Let Them Vibe-Code the Thesis

Διαβάστε το και στα ελληνικά.

I owe you a post about seniors and juniors. It's still coming. But something jumped the queue: the Senior Thesis is having a moment. Right now, institutions in a lot of places — including the one where I teach — are talking about updating their final-year project rules for the AI era. And the edits being discussed are quiet, sensible ones: student engagement during the build starts to count — not just the paper, the product, and the video presentation handed in at the end — and the viva gets refocused on two questions: is this genuinely your work, and do you understand what you built?

That's the right direction, and I want to say so before I start pushing. Because the same conversations tend to leave one thing untouched: AI remains something to be used "very sparingly, always referenced" in the build, and not at all in the report — which still gets run through a detector for an AI-generation score. Half of the emerging rulebook is walking toward where software actually went. The other half is guarding a door the industry already walked through.

What the thesis was designed to prove

Call it what your institution calls it — Senior Thesis, Final Year Project, dissertation, διπλωματική. Underneath the name it's the same artifact: proof that one student can take a problem from open question to working system, over months, largely alone. It is a certificate for the heroic solo builder — the "leave me alone in a basement with pizza and coffee and I'll bring you the best code you've ever seen" engineer.

Here's the uncomfortable part: that's the exact job description AI agents are best at. The cement post made the argument for the industry — the mixing and pouring of software, the actual typing, is becoming the crew's job, and the crew is made of agents. The engineer who thrives now is the one who can gather requirements from a confused human in the morning and specify a task precisely enough for a machine in the afternoon. Communication, in both directions, with both species of colleague.

So the thesis, in its classic form, certifies a role the market is quietly retiring — not because students got worse, but because the question got old.

Two theses, officially

I said it in the classroom post and it bears repeating with more force: research theses are a different conversation. Work that tries to push the field itself forward — including building the next generation of these very tools — still needs the slow, from-scratch, nobody-can-shortcut-this rigor. Some students genuinely want that path. Keep it, protect it, and hold it to the standard it has always had.

But most Bachelor's theses were never research. They were "build something real and prove you can." And the students showing up for that kind of thesis have changed. There's a growing crowd of eighteen-year-olds who look at a four-year degree and see a detour — they don't want to research, they want to build, and a free hosting tier plus a weekend gets a site live before Monday's lecture. If the build-style thesis stays a from-scratch typing exam, the university is answering a question those students stopped asking.

So split it formally. A research track that stays sacred. And a build track that is redesigned around AI orchestration on purpose — not tolerated as cheating-lite, but taught, supervised, and graded as the actual discipline it has become.

What a vibe-coded thesis actually teaches

The internet settled on a name for building by directing an AI: vibe coding. I have friends who teach in CS departments — professors and lecturers, in Greece and in the United States — and some of them hear that name and reach for the red pen. Let me put a real project on the table first.

PlanItShare — the app this blog lives on — went from an empty folder in early July to version 1.41.0 by the first of September. Forty-one iterations, more than a hundred database migrations, and not one line of it typed by me.

Now the part that matters: "not one line typed" is not "not one decision made." I've spent a career with SQL, and I read every one of those migrations before it touched the database. The AI proposes interface designs; I give it specific direction and pick the winner — even with a designer persona loaded, taste and judgment stay my job. When a dev server hangs, I'm the one hunting down the orphaned process. This is delegation with your eyes open, not absence.

And the build kept teaching lessons no lecture of mine ever delivered this cheaply. Two of my favorites. The app used to derive a map pin's country from the nearest airport in a bundled dataset, within 150 kilometers. Nearest "airport" to a pin on Jerusalem? Ramallah Heliport, 15.9 km away — Ben Gurion is 41.1. The feature compiled, the tests passed, the demo looked great, and the app was confidently publishing a country label on contested land. A political statement, wearing a feature's clothes. Second: a research pipeline cited thaiembassy.com for a passport rule — a commercial visa agency, one letter away from the real Royal Thai Embassy network at thaiembassy.org — and a careful human approved it, because the review screen showed the URL as one monospace line among twenty-five. Provenance, it turns out, must be checked by code, not by attention.

By the time I wrote them all down there were five case studies and ninety-five classroom-ready lessons, covering thirteen concepts from AI calibration to honest denominators — and almost none of them are about syntax. They fell out of one app, in one summer, as a side effect of building it carefully. (They now live as a Lessons section inside the app itself — it sits behind the free sign-in, so consider this the field trip permission slip.) A student who builds this way — keeping decision records, catching what the model got wrong, writing down why — is not doing a lesser thesis. They're doing the one the job market is currently buying.

The viva was the answer all along

The friends with the red pens have a nightmare, and it's a legitimate one: a student submits twenty thousand lines they cannot explain, the defense becomes theater, and the degree certifies nothing. I've supervised enough AI output to know the model will happily produce those twenty thousand confident lines. Unsupervised, they ship.

But notice what the updated rulebooks already contain: a viva refocused on ownership and understanding, and engagement tracked across the whole build. That is the cure — taken seriously, it makes the AI ban redundant. If a student can explain why the schema has that index, which week the model went confidently wrong, and how they caught it, they have done the job the market pays for, whoever typed. If they can't, you don't need a detector to fail them. You need a second question.

Grade the build-track thesis the way a senior engineer's pull request gets reviewed, plus a defense: the decision log over the months, what was delegated, what came back wrong, what changed and why. That's a harder rubric to write than a percentage. It's also the only one measuring the real thing — we already watched a detector clear two AI-drafted posts on this very blog as 0% and 6.4% "human written." The number was never the thing. The judgment was.

Keep one room sealed

None of this abolishes the hard parts. For some years yet, I'd argue the fundamentals — object-oriented programming, data structures, algorithms — should still be taught by hand and examined in a sealed room: university-managed tablets, no internet, no model whispering in anyone's ear. Not out of nostalgia. Out of engineering necessity: the human in the loop has to understand what's being built, or the loop has no human in it. You don't catch the heliport publishing a border if you don't know what the query is doing. You can't referee an agent writing SQL if a JOIN is a rumor to you.

The sealed exam and the vibe-coded thesis aren't rivals; they're the two halves of one certificate. One proves understanding exists. The other proves it can be deployed through a crew of agents and survive a viva about it.

The square-root horizon

Ask your class to compute a square root by hand, on paper. Their grandparents could. Nobody mourns it — calculators absorbed the mechanics, and bridges did not get less safe. Someday, reading code may follow the same road; the graduates of 2040 might treat a for-loop the way we treat long division, and that will be fine, then.

But the class of 2027 graduates into the transition, not the destination. They need both rooms: the sealed one where fundamentals get into their hands the slow way, and the open one where they learn to run a crew — with the viva standing at the door, asking the only question that ever mattered: do you understand what you built?

One more confession

This series has a habit of confessions, so here's how this post came to be. I was never the essay student. School compositions fought me the whole way, and at university I wrote exactly as many papers as the degree demanded and not one more. What I've always had is ideas, and a teacher's compulsion to hand them to other people — it's the reason I stand in classrooms. For most of my life, those two facts refused to connect on paper: the craft of prose sat like a toll gate between what I understood and what I could publish. You already know from the earlier confession how these posts get made — AI drafts, I supply the stories, the corrections, and the final call. What that changed for me personally is the toll: the road between knowing something and sharing it in writing is, for the first time, open.

Which is why the "no AI in the report" rule deserves one last look before I go. In every cohort there is a student who understands their build completely and dreads the sixty-page write-up more than any part of the code. The old rules grade that student's prose craft alongside their engineering, whether anyone intends to or not. Let the report be drafted the way this post was — and then let the viva do exactly what those updated rulebooks ask of it: check that the understanding is theirs. That isn't lowering the bar. It's finally pointing it at the right thing.

Next stop — and it has now survived two queue-jumps, so it's earned the slot: why so many companies are chasing one AI-fluent senior instead of training three juniors. Still on the itinerary after that: what to hand a curious teenager long before any of this reaches a syllabus, and the line item finance departments mumble about — tokens and subscriptions next to what payroll used to cost.