I came to the a16z higher-education episodes right after writing that libraries should move at Lindy speed. That order matters.
In the library piece, I was trying to separate one sacred function from the university bundle. Preserve the corpus. Preserve the reading room. Preserve the public waiting place for future readers. Let other university functions face markets, performance, and institutional replacement more directly.
Marc Andreessen and Ben Horowitz make the broader decomposition explicit. Across the two episodes, they stop treating the university as one object. They take it apart: credentialing, courses, research, policy expertise, moral instruction, sports, adult transition, dating, campus justice, accreditation, alumni, administration, government subsidy, employer signal, and the prestige machine around all of it.
That is the right first move.
It is also only the first move.
The harshness in the a16z frame is often earned. The risk appears one step later: once a bundle has been named, the venture mind wants every separated piece to become an opportunity. Some pieces are opportunities. Some are public goods. Some are rituals. Some are legal authorities. Some are slow civilizational memory. Some are social insurance for young adults. Some are moral formation, which may be the least startup-shaped function in the whole pile.
Unbundling works first as a diagnostic. It says: stop letting one word hide many jobs.
The old university did more than bundle many functions. It let each function borrow legitimacy from the others. Credentialing borrowed dignity from learning. Athletics borrowed virtue from campus. Administration borrowed urgency from research. Adult custody borrowed seriousness from scholarship. Moral instruction borrowed authority from science. Research bureaucracy borrowed patience from discovery. The library, the most defensible long-clock function in the bundle, lent moral cover to functions that should have been tested on much shorter clocks.
That is the real rot in a hidden bundle. Weak functions hide behind strong ones. Strong functions get taxed to protect weak ones. Every argument about reform turns into a referendum on the whole institution, so no function has to stand alone in the light of its own purpose.
The test I would add to Marc and Ben's decomposition is plain:
Name the function. Name its proof. Name its clock. Name its carrier.
Instruction has a learning clock. A student either reconstructs and transfers or she does not. Close the source. Solve the next case. Use the idea when the room stops helping. A lecture, a studio, a lab, an apprenticeship, a book, and an AI tutor can all compete here. If a model or tutor teaches calculus better in six weeks, the old course does not deserve four months because it owns a building.
This is where a16z is strongest. Their 1:1 tutoring benchmark matters because tuition has become expensive enough to make absurd comparisons clarifying. If the institution charges more than a personal tutor would cost, it has to explain what else it is selling. Maybe it is selling a powerful bundle. Fine. Name it.
Credentialing has a prediction clock. A credential is useful when it predicts a trait that some downstream actor needs. The employer does not need holiness. The employer needs a signal that survives contact with work. If the diploma mostly says an admissions office once selected the student, call it a selection receipt. If it proves persistence, name persistence. If it proves skill, name the skill. If it proves status, price it as status.
This is where the graph around the last credential cohort matters. The credential apparatus often reads the filter and calls it the signal. A brilliant cohort can emerge from MIT or Stanford and still be evidence about what the apparatus can see, not about where the underlying talent distribution ends. AI makes this distinction more important because it lowers the scaffolding cost for people outside the old intake channel. The credential can no longer rely on being the only way talent becomes visible.
Research has a discovery clock. Basic research cannot be governed like a course. Its output is uncertain, its path is nonlinear, and its future use may be invisible for decades. That argues for public money, philanthropy, prizes, institutes, company labs, and patient funding. It does not sanctify every committee, indirect-cost structure, or publication ritual attached to a university.
Marc and Ben are right to press on the research bureaucracy. They are also right that funding the researcher can be wiser than funding the pre-approved proposal. My addition is the clock distinction: discovery deserves patience; bureaucracy does not become discovery by standing nearby.
Libraries have a civilizational clock. They sit outside the amenity category. A library is a public contact surface between future readers and inherited thought. Destroying it should carry a much higher burden than replacing a course-management system or credentialing pathway. Digitize quickly. Destroy slowly. A society can be pro-market and still decide that memory outranks the market because markets depend on preserved knowledge to know what they are pricing.
Adult transition has a developmental clock. Many eighteen-year-olds need a borderland between family and the labor market. A campus can give distance, peers, rehearsal, and a socially recognized place to become less child and more adult. This is real. It may even deserve support. It becomes dishonest when sold as scholarship alone.
The adult-day-care joke in the a16z episodes is funny because it names a function the institution prefers to hide. But hidden does not mean fake. A society that abolishes the borderland still has to answer where young adults go, how they find peers, how they practice independence, and how they fail without falling too far. The military, a company town, a residential apprenticeship, a monastery, a startup batch, and a university can all serve parts of that transition. They are different carriers for a real function.
Moral formation has an authority clock. Every school teaches a picture of the good life. The question is whether the institution has earned that authority and whether the authority is explicit enough to contest. A university can host moral argument. It can teach intellectual history, law, religion, philosophy, literature, science, and civic disagreement. It becomes dangerous when an administrative consensus arrives wearing the robe of truth.
Here I mostly agree with Marc and Ben's suspicion. The university is a confused moral instructor because it carries several incompatible warrants at once: scientific authority, youth custody, credential power, employment translation, donor politics, and national prestige. That does not make moral formation optional. It means moral formation needs a carrier honest about being moral formation. A church, family, civic order, military service, artistic tradition, local institution, or chosen school can be argued with as that kind of thing. Accreditation should not smuggle it in.
Peer formation has a selection clock. Students buy each other as much as professors. They are buying rivals, friends, collaborators, spouses, future cofounders, future donors, and a local density of ambition. This is why online courses made knowledge cheaper without making college disappear. They attacked instruction and left the peer market, credential, adult transition, employer signal, and parental reassurance mostly untouched.
AI tutoring will repeat the MOOC error if it treats explanation as the whole school. Explanation matters. A model that teaches any student at any hour is a serious event. But education is a product that modifies the user, not content that passes before the user. The accountability test asks whether the learner reconstructs, whether the credential predicts, whether the peer group changes action, whether the user becomes more capable outside the product.
That is where Hari's graph changes the a16z opportunity map. Before asking which university functions can be spun out into startups, ask which functions have startup-shaped feedback.
Instruction can be startup-shaped when learners and payers can see improvement quickly. Credentialing can be startup-shaped if employers trust the prediction and the market can punish bad signals. Research tooling can be startup-shaped when it serves laboratories or companies with visible workflows. Sports can be company-shaped because entertainment and athlete development have paying publics.
Other functions resist that shape. Libraries need public or endowed time. Basic research needs patient uncertainty. Moral formation needs explicit authority. Adult transition needs care, ritual, and guardrails. Peer formation needs selection without reducing every human relation to a market. A startup can participate in any of these, but the function should not be forced to wear startup clothing just because the old university wore too much.
YC is the useful counterexample because it is a school disguised as a fund. It does not pretend to be a university, and that is part of why it works. Its functions are relatively explicit: selection, compressed curriculum, peer density, apprenticeship through alumni, capital, legitimacy, and a charter long enough for students to become faculty. The credential is the batch. The classroom is the company. The exam is the market. YC bundled functions, but the bundle stayed close to the reality test it claimed.
That is the standard for a new education institution. Not unbundled by default. Honest by design.
A new school can bundle. A research lab can teach through apprenticeship. A residential cohort can make instruction stick. A credential can summarize a demanding path if the path actually demanded the trait. A campus can give a young adult the distance from home required to become reachable by serious work.
The crime is the hidden bundle: the place where the sports league, moral office, credential mill, research bureaucracy, adult holding pen, and library all point at one another whenever asked to justify themselves.
The next education system should make a different bargain. Bundle where bundling strengthens the function. Separate where separation clarifies it. Fund what needs a long horizon. Market-test what has a short feedback loop. Name moral authority when moral authority is being exercised. Make credentials say what they predict. Make instruction prove reconstruction. Make research institutions distinguish patience from administrative self-protection. Make libraries slow to destroy.
This is the bridge from the a16z episodes back into the graph. Marc and Ben are right that the university is a bundle. Hidden bundles cross-subsidize, launder legitimacy, and confuse clocks. The library argument adds the exception: not every function should move at startup speed. The credential argument adds the warning: every filter has to prove it is a signal. The learning argument adds the test: explanation has to become reconstruction. The YC case adds the carrier: a school can hide inside another form if the functions remain honest.
The classroom was never the whole school. The school was the social wrapper that told a society how young people become trusted.
Repair starts when the wrapper stops pretending to be one thing.