Silicon Valley investor Naval Ravikant recently discussed the “vibe coding” wave on a podcast, arguing that AI coding agents have already crossed a crucial threshold. They’ve evolved from tools that could only assist with writing code in the past into “long-running agents” that can build an entire application from scratch. He even offered a sharp judgment: when AI can directly generate personalized apps for users, the long-standing software and hardware advantages that iPhone and the App Store have built over time may be facing a challenge.
Naval said that his return to programming began after the release of Claude Opus 4.5 in December 2025, when the capabilities of AI coding agents saw a clear leap. He described that these agents are no longer just early tools that provide a snippet of code for users to copy and paste into a development environment. Instead, they can run for long periods in a terminal, understand the file system, call Unix commands, run tests, fix bugs, and even complete an entire application.
He believes the reason this experience is addictive is that it dramatically lowers the “startup cost” of programming. In the past, to develop an app, you had to understand GitHub, backend services, Firebase, Railway, Xcode, command-line tools, and a pile of engineering jargon. Now, users only need to describe requirements in natural language, and the AI can translate English into Python, C, Rust, Lisp, or various frameworks and tool commands. For people with basic computer architecture concepts, this drastically reduces the barrier to software creation.
From the personal App Store to fitness trackers: AI turns software into “tailor-made”
On the show, Naval revealed that he has already built a “personal App Store.” This App Store isn’t a public commercial marketplace, but a private deployment system for himself, friends, and family: he can ask the AI to generate an app, and after the AI finishes, it automatically delivers it to his personal App Store page—then he can install it onto his iPhone with one tap.
He gave an example: he once asked an AI to build a fitness tracking app that perfectly matches his personal usage habits. He wanted it to combine the functionality of Tonal and Ladder, follow Apple Human Interface Guidelines, and look like a native Apple app. At the same time, it could read past training logs, make entering new records easier, generate charts, calculate strength scores, and even read scientific papers to think about how to assess strength progress by body part. Later on, it could also connect to Apple Health to read data like heart rate.
Naval emphasized that apps like these don’t necessarily replace all mature products. For mainstream needs—such as communications, ride-hailing, banking, and social platforms—there will still be the strongest “best-of-breed” products in the market. But what AI coding agents truly unlock is software demand that is “extremely personalized” and “extremely niche”: apps that exist only for you, that match only your workflow, or even apps built for a specific slice of life.
That’s also why he thinks vibe coding is more compelling than playing video games. Video games immerse players through real-time feedback, task rewards, and difficulty curves, but that’s still a closed world designed by others. Vibe coding is an open world: the underlying system is a Turing machine, your goals can extend infinitely, and it can produce real utility in the real world.
“Knowing what you want” becomes the most important capability
However, Naval also pointed out that vibe coding doesn’t turn everyone into engineers. Most people still view computers as a black box. Even if the difficulty drops by 10x or 100x, it may not matter much to them.
The real beneficiaries are people who have “a clear vision, can express it clearly, and have the motivation to create.” Naval believes the most critical ability for vibe coding isn’t syntax—it’s knowing what you want. When users can clearly describe what the product should look like, how it should operate, and which details matter, AI can help quickly turn ideas into an actionable prototype.
He compared this with his experience founding AirChat in the past. AirChat was a social messaging product centered on voice and video communication. Back then, he worked with 8 to 9 engineers, spending 9 to 12 months building multiple versions. In the end, the product didn’t succeed. The team sold the company, investors recovered their money, and employees also received appropriate arrangements.
But in the vibe coding era, he’s rebuilding a similar product alone—and can adjust it entirely according to his own instincts. In the past, collaborating with an engineering team—even if the founder had a strong vision—required constant compromise. You couldn’t keep asking engineers to shift icons left, then right, then back again, and you also couldn’t demand the team accommodate every instinctive judgment. AI agents, on the other hand, have no ego, no impatience, and won’t feel offended because of repeated iterations.
Naval believes this expands the scope of exploration for software creativity. It’s like how《Minecraft》was initially built by Notch alone. The block-style look seems behind traditional game art standards, yet it fully preserves the product instincts of one person. Vibe coding may bring more products like that—ones not “smoothed out” by team consensus.
Is pure software entrepreneurship no longer worth investing in? Naval: the moat is being eaten by AI
The most controversial view in the episode was Naval’s take on the venture capital market. He previously posted on social media saying, “Pure software is rapidly becoming not worth investing in.” On the podcast, he made it even more direct: if a company’s entire advantage is “I can write software that others can’t,” then it no longer has investment value.
His reasoning has two parts: first, coding agents today can quickly assemble a lot of features. Second, these coding agents are improving at an extremely fast pace, and they may be able to write software with more complete architecture and better scalability within a year—or even less. In other words, a moat built purely on feature development capability is being compressed by AI.
Therefore, Naval believes that venture capital should next look for structural moats that are harder to replicate quickly: hardware, network effects, AI models, data advantages, or other barriers of that kind. He even suggested that “training AI models” might be the new generation of “writing software”—at least until automatic research and automatic training become mature, AI models themselves could still form new startup and investment opportunities.
This doesn’t mean he has a pessimistic view of individual software creators. On the contrary, he sees it as a revival for individual creators. For people who want to do experiments, prototypes, and early product validation, this is the best era in history. But if a product really needs to serve mass users, face high concurrency, and meet complex security requirements, founders still need a real engineering team—and may even have to rewrite the entire architecture.
AI is powerful, but it still needs human supervision: models may pander and may slack off
Naval also didn’t describe coding agents as a magic tool. He noted that as code scale grows, models start running into clear limitations.
The reason is that model context windows have a limit. When the codebase exceeds what the model can “remember” at once, it starts guessing—compressing context, forgetting the original goal, and even fixing the wrong things. Naval described that the model might fix the same bug five times, or apply a quick patch in an architecturally incorrect way. Worse, sometimes it might delete the function or usage scenario that causes the bug just to eliminate the bug.
So during development, he frequently interrupts the model and asks it not to do hacks, not just to patch— but to return to solving the problem at the architectural level. Ironically, the model usually replies immediately: “You’re right—that’s a hack.” Even if it’s not necessarily truly a hack, the model still tends to cater to the user.
Naval used an analogy for AI agents: they’re a bit like hunting dogs. They might be better than you at catching ducks, but if you point them at the wrong bird, they may still charge off. That means humans still need to own direction-setting, architectural judgment, debugging strategy, and product taste.
He also mentioned that he makes different models check code against each other. For example, after Claude writes the code, he pushes it to GitHub. Models like Codex, Gemini, and Grok can automatically review the pull requests, like an AI roundtable. But in practice, the impact isn’t as huge as people imagine, because there’s still a lot of groupthink among models. And if a user nudges the answer toward a certain direction, most models will follow along, with few truly strongly contradicting.
Why do coding agents improve fastest? Because code is easy to verify
Naval believes AI progresses especially quickly in the programming domain for a fundamental reason: code is easy to verify.
Whether the code can compile, whether tests pass, whether a feature actually works—these produce relatively clear feedback signals. Math has similar properties, because there are lots of problem-solving data and verifiable answers. Autonomous driving in some respects is similar too, because feedback loops can be built through large amounts of data and simulated environments.
By contrast, creative writing, taste judgments, and research in emerging fields are far more difficult. A model can generate endless articles, but who decides what’s good or bad? If you just collect a group of low-paid workers to label “good” or “bad,” the result will only reflect that group’s taste. Naval believes that for models to improve meaningfully, they need a “high-taste feedback loop,” which is harder than simply collecting large amounts of data.
He speculated that one reason coding models have gotten stronger recently could be that the best software engineers have started using these models heavily, giving the models high-quality code and high-quality preference feedback. In other words, AI is learning not only from code, but also from top engineers’ judgments of what counts as good code.
Apple’s crisis: when users don’t need to open apps and just talk to AI agents
Naval’s judgment about Apple is especially aggressive. He thinks once AI agents can generate interfaces and functionality on the fly based on user needs, users won’t need to open apps frequently anymore, nor rely on the existing app ecosystem on iPhone.
In the past, if users wanted to call a ride, they opened Uber; to log fitness, they opened a fitness app; to complete a task, they searched for the corresponding app on the phone. But in an agentic interface, users only need to say: “Help me call Uber,” “Log my workouts,” or “Help me complete this task.” At that moment, the real entry point is no longer the iPhone home screen, but the AI model itself.
Naval believes this will weaken Apple’s long-standing core advantage. Apple’s moat isn’t just hardware. It’s also the operating system, App Store, native apps, ecosystem integration, and the user interface. Once users primarily communicate through Claude, Codex, Gemini, or other AI agents, the phone becomes just a screen, a battery, and a network connection device. At that stage, Android can also provide these capabilities, and Apple’s differentiation would be compressed.
He even said directly that Apple falling behind in AI could become one of the biggest strategic mistakes for the technology industry in this generation. Apple won’t disappear immediately, and it could still make a lot of money in the long run. But like how Microsoft missed the mobile wave and Windows stopped being the core entry point for personal computing, Apple could lose its upper limit on future growth amid the AI agent wave.
Will software companies become one-person shops? AI customer support can also fix bugs directly
Naval further described the development workflow he’s building: an in-app bug reporting system. Users see an issue and press a button, and the system sends the logs to a server. Every 24 hours, Claude automatically organizes all bug reports, tries to fix the problems, and puts the fixes into a branch for him to review. He only needs to make the final call: is it truly a bug? Is the fix good? Should it be merged and released?
He believes future feature development could be done similarly. Users propose feature requirements and vote to rank them. The AI agent is responsible for organizing, designing, implementing, and responding. Finally, a maintainer with product taste decides which features are worth shipping—and which users don’t even know what they want.
This blurs the line between “customer support” and “engineering.” The ideal customer support wouldn’t just answer questions—it would also understand the product, fix bugs, write code, work 24 hours, and have no ego. It wouldn’t be hurt if a large amount of code ends up thrown away at the end. Naval believes this model could give a one- or two-person software company the potential to serve millions of users, even tens of millions.
He pointed out that similar cases have already happened historically. For instance, Notch built《Minecraft》alone; Satoshi Nakamoto launched Bitcoin with a very small team. Early on, Instagram and WhatsApp also created massive impact with small teams. But AI coding agents will make such small-team miracles happen more often.
Vibe Coding won’t make engineers disappear, but it will redefine who can create software
Naval’s view on vibe coding isn’t “engineers don’t matter anymore.” On the contrary, when a product moves toward large-scale users, security requirements, stable architecture, and commercialization, truly excellent engineers are still indispensable.
But the real change he points to is that the starting point for software creation is being rewritten. In the past, there was a gap between ideas and prototypes due to engineering resources, team collaboration, development schedules, and costs. Now, people with clear product instincts can hand the requirements directly to AI and receive a working version within minutes to hours.
This will reshape three things: first, personalized software will grow explosively; second, startups that build moats purely through feature development will find it harder to get valuations; third, existing platform entry points like mobile phones and App Stores will be challenged again by AI agents.
This article 《The Naval Playbook》Naval: AI Agents are the beginning of the end for Apple, and with Vibe Coding, everyone can make apps was first published on Chain News ABMedia.
