The lesson of the magenta line
In 1997 American Airlines captain Warren Vanderburgh gave his company’s pilots a lecture called “Children of the Magenta Line.” Young pilots had grown used to following the route line the flight computer draws and flying the automation instead of the aeroplane. Vanderburgh went through several years of incidents, and more than two thirds of them grew out of people failing to cope with the automation (chapter “Where Will the Seniors Come From”). A recording of the lecture is on the page “Captain Vanderburgh on automation dependency”.
On the night of 1 June 2009 the airspeed sensors of an Air France Airbus flying from Rio to Paris iced over, and the autopilot handed control to the people. The task was solvable, but the pilot at the controls had never practised flying such a machine by hand at such an altitude. Four and a half minutes later the aircraft hit the ocean, and two hundred and twenty-eight people died. In January 2013 the American aviation authority advised airlines to have pilots fly by hand more often. The same thing is happening to programming now, only without a cockpit and without black boxes.
What the data already show
In the summer of 2025 Stanford economists led by Erik Brynjolfsson looked at the payroll records of millions of Americans. In the occupations most exposed to AI, employment among twenty-two- to twenty-five-year-olds had fallen by thirteen percent since late 2022 relative to everyone else, and among young software developers by roughly a fifth. Experienced people were barely touched. The paper is called “Canaries in the Coal Mine” (same chapter).
In early 2026 two Anthropic researchers had fifty-two mostly beginner developers learn an unfamiliar Python library, half with an AI assistant, half by hand. On a test taken without the assistant, the first group averaged fifty percent and the second sixty-seven, and the assistant group did worst at debugging. Within that group, those who asked the model why things worked the way they did scored sixty-five percent or more. Those who asked it to write the code for them did not reach forty (same chapter).
Why the ladder broke
For centuries people became masters the same way. The apprentice swept the workshop, the journeyman did harder work under supervision, and after five to seven years you had someone who could be trusted near the load-bearing walls. In programming the sweeping was the small tasks: fix a form, finish a report, chase a dull bug. Those are exactly what agents now do better and faster than anyone (same chapter).
What a beginner does not know is easiest to show with an example. Two clients book the same manicurist for the same slot at the same moment, the program sees the slot as free for both and books them both. That is a race condition, and an administrator testing the site alone will never see it (chapter “Why a Beginner with Claude Did Not Become a Senior”). You can step through the race in the experiment “One chair and a double tap”. The model knows everything about race conditions; it just will not bring them up unless asked. That is why I keep repeating the formula: AI amplifies the one who understands and speeds up the disaster of the one who does not, and the difference between them comes down to the list of questions a person thinks to ask (same chapter).
What I suggest
Start where the pilots started, with flying by hand. In the early 2000s, when I worked at Nienschanz, Kostya and I spent our breaks building Linux from the book Linux From Scratch, everything from source. There was no practical point to it, but afterwards the operating system stopped being a black box for me for good. A beginner needs their own Linux From Scratch: a few things built entirely by hand (same chapter).
After that, the machine can be the best teacher beginners have ever had, as long as you ask it “why.” The school becomes reviewing agents’ code under someone who knows where to look, and the crash simulator is servers switched off on purpose, a restore from backup within an hour, a deliberately broken migration and other people’s post-mortems such as GitLab and Knight Capital (same chapter).
Money works against the twenty-two-year-olds here. Every company is better off waiting for someone else to raise the beginners, and if everyone reasons that way, in ten years there will be nobody ready to hire. Economists call this the tragedy of the commons. Someone will have to decide that growing their own people is cheaper and start before the competition (same chapter). The masterpiece of the new master, I think, will be a system they designed, fenced with prohibitions and checks, left to run unsupervised, and which broke nothing around it.
