What to build in four years
Notes from a talk to first-year IT students: what a degree still buys, what AI actually changed about the job market, and how to learn when a machine will answer for you.
TL;DR
A degree still moves your employment odds. AI has changed the shape of most IT jobs without deleting them. The people who do well over the next four years will be the ones who can still think when the assistant is switched off.
This entry is evergreen. It is maintained and revised, not finished. Where a claim rests on evidence, the evidence is linked into the Brain so you can check it rather than take my word for it.
Here is the shape of the argument, so you can see where each part is going before you read it:
A degree moves your employment odds, and AI reshapes the work rather than deleting it. Both push value toward judgement. Judgement is built by learning that survives the assistant being switched off, which is practised by building real projects, and held together by acting before you are certain.
01 / University is not the same thing as work
The honest version: university teaches you to produce correct answers to questions someone else has already decided are worth asking. Work gives you an unclear problem, insufficient information, a deadline, and other people with different incentives — and asks what you are going to do about it.
That is a genuine discontinuity, and nobody smooths it over for you.
What a degree does still do is measurable. Across OECD countries in 2024, employment rates for 25–64-year-olds rose steadily with attainment.
Adults aged 25–64. The gap between tertiary and below-upper-secondary attainment is 27 percentage points.
This is an association across populations, not a promise to any individual. Attainment travels with prior advantage, family resources, and country labour market conditions. It tells you the odds moved. It does not tell you the degree moved them on its own.
So the credential is worth something. It is also not the thing that makes you good. Both are true at once, which is why the useful question is not “is university worth it” but what education is actually for — and why credentials can still be useful without being sufficient.
02 / What AI did to IT work, and what it did not
The loudest version of this story is that AI is coming for the jobs. The measured version is more useful.
The ILO’s 2025 refined global index assessed nearly 30,000 tasks to estimate occupational exposure to generative AI. Roughly one in four jobs worldwide has some exposure. The share sitting in the highest exposure tier is far smaller.
Exposure measures how much of an occupation's task content generative AI could touch. It is not a forecast of job loss.
The ILO's own reading is that transformation is more likely than replacement for most exposed occupations, because most jobs retain tasks that still need a person. Exposure is a measure of technical potential. Adoption, cost, regulation, and how organisations choose to reorganise all sit between exposure and any real employment effect.
The distinction matters more than it sounds. AI exposure is not job replacement. What it reliably does is move where your value sits: away from producing the artefact, toward deciding what should be built, judging whether the output is right, and carrying the consequences when it is not.
Which is also the honest case for working for yourself. Solopreneurship is not freedom from work — it is taking ownership of the consequences in exchange for choosing the direction. Tooling has made the surface area a single person can cover much larger. It has not made the responsibility smaller.
03 / How to learn when a machine will answer for you
This is the part I would most want a first-year to hear, because the trap is invisible while you are inside it.
Two findings, put side by side.
Freeman et al. meta-analysed 225 studies of undergraduate STEM teaching. Across the 67 studies that reported failure outcomes, average failure rates were substantially lower under active learning than under traditional lecturing.
Undergraduate STEM courses. From the 67 of 225 meta-analysed studies that reported failure-rate data.
Exam performance also rose by about 0.47 standard deviations, roughly 6%. This supports the narrow claim that retrieving, explaining, and applying beats receiving. It does not show that anything labelled 'active' works, or that learners should be left without guidance.
Now the second finding. Recent work on generative AI in learning settings shows that unguarded assistance can raise your visible output while you are using it, and leave you worse off once it is removed. The work looks done. The learning did not happen.
Put those together and you get the rule that matters for the next four years:
That is the whole trick. AI assistance is not the same as learning. The assistance is real and worth using — the failure mode is letting it stand in for the retrieval that actually builds capability.
Practically, for coding specifically: write the broken version yourself before you ask. Read the generated code until you can explain every line to someone else. Delete it and rewrite it from memory once. It is slower for a week and compounds for a decade.
04 / Three projects, four years
If you do one thing with this: pick three projects you will build for yourself before you graduate. Not for a grade, not for a client. For you.
Not three ideas. Three things that meet reality — that get finished badly, get used by someone, and teach you something a syllabus cannot.
- One that solves your own problem. You are the user, so you cannot lie to yourself about whether it works.
- One that someone else has to use. The moment a stranger touches it, you learn everything a spec never told you.
- One that is too hard. You will probably fail it. That is the point — it is where you find the edge of what you can currently do.
The method underneath all three is small and unglamorous: make the idea meet reality early, run small experiments rather than one large bet, and write down what you expected before you find out. That last one is the cheapest habit with the highest return, because it is the only way to tell later whether you were right or merely lucky.
05 / The one that actually matters
Embrace uncertainty and imperfection. It is part of living your life.
I do not mean that as encouragement. I mean it structurally.
You will not get certainty before you have to act — not about your career, not about a build, not about whether the thing you are making is worth making. Evidence reduces uncertainty without removing it. If you wait for the feeling of being sure, you will wait past the point where the decision was still yours to make.
And when it goes wrong, separate the two things most people fuse together. A good decision can still have a bad outcome, and a bad decision can be rescued by luck. Judge your reasoning by what you knew at the time, not by how it landed. Otherwise you will learn the wrong lesson from every result — punishing sound thinking that got unlucky, and rewarding recklessness that happened to work.
Four years is enough time to build something real. It is not enough time to become certain first.
Start anyway. Finish badly. Revise.