Karpathy Said App Stores Are Dying. Here's What Has to Exist for That to Be True.

Published Feb 22, 2026 7 min read Nicholas Y., PhD

AI Infrastructure Future of Apps Community Data

Earlier this week, Andrej Karpathy — neural network pioneer, former Tesla AI director, one of the most credible voices in the field — published a post that sparked genuine debate across the AI world.

His argument: app stores may soon be obsolete. Instead of downloading a "Cardio Tracker" app, you describe your goal — "help me track my Zone 2 training for the next 8 weeks" — and an AI agent builds you a custom dashboard on the spot. In about an hour, using vibe coding with Claude, Karpathy did exactly that. He then asked the more interesting question: why did it take an hour? It should take one minute.

He's right about the direction. But buried in his post is the most important implication — one he names but doesn't fully unpack. The bottleneck isn't the AI model. It's the infrastructure underneath it.

What Karpathy Actually Said

The post is worth reading in full. Here is the core argument, distilled:

Software is shifting from "standardized products" to "on-demand, disposable tools" — generated fresh for each use case.
This requires hardware and services to provide AI-native APIs so agents can interact with them directly, not through UIs designed for humans.
The biggest obstacle right now is not the AI's capability. It is that "99% of products and services still have no AI-native CLI."
If devices offered agent-friendly APIs, if there were mature skill libraries, if AI held persistent personal context — generation time drops from 1 hour to 1 minute.

His frustration with the Woodway treadmill is telling. The machine is essentially a sensor — it converts physical data into digital information. But instead of offering an AI-readable API, it maintains a consumer web interface. Karpathy's agent had to reverse-engineer the treadmill's cloud API to extract raw data, debug unit conversion errors, and fix calendar logic — work that should not exist.

"99% of products and services still have no AI-native CLI. 99% of products and services still maintain .html/.css documentation, as if I won't immediately copy-paste the content to my agent to accomplish a task."
— Andrej Karpathy, February 2026

He's describing an infrastructure gap, not a model gap. And that gap is exactly what makes his prediction conditional: app stores will die when the infrastructure exists to replace them. We are not there yet.

The Missing Layer

Think about what has to be true for a parent to say: "Find something fun for my 4-year-old this Saturday morning in Cary" — and get a genuinely useful, accurate, real-time answer.

The AI model doesn't need to be smarter. GPT-4 can reason about toddler-appropriate activities. Claude can write the interface. The intelligence already exists.

What doesn't exist — what is almost entirely missing — is the structured, verified, real-time data layer that the AI needs to answer the question truthfully.

Which splash pads are open today vs. closed for maintenance?
Which library storytime is actually running this week (not cancelled due to a staff shortage)?
Which indoor playground has a toddler-only section — and is it too crowded on Saturday mornings?
Which swim school has an opening in the Tuesday 10am class?

None of this is on Google. Most of it was never on the public internet at all. It lives in parent Facebook groups, text threads, paper flyers pinned to bulletin boards at community centers, and the personal knowledge of the neighbor who's been here for ten years and knows everything.

Karpathy's treadmill problem and this problem are the same problem at different scales. The treadmill should have an API. The splash pad should have an API. The local swim school should have an AI-queryable interface. Almost none of them do. That is the missing layer.

Why This Doesn't Solve Itself

You might expect that as AI gets better and the ecosystem matures, this infrastructure gap will close automatically. It won't — for three structural reasons.

First, local knowledge is inherently community-maintained. Whether the splash pad is open today is not a fact that any technology company can scrape, infer, or model. It requires a human who was there, or who knows. That knowledge has to be captured, structured, and verified by real people in real communities. No amount of model improvement changes this.

Second, the web was built for human eyes, not agent queries. A splash pad's status might be posted on a parks department website — but buried behind three navigation clicks, formatted as unstructured prose, and updated on an irregular schedule. An AI agent scraping that site will fail unpredictably as layouts change. The only durable solution is structured data designed to be machine-readable from the start.

Third, the economic model for building this infrastructure doesn't yet exist at scale. Large tech platforms have no incentive to index hyper-local, rapidly-changing community knowledge — it doesn't serve their advertising models. Small community organizations don't have the resources to build APIs. The gap persists because no one has figured out how to make maintaining the data layer sustainable.

What "AI-Native Infrastructure" Actually Means for Local Knowledge

Karpathy talks about AI-native CLIs and agent-friendly APIs for devices and services. The same concept applies to community knowledge — it just looks different.

An AI-native layer for local family activities means:

Structured event data with real fields: date, time, location, age range, cost, open/closed status — not a paragraph of prose on a flyer.
Community-verified freshness — information that is actively maintained by people who use and attend these events, not scraped once and forgotten.
An API layer that AI agents can query with structured requests and receive structured, reliable responses.
A trust layer — provenance that lets an AI agent (and its user) know whether the information is recent, who submitted it, and how reliable it has been in the past.

This is what we are building with MyKidSpots. Not another app to download. A data layer — community-maintained, AI-queryable, structured from the start — that makes Karpathy's one-minute future possible for the questions parents actually ask.

The 1-Hour to 1-Minute Gap Is an Infrastructure Problem

Karpathy's post is framed as a prediction about software distribution. But read carefully, it is actually a diagnosis of an infrastructure deficit.

He spent most of his hour not on AI reasoning — not on describing the goal or structuring the logic — but on compensating for missing infrastructure. Debugging unit conversions because the treadmill didn't expose clean data. Fixing calendar logic because the API wasn't designed for agent consumption. Correcting errors that a well-designed data layer would have prevented entirely.

That friction is not unique to fitness tracking. It is the universal tax of building on infrastructure that wasn't designed for agents. And it will remain until someone builds the infrastructure that was.

The models are ready. The agent frameworks are ready. What is not ready — what the entire ecosystem is waiting for — is the data layer. Community knowledge, local expertise, real-time ground truth: structured, verified, and accessible to the AI that needs it.

That is the gap Yapplify exists to close. One community, one vertical at a time. Starting with the question every parent asks every weekend: what should we do tomorrow?