Nobody knows how to train juniors anymore.

Under letter #100, a tech leader left a comment that I haven't stopped thinking about.

They wrote that most of the conversation about AI is about what it saves, and that the cost shows up later. If all the real knowledge stays with a handful of senior people, we are creating a gap we'll pay for down the line.

Then she asked: "Have you seen any teams getting this right?"

I answered: "Not yet, including me."

That's the honest answer. And I think it deserves a full letter, because "not yet" is not a very satisfying thing to say to someone who is building teams right now.

In January I wrote "Why I banned AI for my juniors (for 6 months)." The protocol was simple. Manual first: juniors, or anyone learning a new domain, do the core work without AI for three to six months. Then the Reverse Turing Test: I point at a random line of AI-assisted work and ask "explain exactly how this works and why the alternative was rejected." If you can't, you lose your AI license for the week.

I still believe the rule behind it: you are not allowed to automate what you do not understand.

But a rule protects one junior, on one team, with one senior who cares. It doesn't explain why almost every company is doing the opposite, even when their own CEOs say it's a terrible idea.

That's what I want to understand this week.

Everybody agrees, and nobody does it

Here's the strange part.

The AWS CEO, Matt Garman, was asked about replacing junior employees with AI. He called it "one of the dumbest things I've ever heard." His reasoning: they're your cheapest people, they're the most comfortable with AI tools, and in ten years you'll have nobody who has learned anything.

I agree with every word.

And yet.

A Harvard team looked at data covering about 285,000 US companies and 62 million workers. After early 2023, in the firms that adopted generative AI, junior employment fell by roughly 7 to 12 percent compared with firms that didn't, while senior employment kept growing. Almost none of it came from layoffs. Those companies simply hired fewer juniors.

(Same pattern I described in Letter 85. Nobody gets fired. The next batch just never gets hired.)

A smaller study this summer, from researchers who interviewed junior and senior engineers in South Korea, gave it a name: absorption. AI expanded what seniors can do alone, so the entry-level work got absorbed into the senior's workflow. One startup founder told them: "We let a lot of junior engineers go... there was no impact whatsoever."

And here's the detail that stuck with me. The seniors in that study assumed juniors would get the same experience they got. The juniors were describing a path that no longer exists.

So we have a market where everyone says the right thing and does the other thing. When that happens, look for the incentives.

Juniors used to pay with their work

This is the part I had never fully understood, and an economist at IESE, Enrique Ide, explains it beautifully in a paper on how AI affects the transmission of knowledge between generations.

The knowledge that makes a senior valuable is mostly tacit. You can't write it all down. You can't verify it on paper. And you can't sell it in advance.

So think about the junior's side of the deal. The best way to learn is to sit next to someone excellent and watch how they think. That access is worth a fortune over a career. But a 23-year-old can't take out a loan to pay a great senior for it, because there's nothing to borrow against. The skill doesn't exist yet, and once it does, it lives in their head.

So how did juniors pay for it?

With their work.

The boring tasks. The first drafts. The four hundred contracts. The fifteen versions of the same model. The long hours at the top law firms and in medical residency, at pay that makes no sense for the hours. That was the deal: you do the work I don't want to do, and in exchange you get to watch how I think.

Ide puts it in one line that I find brutal: juniors are forced to finance their mentorship solely through their labor, "the very currency that an improvement in entry-level automation devalues."

That's it. That's the whole mechanism.

In Letter 85 I wrote that the grunt work was the gym. There was a second half I didn't see. The grunt work was also the payment.

AI removed the reps, and it also made the junior's only currency nearly worthless. The senior doesn't need their first draft anymore. So the deal quietly breaks, and nobody has to decide to break it.

And it gets worse. The value of training a junior shows up years later, in the senior that junior becomes. Nobody on this quarter's budget gets paid for that. So, as Ide shows, a market left alone automates early-career work more than is good for everyone, even when every single company is acting rationally.

That is why "everybody agrees and nobody does it." Each company is making the right decision for itself, and the wrong decision for the next ten years.

(One more subtle point from the paper that I find very practical: you don't even need to cut junior jobs to break the pipeline. If your juniors end up working next to your least skilled people, because your best people are too busy, the best practices spread more slowly, and the damage compounds over generations.)

The good news is in the same data

Now here's the part that gives me hope.

In January, Anthropic ran a small experiment with 52 mostly junior engineers. They had to learn a Python library they had never used. Half had an AI assistant, half coded by hand. Then everyone took a quiz on what they had just used.

The AI group scored 50 percent. The hand-coding group scored 67 percent. The biggest gap was on debugging, which is exactly the skill you need to catch AI's mistakes.

But look inside the AI group. The people who handed the work to the AI scored below 40 percent. The people who used it to ask conceptual questions, or generated code and then asked for explanations, scored 65 percent or more. Almost as well as the people with no AI at all.

Same tool, same task, and the result depends entirely on how the junior used it. (AI is an amplifier. It amplifies learning or it amplifies avoidance.)

Ide's model has the same twist. AI can also let your best experts run bigger operations. If that means each of your best people works with MORE juniors, knowledge spreads faster, and growth goes up.

So the technology can go both ways. It can be the thing that closes the apprenticeship, or the best apprenticeship tool we ever had. The difference is entirely in how we organize the work around it.

That's why nobody has got it right yet. Including me. The default only goes one way, and the right way has to be designed on purpose.

What I would do on Monday

I'll keep the January rules. You are not allowed to automate what you do not understand, and the Reverse Turing Test is still the cheapest diagnostic I know.

But the economics changed how I think about the rest.

Pay juniors in a new currency. If their work is no longer worth much, their understanding has to be. Make learning the explicit job, and evaluate it. Ask what they understood this week, where they got stuck, which gap they found in their own knowledge. Researchers in the Korean study recommend exactly this: judge juniors on how well they recognize what they don't know, not only on what they ship.

Put juniors next to your best people. Your least busy senior is often the wrong mentor. And your best people are now proofreaders (Letter 94), drowning in AI output to review. So turn the review into the lesson. When a senior reviews AI-generated work, the junior sits in and the senior thinks out loud: what looks wrong, what they check first, why they'd reject this. That's the tacit knowledge, finally visible. You are paying for that review anyway.

Teach the questions, ban the delegation. The Anthropic data is clear: "explain this," "why this and not that," "what would break" build skill. "Just do it" destroys it. Make that the house style for anyone in their first years.

Protect a few real tasks on purpose. Aviation does it: pilots still fly by hand to keep the skill, even when autopilot could do it better. Keep some small, real, low-risk tasks for juniors to do end to end, even when AI would be faster. It will look inefficient on a dashboard. It's the cheapest insurance you'll ever buy.

None of this happens by default. Someone in the company has to decide that it matters, because the market has shown it won't decide for you.

Back to the question

"Have you seen any teams getting this right?"

Still not yet. But I can tell you now what "right" looks like. A team where the junior's understanding is worth something again, where the best people get more juniors around them instead of fewer, and where AI is used to make thinking visible instead of making it unnecessary.

If your team is trying something like this, hit reply. I'd genuinely like to know what's working. That might be the best letter I write this year, and you'd be writing it.

The grunt work was the tuition.

Somebody has to pay it now.

Have a great weekend.

Stay sharp.

— Charafeddine (CM)