Who Teaches the Juniors Now?

Every senior engineer I know built their judgment the same way. Years of writing the boring code, getting it torn apart in review, and slowly learning why. Nobody taught it. There was no course. It accumulated as a byproduct of grunt work: the CRUD endpoints, the off-by-one bugs, the review comment that stung for a week. That was the tuition.

AI just automated the tuition away.

It does the grunt work better than a junior does, which makes not hiring the junior the economically rational move, and companies are making that move right now. Look at any team’s hiring this year: senior head-count holding, entry-level quietly gone. Each individual decision is defensible. Why pay a first-year to write the endpoint the model writes faster and cleaner? But stack the decisions up and the industry has done something strange: we kept the seniors and deleted the process that makes them.

Nobody decided this. That’s what makes it interesting. It’s what happens when every company optimizes for this quarter’s output, because every one of them is defecting against the industry’s supply of next decade’s judgment. A classic commons problem, and there is no one whose job it is to fix it. The senior shortage of 2035 is being manufactured right now, one rational hiring freeze at a time.

The counterargument deserves a full hearing, because it might be right. Maybe the apprenticeship was always an inefficient hazing ritual, and we’re nostalgic for it the way people get nostalgic for hard winters. A junior with an infinitely patient tutor that explains every mistake the moment it happens might learn faster than one waiting a week for a grumpy senior to get to their branch. Under that reading, AI didn’t delete the tuition. It cut the price and improved the instruction, and the juniors who do get hired will grow faster than we did.

The counter to the counter, and the crux of the whole question: explanation is not experience. You don’t learn why the abstraction was wrong by being told. You learn it by living inside the consequences of the wrong abstraction for six months, by being the person paged when it breaks. Whether AI-compressed learning transfers judgment, or only transfers knowledge, is genuinely open. I’d write both sides at full strength because I don’t know which one wins, and neither does anyone selling you certainty about it.

There’s a personal version of this question I can’t dodge: would I hire a junior today, for my own products? The honest answer is: not for the grunt work that used to be the job description. Maybe for judgment under supervision, if I can define what that looks like when the model already writes the endpoint. I don’t have a clean hiring plan that solves the commons problem. I have a worry that the path I took is closing behind me.

And there’s an uncomfortable connection to my own situation that I’d rather name than have pointed out. The solo-with-AI model I run works because I already have twenty years of judgment to spend. The tools multiply what the years built. That path isn’t available to someone starting now: the work that would build the judgment is exactly the work the tools absorbed. I argued in Taste Is the Last Moat that judgment is the durable asset. This is the other half: we may have dismantled the factory. I could be wrong that nothing replaces it. I hope I am.

The generation that benefits most from AI may be the last one trained without it.