Karpathy: Apps Should Never Exist, Human Understanding Is Our Last Moat, CPUs as Sidekicks

LLMs as a New Computing Paradigm

Andrej Karpathy frames large language models (LLMs) as a new kind of computer: the context window functions as memory, prompts act as a programming language, and the model itself serves as an interpreter. This view goes beyond treating LLMs as speed‑up plugins for existing software.

Software Evolution: 1.0 → 2.0 → 3.0

He revisits the familiar progression:

Software 1.0 : developers write explicit rules and shell scripts that encode paths, dependencies, and environment variables.

Software 2.0 : developers train models and let them generate code.

Software 3.0 : developers craft prompts that program the model directly.

In this framing, the prompt becomes the executable interface, and the model’s interpreter replaces traditional command‑line tools.

Concrete Example – Menu Gen

Karpathy’s “Menu Gen” project originally required a pipeline of OCR, database mapping, image generation, UI rendering, and multiple service calls. By uploading a menu photo and prompting the model with “overlay dish images onto the original menu”, the model directly outputs the final image. The input and output are both images, and the entire software stack that previously handled data flow disappears. He argues that the full Menu Gen application is redundant under the new paradigm because the task can be expressed as a single prompt.

Replacing Installation Scripts with Natural‑Language Instructions

Traditional installation scripts encode every path and dependency (Software 1.0). Karpathy demonstrates that a plain English description—e.g., “install the package according to these steps”—can be fed to an LLM, which then interprets the environment, performs the installation, and conducts real‑time debugging. The language itself becomes the execution layer, suggesting that future software interfaces may be legible to LLMs rather than human‑focused GUIs or APIs.

Jagged Intelligence

Karpathy introduces the term Jagged Intelligence to describe the uneven growth of model capabilities. In domains with clear verification signals—code, mathematics, games—models achieve rapid, near‑escape‑velocity progress because labs can apply reinforcement learning with measurable rewards. In contrast, commonsense tasks lacking easy verification (e.g., “walk to a car‑wash 50 m away”) lag behind due to limited data coverage and lower laboratory priority.

Economic Perspective: Laboratory Focus

Laboratories have finite compute and data‑engineering resources. They allocate these resources to high‑value, quantifiable problems, creating “data bias” that accelerates performance in verified domains while leaving other areas under‑developed. This economic behavior explains the jagged performance curve.

Entrepreneurial Implications

When asked whether newcomers can compete in verified domains, Karpathy answers affirmatively, emphasizing vertical markets. The key steps for a startup are:

Identify a task that can be framed as a verifiable problem.

Build a feedback loop to collect domain‑specific data.

Fine‑tune an LLM on that data using reinforcement learning.

Even if large labs have not yet tackled the niche, this approach can yield measurable gains. Ownership of real data, task workflows, and evaluation standards becomes the competitive advantage for vertical AI systems.

Agent‑Native Economy

Karpathy predicts that future products will be composed of sensors, actuators, and logic that are directly understandable by LLMs. Current software documentation still instructs humans (“click this”, “run this command”), which he sees as a mismatch for agent‑centric systems.

Extending the Menu Gen example, the bottleneck shifts from coding to deployment. An agent receiving the prompt “build Menu Gen” could automatically provision DNS, configure services, and launch the application without human intervention.

He coins the emerging profession Agentic Engineering , which focuses on orchestrating multiple agents to accomplish complex tasks while ensuring quality, safety, and reliability. Evaluation of talent may move from algorithmic puzzles to the ability to leverage agents for real‑world projects.

Future Competitive Edge

The most important future advantage, according to Karpathy, is making data structures, process descriptions, and system interfaces as legible as possible to LLMs. When models can directly read and execute these specifications, they become the primary computation layer, relegating traditional CPUs to auxiliary deterministic tasks.

Human Role in an Agent‑Dominated World

Even as agents automate generation, comparison, and execution, they cannot replace human “understanding”—the ability to decide what matters, what is true, and what risks are acceptable. Karpathy stresses that personal knowledge bases, wikis, and tools that reorganize information help preserve and enhance this understanding, providing a strategic edge as AI takes over routine execution.

Code example

.md