There is a question worth sitting with before the next model drops.
Not “what will AI be able to do?” — that question answers itself, reliably, every few months. The more interesting question is: when production costs approach zero, what does human intelligence actually become?
The AI app-building market is offering a live answer right now, and it isn’t comfortable.
A company called Lovable raised $330 million at a $6.6 billion valuation in early 2026. It lets anyone build a working software application — with real code, real functionality, a real database — through a conversation. No programming required. Users are creating 100,000 new projects every single day.
It is, by almost any measure, an extraordinary product. And the space it operates in is, by the same measures, collapsing.
The reason is straightforward once you see it. Lovable, and most of the companies like it, are thin layers on top of foundation models — Claude, GPT, Gemini. Their actual technology is a user interface. Their moat, the thing that stops a competitor from replicating what they do, is approximately one week of development time. In the age of AI-assisted coding, that is not a moat. That is a weekend.
The companies that are not collapsing — Vercel, Replit, Notion — are not surviving because they trained better models. Vercel survives because it owns the deployment infrastructure that already runs production applications for OpenAI, Anthropic, Nike, and PayPal. Replit survives because it owns the runtime environment where code actually executes — something Claude cannot do on your behalf. Notion survives because it holds the accumulated knowledge graph of 100 million users’ organisational information. The model, for Notion, is interchangeable. GPT, Claude, Gemini — pick any. The context is not interchangeable. That lives in Notion, and it is yours. However if you leave, you cannot bring that context with you.
The pattern in every survival case is the same: they own something structural that infinite production cannot replicate.
What that structural something turns out to be is illuminating.
It clusters into five categories. Trust — the verified, accountable layer that allows transactions to happen when anyone can generate a convincing-looking interface in thirty seconds. Context — the specific, accumulated knowledge of your situation, your history, your relationships, which AI is general enough to lack and particular enough to need. Distribution — the audience, the platform, the curation layer that routes attention when supply is infinite and discovery is the only scarcity left. And Liability — the accountability that a system needs someone to hold when it acts in the world on your behalf.
The fifth category deserves particular attention: Taste.
Taste is the human judgment about what to build, for whom, and why. Not aesthetics — though that’s part of it. The deeper thing: the ability to look at a problem and know which part of it actually matters, to look at an audience and know what they need before they can articulate it, to look at a finished product and know why it isn’t finished yet. When production was expensive, taste was one input among many. When production is free, taste is the entire game. The vibe coder who ships an app in thirty minutes hasn’t done the hard part. They’ve cleared the ground for it. Every prompt, every workflow, every agent system still requires someone to decide what success looks like — and to recognise, honestly, when it hasn’t been reached. In an AI-assisted world, that judgment doesn’t diminish. It compounds. The tools for expressing taste change with every model release. The need for it doesn’t move.
These five things share a property. They all become more valuable as AI improves, not less. Every time a model gets better at producing, trust matters more because the cost of faking legitimacy falls. Every time a model gets better at reasoning, context matters more because better reasoning applied to the wrong situation produces better-reasoned errors. Every time more content is generated, distribution matters more because attention is finite and curation is the only way through.
The instinct — reasonable, understandable — is to worry that better AI makes human contributions less valuable. The five-layer structure suggests the opposite: what AI cannot replicate is not a residual category of tasks waiting to be automated. It is the structural core that makes everything else function.
There is a sharper version of this point for individuals, not just companies.
Building an application used to be a years-long specialization. Now it is a conversation. That is genuinely extraordinary, and it is not nothing. But as one analyst put it recently: building an MVP in a day means your real work starts the next morning. The oldest problem in business — how do you find the people who need what you’ve made? — has no model update. It requires exactly what it always required: judgment about who you’re building for, discipline about what matters, and the accumulated trust of an audience that believes you understand their problem.
When anyone can make anything, making things stops being the hard part.
The hard part becomes knowing what’s worth making. Knowing who it’s for. Having the judgment to recognize the difference between something that works and something that matters. That is taste — not as a soft, decorative quality, but as a cognitive discipline. The ability to hold a standard in your head and measure everything against it, at every step, regardless of how fluently the machine produces.
That is not a residual human task. It is the cognitive work that production was always in service of. AI has just made it visible by removing everything around it.
Text summarized and optimized using Anthropic’s models and reviewed by a human.