Is a Computer Science Degree Still Worth It When AI Writes the Code?

Every family we work with asks some version of this now, usually in the same nervous half-sentence: is it still worth it, with the AI stuff?

The answers on offer are useless in both directions. One camp says nothing has changed and the degree is as good as it ever was, which is not what the labor data shows. The other says the field is finished, which is not what anyone's payroll shows. Both are arguments about vibes, made by people who have not put the numbers side by side.

Our answer is yes — emphatically yes, and we would send our own future kids — but the reason is not the reason it was yes ten years ago, and the difference between those two reasons is the entire decision.

The part that really did collapse

Start inside engineering organizations, because this is where the exaggerations live and the exaggerations are doing real damage to seventeen-year-olds' choices.

Stripe's internal coding agents — Minions — open more than 1,300 production pull requests per week. Not autocomplete, not suggestions: end-to-end tasks on a codebase clearing over a trillion dollars a year in payment volume, containing no human-written lines. That figure came from Stripe's own engineering org in early 2026.

Ramp built an internal coding agent that now accounts for roughly 30 percent of its engineering pull requests. Meta's reported internal targets put 55 percent of engineering code changes as agent-assisted company-wide, with some organizations pushing selected teams toward 75 percent during 2026.

Those are the published numbers. The private ones are higher. Ben, one of our co-founders, sees directly everyday in his own team, and broader orgs at Meta, that on their teams the practical figure is closer to everything — that essentially all committed code now originates from an agent, with the human work being direction and review. Treat that as what it is: credible first-hand reporting, not an audited statistic, and definitionally slippery in the way all of these numbers are, since "originates from an agent" and "shipped exactly as generated" are very different claims. But every engineer we know inside a large company tells a version of the same story, and none of them tell the opposite one.

You do not need the highest number for the conclusion to hold. Producing a working implementation from a clear specification is no longer scarce. It was scarce for about fifteen years, it made a great many people wealthy, and it is over.

If your model of a CS degree is vocational training in producing working implementations, that model is obsolete, and no college brochure changes it.

The interview repriced before the curriculum did

The clearest evidence this is real rather than projected is that hiring already moved, and hiring is the most conservative process at any large company.

Meta introduced AI-assisted coding interviews in October 2025: a CoderPad session with frontier models sitting in the window, initially for mid-level and senior software roles, extending through 2026. Shopify, LinkedIn, Rippling, and Canva have gone the same direction. The candidate is handed the tool and watched using it.

Consider what that means as a measurement decision. For two decades the technical interview asked you to reproduce, from memory and under time pressure, the solution to a problem with a known answer. That was never interesting in itself. It was a cheap proxy for can this person think structurally, and it worked only while producing the code was hard.

Once every candidate can produce the code, the proxy measures nothing. So the interview stops asking for code and starts asking for what survives: can you specify the problem precisely, can you tell when a confident answer is wrong, can you say why this implementation is correct and what it costs at scale, do you notice the edge case nobody prompted for.

I would not tell a high schooler that competitive programming is worthless — the reasoning it builds is real, and students who love it should keep going. I would tell them its value moved from credential to training, and that the move is already priced in at exactly the companies they are aiming at.

The labor data, both halves of it

Here is the part nobody quotes in full. The New York Fed tracks outcomes for graduates aged 22 to 27; its February 2026 release, on 2024 census data, contains two findings about computing majors that are usually cited separately by people arguing opposite sides.

Computer engineering has the highest early-career median wage of any major in the dataset and the second-highest unemployment rate of any major in the dataset. Computer science is just behind on both. Special education, with essentially no unemployment, pays a fraction of either.

Read only the left column and computing is a trap. Read only the right and it is obvious. Both readings are wrong, because they are reading one number where there are two.

That table does not describe a good bet or a bad one. It describes a high-mean, high-variance bet, and the two properties have one cause. Employers are paying more than ever for the engineers they want and hiring fewer of them, because the work a marginal junior engineer used to do is the work that automated first. The distribution did not shift down. It spread out.

Underemployment is the number I actually watch, and it is worse than the headline: 42 percent of all recent graduates hold jobs that do not require a degree, the highest share since 2020. Unemployment tells you whether you have a job. Underemployment tells you whether the degree bought you anything.

What decides which tail you land in

Go back to the Stripe number and to the clause everyone drops when they quote it. Thirteen hundred agent-written pull requests a week — every one human-reviewed before it merges.

That is the next decade in a sentence. The cost of producing a candidate solution fell to roughly zero. The cost of deciding whether a candidate solution is correct did not fall at all. When one side of a pipeline gets a thousand times cheaper and the other does not, the expensive side becomes the whole job. Stripe did not stop needing engineers when the Minions started shipping. It needed engineers who could read a thousand plausible diffs a week and find the one that quietly breaks idempotency on a payment retry.

You do not do that by prompting. You do it by knowing what a race condition is, what a transaction actually guarantees, what the complexity of the thing you just approved is, where floating point will bite, what happens when the network partitions. That knowledge has a name. It is the second half of a computer science curriculum — algorithms and complexity, operating systems, computer architecture, distributed systems, databases, programming languages and types, numerical methods — and it is exactly the half students have spent a decade treating as the annoying prerequisite between them and an internship.

The unpopular half of the major is now the entire product.

And then there is the half nobody puts in the brochure

Skills are one of two things a great department sells. The other is who is standing next to you, and it has become the more valuable of the two, not the less.

PitchBook tracks it directly. Its 2025 university rankings counted more than 173,000 venture-backed founders worldwide and sorted them by undergraduate institution, over roughly a decade of funding activity:

Five undergraduate programs, roughly seven thousand funded founders, and something north of four hundred billion dollars of capital between them. The pattern repeats at the front of the funnel: in Y Combinator's Winter 2026 batch, Berkeley alone placed 30 founders, Stanford 22, Harvard 18 — three schools supplying about 16 percent of a batch drawn from the entire world.

But the aggregate understates it. Skip the tables and read down the roster of companies your student actually names when you ask what they want to build:

That is not a list assembled to prove a point. It is what you get by writing down the best-known companies of the last fifteen years and looking up where the founders went. The set of schools does not run out. It repeats — and the last row is there because it is the honest exception, and there are others.

And it repeats in a specific way that is the actual finding here. Ilya Strebulaev's group at Stanford, which studies unicorn founders against a matched random sample of VC-backed founders, counts 147 unicorn founders from Stanford, 117 from MIT, and 86 from Harvard. The six leading engineering schools — Stanford, MIT, Berkeley, Georgia Tech, Carnegie Mellon, Caltech — produced 16.3 percent of all unicorn founders, more than the entire eight-school Ivy League at 13.8 percent.

Then the number that says the quiet part: among schools with twenty or more unicorns, roughly 29 percent of MIT's and 28 percent of Stanford's had two or more co-founders from the same school. Nearly a third of those companies were started by people who met on campus.

That is not credentialism. A diploma does not co-found anything. What that statistic measures is proximity — that the person who becomes your co-founder is disproportionately someone you sat next to in a systems class at nineteen.

Jason made the parallel argument about outcomes in The Ivy-Plus Premium Is Narrower Than You Think, and the honest caveats there apply here with force. These are absolute counts, not rates: Berkeley graduates an enormous number of people, so a large founder count is partly a large denominator, and PitchBook credits a founder to every school they attended. Read the capital column against the founder column and you can watch the confound work — Penn produced fewer founders than Berkeley and nearly twice the capital.

The rate data cuts against the simple story even harder, and it deserves saying plainly. On Strebulaev's odds ratios — a school's share of unicorn founders divided by its share of ordinary VC-backed founders — the leaders are not the famous names. Utah runs 3.7×, Washington 2.6×, Arizona 2.2×; MIT is 1.8× and Stanford 1.7×. About a quarter of US unicorn founders earned their degrees abroad. And 84 percent of that YC batch came from somewhere other than the top three schools. Plenty of people build enormous things from everywhere, and any consultant who tells you otherwise is selling.

So do not read this as attend one of these five and you will raise a round. Read the narrower thing it actually supports: the people who start and fund technology companies are not distributed evenly across campuses, and they never have been. Founders meet co-founders in dorms, labs, problem sets, and clubs. Capital flows along paths people already know. That is not a moral fact about merit; it is a structural fact about how trust and information move, and it is the single hardest thing on this page to acquire any other way.

Which matters more now, not less. When the mechanical part of building software costs nothing, the scarce inputs become the problem worth solving and the people willing to solve it with you. A degree cannot give you the first. A strong department, at a school where the person in the next seat is building something at midnight, is still the most reliable way anyone has found to buy the second.

You do not have to found anything for this to matter

Founding is the vivid case, and it is also the rare one. The same concentration shows up in ordinary hiring, which is where nearly all of your student's classmates will actually land — and at the top of that market the numbers have gotten strange.

Look at quantitative trading, the highest-paying destination available to a twenty-two-year-old with a computing degree. Jane Street new-grad traders are reported at $250k–$375k base with sign-ons up to $200k, first-year totals landing roughly $450k–$650k. Citadel Securities lists $250k–$350k base for quantitative researchers, with first-year totals reported in the $400k–$700k range once the discretionary bonus arrives. Hudson River Trading sits in the same neighborhood. These are compensation figures for people twelve months out of a bachelor's degree, and they are being paid at firms where the models, the execution, and increasingly the research tooling are all AI-driven. The machines did not depress the pay. They raised the ceiling on what a small number of very well-trained people can be worth.

Now look at where those firms recruit. The reporting is consistent and unglamorous: MIT, Harvard, Princeton, Stanford, Caltech, Carnegie Mellon, Berkeley, Chicago, Waterloo. Citadel's intern class is described as drawing disproportionately from that same short list. The pattern holds outside finance — Palantir recruits hard at Stanford, MIT, CMU, Penn, Duke, and Cornell for a new-grad class of ~100 people in a pool of tens of thousands of applications; Google and Amazon run structured campus pipelines concentrated on a comparable set of research universities and engineering flagships.

I want to be careful about what is and is not public here. Firms do not publish per-school hiring counts, so anyone quoting you "Palantir takes X percent of its class from Penn" is estimating. What is observable is the recruiting calendar: which campuses get the on-site rounds, the info sessions, the sponsored clubs, the returning alumni who forward a résumé. That list is short, it is stable, and it is the mechanism.

It is also not a wall. The same recruiters say plainly that demonstrated ability — olympiad and competition results, real research, a project with users — overrides pedigree, and Palantir now runs a fellowship that hires talented people straight out of high school with no degree at all. Polymarket's founder left NYU in his first year. If your student is exceptional and unlucky in admissions, the door is open; it is just narrower, and they will have to build the evidence themselves that a target-school student gets handed by the calendar.

Where the argument is genuinely weak

The strongest objection deserves naming rather than burying.

Reviewing is a senior skill. Historically people became good at judging code by writing a great deal of mediocre code first, on low-stakes tasks, with someone more experienced correcting them. That apprenticeship ran on precisely the work now automated. If the bottom rung is gone, it is not obvious how this cohort reaches the rung above it, and "the fundamentals will carry you" is a claim about a pipeline that has not been tested end to end. Anyone who tells you they know how it resolves is guessing.

The market is already pricing that uncertainty. CRA's 2025 Taulbee Survey, published this spring, found bachelor's-level computing enrollment down 3.1 percent among units reporting in both years — a near ten-point swing from a 6.6 percent increase the year before. Students read the same headlines you do, and they are backing away.

Read that number twice, though, because it cuts the other way for anyone still willing to work. The founder tables above did not thin out. The capital did not leave. What thinned out is the number of eighteen-year-olds prepared to take on a hard major with an uncertain first job — which is a description of reduced competition for the seats, the labs, and the collaborators that produce everything in the section above.

It also produces the strange condition of this cycle: the applicant pool for computing is cooling while admission to computing is not. Direct-admit CS programs, capped majors, and internal-transfer restrictions were all built during the boom, and institutional policy lags reality by years. The seats did not expand. The old workaround — apply to the easier major and switch in sophomore year — is closed at a growing number of schools, and closed most firmly at exactly the public flagships families treat as the safe option.

What we tell our students

Choose the department, not the label. Two programs with identical names differ enormously in how much theory, systems, and proof they require. Read the actual degree requirements on the department site, not the admissions page. A core heavy on architecture, systems, and formal reasoning is selling the durable half. One that has drifted toward framework tutorials is selling the half that just got automated.

Verify major access before you build the list. Per school: is CS direct-admit, is it capped, what is the real internal transfer rate. That last number is often published and often brutal. It is the most expensive common mistake we see in a STEM list.

Treat the network as coursework, not as luck. Nearly a third of MIT's and Stanford's unicorns had two or more co-founders from the same campus, and that number is not produced by enrolling. It is produced by research groups, project teams, hackathons, and upperclassmen who graduate two years ahead of you into rooms you want to be in. A student who leaves at twenty-two with a strong transcript and nobody who would vouch for them has bought half of what they paid for.

Change what your evidence looks like. An activities list built on contests and coursework was legible in 2019. What reads as real now, to an admissions officer and later to an interviewer, is narrower and harder to fake: you built a thing someone else uses, you measured it, it broke, you found out why, and you can explain in plain language why the fix was correct. That is a verification story told at seventeen, and it is the same skill the labor market is now paying a premium for.

The degree is worth it. It is worth it for judgment and for people — the two things that got scarcer, not cheaper, the moment the code stopped being the hard part. And it is worth it right now, at the exact moment a measurable share of your competition has decided it is not.

Sources: Stripe engineering via InfoQ · Ramp coding agent, InfoQ · NY Fed, via Forbes · PitchBook university rankings 2025 · Strebulaev, best US universities for unicorn founders · YC W26 batch analysis · CRA Taulbee Survey 2025 · HFT hiring and compensation, 2026 · Palantir campus recruiting, SF Standard · Meta AI-assisted interviews