Who will train our kids as AI takes over junior work

What I tell my kids about finding work that will teach them something.

TL;DR: AI is taking over some of the work that helped beginners build skills and show they could be trusted. I’d put a manager who teaches ahead of a famous employer. Young people need chances to make decisions, with someone close enough to help them learn from the result.

My oldest graduated from the University of Oregon in May. My middle graduates this coming May. My youngest starts college next fall. My wife asked me this question at dinner last night: what does a career even look like now?

I’ve been giving this question a lot of thought and I’ve spoken about it at keynotes and client meetings. This week I decided to give the question to the most advanced models in ChatGPT, Gemini, and Claude and had them challenge each other’s answers. Three rounds, each one reading the others’ arguments and pushing back. Where they disagreed turned out to be the useful part.

The research group METR tracks how long a task AI can complete. In October 2024, Claude 3.5 Sonnet could handle software and research tasks that take a skilled person about twenty minutes, succeeding about half the time. By this spring, an early version of Claude Mythos Preview had reached roughly seventeen hours. Those are contained tasks with clear grading, and the hours describe how long the work would take a person, not how long the model ran. Models also hit gold-medal scores at the International Math Olympiad in the summer of 2025.

Last winter I ran a demo for a law firm of more than 100 people. Claude produced research and a memo draft in about eight minutes, work I’d have expected to take a first-year associate a day. It still needed review. The partner asked, “So what do we hire the new people to do?”

I’ve been thinking about what that assignment used to accomplish. The firm got a memo. The associate got practice, including corrections from someone who knew more. Over time, the partner learned which decisions that associate could handle.

The models fought about that last part. Gemini called the old grunt work “an institutional trust and hazing ritual,” a way for partners to test a kid’s reliability before trusting them with a client. ChatGPT pushed back that tedious work “could reveal reliability without having been designed as a character test.” I’m closer to ChatGPT. Some grunt work taught people, and some was just grunt work. I have no interest in defending two years of avoidable drudgery. But when a firm removes a task, it should ask whether the task was teaching anything, or showing anyone who could be trusted.

If the firm gives the assignment to AI, it still needs a way to do both. That part takes effort, and the savings from automating the work arrive first.

The training has to come from somewhere

ChatGPT wrote one sentence I keep coming back to: “The scarce educational resource will be permission to learn at someone else’s risk.”

That’s an uncomfortable thing to buy. AI can explain a lot to a beginner, and cheaply. What it can’t do is hand a beginner an assignment that matters and leave room for some mistakes. A manager has to do that, then stay involved long enough to explain what happened.

Hiring someone who already knows the work costs less than teaching a beginner, and a firm that trains a beginner creates someone its competitors can hire. If enough firms make the cheaper choice, they leave fewer places where beginners can become those people.

Training could become a competitive advantage for the firms that see it. I wouldn’t assume enough of them will in time for my kids.

So here’s my bet. If firms keep skipping the training, then by the early 2030s a thirty-year-old who has made hard decisions and lived with the results will be hard to find. I notice I want this to be true, which is a reason to be careful with it. Senior people may use agents to stay in their seats longer, and firms may need fewer experienced people too. Still, I’d want my kids getting that experience wherever they work.

Hiring can change before the whole firm changes

On October 16, Anthropic published a report on its own workforce. By August, AI agents directed by individual employees were logging about six hours for every human hour. The junior share of its technical hires fell from 27 percent to 9 percent since early 2025, even though total technical hiring kept growing. It’s one company, but it’s the one building the tools.

Stanford researchers reported a broader pattern in August. Using payroll data through June, they found employment for 22-to-25-year-olds in highly AI-exposed occupations about 19 percent below where it would be if it had kept pace with less-exposed peers. The adjustment came mainly through reduced hiring. The authors describe the evidence as descriptive, not causal. They also found no widespread, economy-wide displacement.

Now extrapolate. Over the long run METR has found task length doubling about every seven months, and lately it has moved faster. Even at the slower pace, agents would handle bounded tasks that take a skilled person a full work week by 2027, and a full month by 2029. That’s still at the “works half the time” level, on tasks with a clear definition of done, and the trend could slow. My working assumption is that AI will handle substantially more junior work by the time my youngest graduates, and check more of it too. I wouldn’t build a career plan around humans always being better at review.

Organizations will move unevenly. A general counsel and a VP of Sales can want opposite outcomes, and neither may put the real objection in writing. But a firm can stop replacing departing juniors well before it has worked out how to run the whole business differently.

I wrote last year that the career pyramid was becoming a diamond, with a thick middle of experienced people running the tools. I no longer trust that picture. I expect agents to take over more of the coordinating, so the middle may thin too. I wouldn’t treat prompt engineering or coordinating agents as a lasting career advantage either.

Producing more gives a young person no automatic claim on the savings. If one analyst with agents does the work of four, the firm may simply hire one. So ask who keeps the savings when the work gets cheaper. I’d look for work that puts you close to decisions and the people affected by them. That can be customer work or an internal project, and a thirty-person company may hand you real decisions sooner than a famous firm. The job title won’t tell you how much you’ll learn.

And I wouldn’t choose a profession on the promise that someone will always have to take responsibility. One person with agents may sign for work that used to take a team, and cheaper drafts could put pressure on fees. None of that tells a twenty-two-year-old which job is safe.

Learn enough to explain your own work

For a skill you’re learning, start with an attempt of your own. Then compare it with the AI’s answer, and work through the differences with a teacher or colleague who can tell you where your reasoning failed. Pilots still hand-fly approaches with the autopilot available, because the day it fails is the wrong day to learn. You need to know what you understand before you start delegating it.

An Anthropic study published in January tested 52 engineers learning an unfamiliar software library. The group with AI assistance scored 50 percent on the mastery quiz. The group that coded by hand scored 67 percent. The largest gap was in debugging. Some AI users who asked for explanations and checked their understanding scored much better than participants who handed the work over. That part of the study was observational, so I hold it loosely.

Gemini wanted me to tell my youngest to struggle with a problem “until your head hurts.” ChatGPT objected that “pain is a poor measure of learning.” The study pushed me toward ChatGPT. What matters is what you do while the tool is open. Time with it switched off can be useful, but hours alone won’t tell you whether you learned anything.

So explain why your answer works, then try a changed version of the problem without help. Ask the AI for a hint before you ask it for an answer, and have it challenge your explanation. You still need feedback you can trust, especially when you’re too new to recognize a confident mistake.

Practice alone won’t get you promoted, though. As Gemini put it, “nobody promotes you for practicing alone in your apartment.” Practicing on your own builds the skill. Keeping a promise where people can watch is what gets you trusted with bigger work.

What I tell each of them

To my oldest, five months in: look at what your job is teaching you. Are you getting harder decisions? Does anyone explain why they changed your work? Internal tickets can teach a lot, and a customer-facing job can still give you no authority. Look at the assignments, and volunteer for the messy problem nobody owns. If a year passes without more responsibility or useful feedback, start looking for a better place to learn. Keep paying the bills while you look.

To my middle, graduating in May: ask every interviewer one question. “What did your last junior hire learn to own in six months, and who taught them?” Then find that hire and check. Ask what happens when a beginner makes a mistake. I’d take a boss who teaches over a logo on the resume, assuming the pay and circumstances work. Bring something you built and be ready to explain where the AI broke.

To my youngest, choosing a school: pick subjects that interest you enough to work through the difficult parts. Compare the cost of the actual programs, and ask how students get paid experience in their field. Do an honest attempt at the problem set before asking the model to help. Read whole books. Spend time with people who know more than you and are willing to explain their thinking. I want you able to work without the model, and able to use it well when it’s available.

For all three: as more resumes and writing samples start to read like the same model, keep a few examples of work you can explain in your own words. What decision did you make? What changed after you made it? Hearing you think out loud through those questions tells me far more than a polished output.

Start where you have access

Supervised responsibility often depends on who will give you a chance. My kids have me. A lot of kids don’t have a parent who can make the call, and I’d rather say that than pretend the transcript decides everything. If polished applications start to look alike, I’d expect personal referrals to matter even more.

If you need income, the job you have is a reasonable place to start. Ask to handle one recurring problem during paid hours. It could be a customer complaint that keeps coming back, or a handoff between colleagues that keeps going wrong. Agree on what you can change and ask your manager to review the result after two weeks. Keep a short record of what you did. That gives you something specific to point to when you ask for the next assignment or apply elsewhere.

If you’re still in school, ask a department adviser about paid placements and who supervises them. Use the formal application process, and ask a professor or a former boss to make a call for you too. A small assignment won’t fix a weak job market, but it gives you evidence of what you can do without months of unpaid work.

Where I’d put a parent’s effort

When one of my kids compares offers, we’ll ask how often someone will review their work, and we’ll look at the pay and the cost of living. A job has to be workable before it can teach anything.

If I make an introduction, I’ll ask about a specific opening or a chance to learn. An introduction to a job where nobody has time for a beginner isn’t worth much. Then my kid handles the conversation, because asking for responsibility takes practice too. And if the first search is slow, I won’t treat it as a verdict on them. Right now it may say more about how firms are hiring.

If you run a firm, ask how much of the time AI saves goes back into teaching beginners. Fund one paid apprenticeship with a named supervisor. Publish the criteria for who gets it and what the person should learn, and give the supervisor time to teach. Open that door to someone whose parent doesn’t know you.

I expect employers to keep finding ways to get more work out of fewer people, and I don’t expect the training to come back on its own. Before my kids accept a job, I want us to know which decisions they’ll learn to make, and who will be there when they get one wrong.

Thank you for reading! If you enjoyed this, please share it with others. I’m in NY for a couple of days so sorry - no new Magnus pictures.