Key takeaways
- Andrej Karpathy believes app stores are becoming outdated in favor of AI-generated applications.
- He successfully created a custom fitness dashboard in under an hour using an AI coding agent.
- Karpathy argues that specialized apps shouldn't exist when LLMs can generate code on demand.
- The industry should shift towards services designed for AI agents, not human users.
- Karpathy emphasizes the need for better APIs for devices to simplify data access and interaction.
Andrej Karpathy, the former OpenAI researcher, says the idea of browsing and installing apps from an app store is increasingly becoming outdated.
In a recent post on X, Karpathy recalled building a custom dashboard to track an eight-week cardio experiment aimed at lowering his resting heart rate from 50 to 45 beats per minute. Instead of downloading a fitness app, he used an AI coding agent to generate a small web application that pulled data from his Woodway treadmill’s cloud service, processed it, and displayed progress against his goals.
The project took about an hour and required manual fixes, including correcting unit conversions and calendar alignment. Karpathy noted that two years ago, the same task would have taken roughly ten hours. But he argues it should ultimately take closer to one minute.
“There will never be (and shouldn’t be) a specific app on the app store for this kind of thing,” Karpathy wrote. “I shouldn't have to look for, download and use some kind of a ‘Cardio experiment tracker’, when this thing is ~300 lines of code that an LLM agent will give you in seconds.”
He added that the idea of an app store as a catalogue of a discrete set of apps feels somehow wrong and outdated when LLM agents can improvise the app on the spot with personalisation.
Karpathy also argued that the industry must reconfigure around services built for agents, not humans. In his view, products should function as sensors and actuators with agent-native ergonomics.
He uses his Woodway treadmill as an example. It is a sensor that converts physical activity into digital data. It should not depend on a human-readable frontend, nor should an LLM have to reverse engineer its web interface to access that data. Instead, it should expose a clean API or command-line interface that an agent can call directly.
He said the progress toward this model is slow. Most products and services still lack AI-native CLIs (command line interfaces). Instead, they maintain HTML and CSS documentation designed for people to read and follow manually. Users are expected to open URLs, click through dashboards, and execute step-by-step instructions themselves.
“Today I am impressed that this random thing took 1 hour (it would have been ~10 hours 2 years ago). But what excites me more is thinking through how this really should have been 1 minute tops,” he added.
However, his post attracted numerous counterarguments. Sebastian Raschka, author of Build a Large Language Model From Scratch, stated that while the model is shifting, app stores can still serve a role as an added security layer, especially for complex apps that communicate with the web.
He suggested a possible “App Store 2.0” model in which users download a base version of an app, such as a fitness tracker, and then customise it through prompts at the operating system level.
