Generated by Claude. See Beating Pangram and AI detectors.

AI is cutting the entry-level jobs that used to train tomorrow’s architects. Aviation, chess and bookkeeping have run this experiment before, and solved it three different ways.

Zoho’s Sridhar Vembu posted something that’s been bugging me: AI makes senior architects more productive and cuts the need for junior engineers. Then the harder line: if nobody starts junior, how does anyone become an architect?

The data backs his worry. Stanford found that since late 2022, employment for 22-to-25-year-olds in AI-exposed jobs like software fell 16 percent relative to older workers in the same roles. Matt Beane at UCSB watched this happen in robotic surgery before AI arrived: robotic consoles let surgeons do what residents used to do, so surgeons stopped bothering to train residents.

I don’t have a clean answer. But history has three, and which one applies depends on what happens when the entry-level worker gets it wrong.

Switch, when being wrong is cheap. Spreadsheets made manual ledgers redundant. But manual bookkeeping used to teach accountants how errors happen. So the job changed: fewer people checking arithmetic, more people designing the checks that catch it. There are more accountants today than before spreadsheets existed.

Enforce, when being wrong is expensive. Autopilot cut junior pilots’ hand-flying hours until Air France 447 went down. Regulators responded with the FAA’s 2014 rule mandating manual-flight training. Surgical simulation centers do the same job now: they force the rare-case practice that robotic consoles took away.

Level-up, when the work is the point. A free phone app now beats any grandmaster. Chess should be dead. Instead, engines became coaches, and today’s young grandmasters are stronger than any generation before them.

[Suggested visual: three-panel strip - cockpit dash, chess board, ledger/spreadsheet - captioned “Switch, Enforce, Level-up”]

Most commercial software sits in the switch bucket. A broken deployment, a weak first draft: you catch it and move on. So I’ve stopped hiring junior developers to write code. I hire them to catch what AI gets wrong. One of my interns records client calls, feeds the transcript to a coding agent, and ships. I told him deliberately not to try to understand the domain. He’s three times faster than someone with five years of experience, because experience was the bottleneck.

For mission-critical software, this doesn’t hold. Wrong code in a trading system or a medical device sits in the enforce bucket: build the skill on purpose, using AI as the simulator, because Bainbridge’s old irony of automation still applies.

“The better the machine gets, the rarer the human intervention, and the harder it is to stay sharp for the one time it matters.”

Which bucket your junior engineers are actually in isn’t obvious. Get it wrong either way and it costs you: over-train for switch work and you waste everyone’s time; under-train for enforce work and you’re one Air France 447 away from finding out.

Vembu is still thinking through how this resolves. So am I.