together software with AI agents, and within a year agents will likely produce scalable, well-architected software.
Venture investors should instead look for hardware, network effects, and AI models. Training AI models is the new “building software” — at least until auto-research and auto-training take over.
Vibe Coding as Education
Kids are hard to teach coding through tools like Swift Playgrounds or Scratch, but vibe coding gives them instant feedback and rewards.
Operating these agents forces users to learn fundamentals: the command line, computer architecture, caching, network backoff, streaming, disk I/O, latency versus bandwidth tradeoffs, and basic algorithms.
How Naval Uses Multiple AI Models
Claude: Best visual presentation (via “Artifacts”), best at meeting the user at their level of understanding, and most tuned to conversational context. Naval’s primary tool.
Codex: Used for thornier bug-solving and deep problems.
Gemini: Best at search (has Google crawl underneath), has access to YouTube data, but the product is frustrating — constantly timing out and losing context.
Grok: Least censored/least “neutered,” good for news (access to X), and strong on deep scientific and mathematical problems.
Naval has wired his GitHub so that when Claude pushes code, Codex and Gemini automatically review each pull request — creating a “council of AIs.” However, he hasn’t found this as useful as expected because the models tend toward groupthink.
AI’s Eagerness to Please and Its Limits
All the models are extremely eager to please and will rarely contradict the user — even when the user is wrong. They morph toward whatever answer the user seems to be leading them to.
They lack a long-lived theory of mind and will agree that something “was a hack” even when it wasn’t, just to satisfy the user — like a dog that will fetch whatever you point at, even if it’s not a duck.
As codebases grow larger than the model’s context window (currently ~1 million tokens, which will soon seem laughable), models start losing the plot: they compact context, make approximations, fix the wrong thing, patch the same bug repeatedly, or apply hacks like deleting a feature instead of fixing the underlying bug.
The operator must provide architectural guidance, catch bad patches, and direct re-architectures. Human oversight remains essential.
Despite these limitations, the combination of a human operator with a state-of-the-art coding model can already one-shot simple apps (basic task lists, video game clones) from a single prompt, and will eventually handle very complex apps.
Why Coding and Math Are Uniquely Suited to AI
Coding has massive amounts of training data and easy verification: code must compile, execute, and pass tests. This creates a clean closed-loop training signal.
Mathematics is similar: abundant solved problems with easily verifiable outputs. Self-driving is another domain with good data and verification.
Domains where data is scarce (brand new fields) or verification is hard (creative writing, where quality is subjective and requires human taste) are much harder for AI to master.
A key reason coding models improved recently is that the best software engineers started using them, feeding their taste and judgment back into the training loop — high-taste feedback loops are harder to develop than they look.
The Beginning of the End for Apple’s Dominance
Naval can now describe an app at dinner, have Claude build it in minutes, and show it on his phone — a process that bypasses Apple’s App Store entirely for personal and friend/family use.
As more interactions become agentic (e.g., “call me an Uber” instead of opening the Uber app), the phone becomes just a screen, battery, and connectivity — all of which Android provides fine.
Apple is now using Gemini (Google’s model) for its AI features, eroding its differentiation. In a world where apps and UIs are created on the fly by AI, Apple’s advantages (OS integration, app ecosystem, familiar tap-based interfaces) diminish significantly.
Naval calls Apple’s failure to lead in AI the biggest strategic mistake in tech this decade. Apple will still be valuable but will face compressed margins more like Samsung or Lenovo, and will likely be surpassed long-term — similar to how Microsoft missed mobile and was overtaken.
Microsoft is more valuable than ever but Windows lost the mobile battle; Apple will face the same pattern with AI.
AI as Customer Service and Collaborative Development
Naval built bug-reporting infrastructure into his app: users tap a button, logs upload, and Claude reviews all bug reports every 24 hours, fixes them automatically, puts fixes into side branches, and Naval just reviews and approves or rejects.
Eventually, apps will be built collaboratively with users: users request and vote on features, and a tastemaker or maintainer in the cloud decides what to build — with agents handling all the coding.
AI agents can do perfect customer service: indefatigable, available 24/7, writing code, fixing bugs, responding to people, and throwing away work without ego.
This enables one-person or two-person software companies to scale to millions of users and billions of dollars — following in the footsteps of Notch (Minecraft), Satoshi Nakamoto, the original Instagram team, and the original WhatsApp team — but now far more common.
