The AI Programming Revolution

A large language model like GLM can replace a skilled programmer for most coding tasks. That is not a prediction — it is the current state of the technology, and we are running it in production every day.

The days of a programmer writing code are over

For most of software's history, the scarce resource was typing skill: someone fluent in the language, the frameworks, and the endless ceremony around them. That era has ended. Given the right inputs, an LLM can produce in a few minutes what a programmer would have taken days or weeks to accomplish — the application code, the deployment configuration, the infrastructure wiring, even the DNS records to point a domain at it.

We do not argue this in the abstract. On this very site, we have written about the Flask Factory: an LLM takes a described requirement and returns a deployed, working web application — and the site you are reading this on was produced that way, in a single working session.

What is still helpful: the senior review

None of this means experience is obsolete. It means experience moved up a level.

It is still helpful — we would say essential — to have an experienced senior programmer review the code and the strategic direction that an LLM generates. Not to write it: to judge it. To catch the confident mistake, the design that works until it doesn't, the shortcut that will cost you in a year. The machine produces at superhuman speed; the human ensures what was produced deserves to exist.

The economics follow the division of labor: one senior review of machine output replaces weeks of junior typing. The programmer's job stops being production and becomes quality, direction, and taste.

The uncomfortable summary

If your business still budgets for code-writing, you are overpaying. The typing is a solved problem. What you should budget for is the judgment to direct it and the experience to audit it — and increasingly, the skill of asking the right inputs in the first place.

That is the revolution: not that machines learned to code, but that knowing what to build, and checking what gets built, became the entire job.