‘Learn to code’ is dead. ‘Talk to code’ is about to take over the world.

When a new technology is developed, we initially try to adapt it to existing patterns. The programmer writes code. The installer runs scripts. The researcher indexes documents in a database to later retrieve them. These habits feel natural; then something shifts, and the habits turn out to have been contingent and temporary. They were, like many arrangements we mistake as permanent, just what we happened to be doing at the time.

Coding agents with large language models have arrived, and they are replacing some patterns with something that looks, on first encounter, suspiciously simple. You describe what you want in English, and the thing gets built.

The history of technology is in part one of faith placed, too quickly, in systems that did not warrant it.

Andrej Karpathy, who has thought about neural nets longer than most, built an application called MenuGen: You photograph a restaurant menu, and the system generates images to illustrate every dish. The application is functional, culturally fluent, and genuinely useful. More notable is the process. Karpathy did not write the code. He described the application, and an LLM wrote both front-end and back-end. He has said, with the equanimity of someone who has made peace with a strange fact, that he does not really know how MenuGen works in the conventional sense.

The AI is his programmer now.

A cyborg language

There is a temptation to read this as novelty, as spectacle, as one more iteration in the long carnival of Silicon Valley announcements. The change here is structural: The barrier between having an idea and building a thing has collapsed to something quite small. Karpathy suggests the barrier may soon be low enough for anyone to publish an AI-driven application as easily as one can now post a video on a social media site.

The installation script is another interesting case because it is modest enough to be revealing. Mintlify, the documentation company, proposed in early 2026 that software should ship with English files instead of shell scripts: install.md, a human-readable checklist that a coding agent can read and execute. The agent detects your operating system, detects your environment, proceeds through each step, and pauses for your approval before running commands. The process is more auditable than a Bash script because anyone can read it. More generally, software can support skill.md files, describing both installation and usage, served at well-known URLs so that any coding agent can load them. Developers are now writing documentation for machines, rather than humans, although humans can easily read and edit it.

Karpathy has noted that digital services could become LLM-friendly: documentation in Markdown, command-line interfaces exposed for commands, APIs that take English. The user and the AI have effectively merged into a single audience. Design for one, and you design for both.

LLMs address the knowledge problem in ways that Vannevar Bush, in 1945, could not imagine. Bush dreamed of the Memex, a personal filing system for all one’s books and communications, a mechanized memory that would allow fast and flexible recall. He was describing, in the vocabulary available to him, what we would now call a knowledge base. The LLM makes this capacity available to everyone without clever indexing.

RELATED: Social media scams are up 700%. Here’s how to stay safe.

Media Trading Ltd/Getty Images

Karpathy’s approach to personal knowledge management involves feeding raw sources to an LLM, which then writes a linked wiki of Markdown documents using summary pages, encyclopedia-style articles, connections between ideas. The AI performs a health check on its own wiki, looking for contradictions and gaps, refining the content. It is, as Karpathy has described it, a full-time research librarian. Any claim can be verified by a human who reads the file. The complexity of the database is bypassed through the simpler expedient of using words.

With great speed comes great care

This design is not infallible. An LLM can hallucinate, conflate, or extrapolate with confidence from insufficient evidence. The wiki it writes is plausible before it is accurate. Error-checking in these systems is not a solved problem. These are not small concerns. The history of technology is in part one of faith placed, too quickly, in systems that did not warrant it.

Yet the direction is clear. We are moving from an era in which computers required exact instruction to one in which they accept intention. Assembly language gave way to high-level languages; high-level languages gave way to graphical user interfaces; graphical user interfaces are giving way to conversation. Each transition lowered the barrier to software development. Each distributed power in its own way, with costs and distortions that took years to fully account for. This transition will be no different in that respect.

The medium, however, is different. Previous transitions expanded access to computation while keeping language in its ordinary role: a wrapper for technical instruction, a label on a button, a comment in a codebase. This transition makes language the primary interface. The program is the sentence. The installation script is the paragraph. The application is its description rendered executable by a model that has read more human text than any human could read in a thousand lifetimes. We have always known that words could build worlds. Such building used to route through human action in a way that is no longer true.