It is no longer unusual to walk through a software shop where almost no one writes code by hand. Wes McKinney, the developer behind the data-analysis library pandas, recently wrote that he no longer types code at all. Instead he pushes more than ten billion “tokens” a month through artificial intelligence (AI) models. A token is the smallest unit an AI uses to process text or code. McKinney is now shipping work written in a programming language he has never touched directly. A task once treated as a craftsman’s manual skill is migrating to machines at speed.
Developers are divided over what this means. One camp favors a setup in which humans set the direction and AI does nearly all of the building. They call it the “dark factory,” after the unlit robot plants that need no lights because no humans work the floor. A planning agent breaks the job into small pieces, separate agents each build and test their piece, and failures simply retry on their own. The human’s job is not to produce the work but to inspect it. The other camp insists that human oversight is still the heart of the work: quality turns on how well you coordinate a swarm of agents and how precisely you tell them what to build.
But inspection is harder than it sounds. Whether the code an AI hands back actually works is usually decided by automated tests. If the tests themselves are shaky, plausible-looking bugs get mass-produced at machine speed. One tool even has the agent record a video of itself using the feature it just built and attaches the clip for a person to review. At a recent developer event, someone framed the danger plainly: is your evaluation lying to you?
There is a telling observation buried here. McKinney has noted that once a codebase reaches roughly a hundred thousand lines, the AI begins to choke on the very pile of code it generated. Producing more has become easy; managing the complexity has not. The moment recalls an insight from fifty years ago. In 1975 the software engineer Fred Brooks argued that adding people to a late project only makes it later. Swap people for AI and the question is unchanged: throwing unlimited hands at a problem does not produce good design.
The enterprise view is different again. Aaron Levie, chief executive of the cloud file-storage company Box, argues that AI does more than automate existing work. It unlocks work a company could never afford to attempt: analyzing stacks of contracts no one ever read, refining processes no one ever had time to touch.
So the conclusion many have reached is that the real bottleneck was never typing speed. The judgment to decide what to build and what to leave out, call it taste, remains the scarce resource. And a question follows close behind. Taste was built over years of touching the code directly. If newcomers no longer write code by hand, where will that taste come from? The craft of writing code is not so much disappearing as shifting its center of gravity, from writing code to designing systems.