Context

Erik Torenberg talks with a16z cofounder Ben Horowitz and Gagan Biyani (cofounder of Udemy and Maven, now CEO of the new Horowitz and Andreessen Academy) about why they're building a project-based school for young people who already know they want to build, rather than a general college alternative. The conversation matters to PMs and hiring managers because it's a concrete argument about which skills (judgment, people skills, tolerance for productive failure) can't be taught in a classroom and have to be developed by actually doing something with real stakes, a question that applies just as directly to how companies onboard and develop junior talent as it does to education.

The Big Idea

The skills people actually need now (judgment, people skills, and the ability to use powerful AI tools to solve real, open-ended problems) can only be developed by doing real projects with real consequences, not by studying about doing them, and most attempts at reinventing education have failed because they optimized for the learning piece alone instead of building the full bundle college actually provides: training, credential, community, and fun.

Biyani's specific diagnosis of why earlier "college alternative" efforts didn't cut through: they treated education as a single, isolated function, when college is actually a bundle of a dozen different things (status, credential, brand, social structure, fun) and a genuine alternative has to competitively replace the whole bundle, not just the classroom instruction part of it.

Key Insights

A viable school for builders needs the same bundle college provides, not just better instruction

Horowitz and Biyani both argue that most attempts to disrupt higher education failed by narrowing scope too far, building just an online course platform or just a credentialing product, without also delivering the social community, residential experience, brand signal, and direct employer connections that make college function as a package. Their explicit design response is to build all of it together: a residential program in San Francisco, cohabitation with real companies and projects, deliberate cross-cohort social structure, and formal pipelines into a16z's network of portfolio companies, on the theory that a partial substitute for college won't beat college, only a comparably complete one will.

Timing depends on a real customer existing, not just a real idea

Biyani frames the Academy's viability using a standard early-stage startup lens: you need a specific, existing customer with a burning, already-present need, not a theoretical one. His evidence that this moment is different from earlier attempts: over the past six months he's been directly embedded with a growing population of young people already building real projects and companies on their own initiative, well before any formal program existed for them, which he treats as the demand signal that was previously missing or too thin to build a viable school around.

Judgment and people skills require real stakes, not lecture content

Horowitz's clearest articulation of why this can't be taught didactically: he compares it to learning football from a book, no amount of reading about how to play quarterback prepares you for the actual experience of a 300-pound lineman rushing at you while you're deciding who to throw to. His own leadership books, he says explicitly, don't teach someone to be a CEO by themselves; they're reference material that only becomes useful once someone is actually in a high-stakes decision and needs to recognize the pattern. The Academy's design response is to put students in situations with real (if lower-stakes than an actual company) consequences, forming a team, driving a project to completion, resolving conflict, so they can begin acquiring judgment through direct experience rather than only reading about it.

Immersion works through forced necessity, not curriculum design

Biyani's example of how the "osmosis" element is meant to function: a student trying to build a specific hardware project needs a particular chip, doesn't know how to write a cold outreach email, and is surrounded by people at partner companies who do, so they have no real choice but to learn that skill in order to make progress on something they already care about. His broader claim is that this kind of forced, need-driven skill acquisition, inside a peer network and physically embedded among working companies, develops capability faster than an equivalent classroom lecture on the same topic (his direct example: a negotiation lecture versus needing to actually get a team aligned behind a real deadline).

AI-resistant assessment means designing problems, not administering tests

Horowitz relays a framing from a Stanford cryptography professor he cites (Dan Boneh) on how to handle AI in assessment: banning AI tools doesn't work, but making the problem hard enough that a student can't solve it without meaningfully using AI does, and that professor reports students now solving problems neither the students nor the professor could have solved unassisted before. The Academy's stated design principle follows this logic directly: assess "did you build something interesting" rather than "can you answer questions correctly," since the tools that can now trivially answer most test questions can't substitute for the judgment needed to actually complete an ambitious, open-ended project.

Productive failure is defined narrowly: you have to actually learn something specific from it

Horowitz distinguishes Silicon Valley's supposed tolerance for failure from a vaguer "failure is fine" attitude: the failure that has real value is the kind where you attempt a genuinely hard problem, don't solve it, and come away knowing specifically why it was hard, which then reveals the actual sub-problems or adjacent problems worth attacking next. He frames this as close to how scientific discovery and entrepreneurship both actually work in practice, where the original hard problem often isn't the one that ends up mattering; a different, related problem discovered along the way frequently becomes the real opportunity. His explicit takeaway for young builders: don't settle for a smaller, safer idea just because the big one didn't immediately work, use the specific knowledge earned from the failed attempt to find the adjacent problem worth solving instead.

Mental Models & Frameworks

The completeness bundle test for any "alternative" product

Biyani's framework for evaluating whether an alternative to an entrenched, multi-function incumbent (college, in this case) can actually compete: identify every distinct function the incumbent bundle serves (here: job training, status and credential signaling, social community, and enjoyment), and check whether the alternative product delivers a competitive version of each one, not just the function that seems most obviously improvable. A product that wins on one dimension (better, faster instruction) while ignoring the others (community, credential, fun) will lose to the full incumbent bundle even if that one dimension is genuinely superior.

Make the assessment problem AI-resistant instead of banning AI

A direct, reusable design principle for any evaluation process in an environment where capable AI tools are available: rather than trying to detect or prohibit tool use, design the problem itself to be hard enough that clearing the bar requires genuine skill in directing the tools well, not just having access to them. Use this whenever you're designing an interview process, an academic assessment, or an internal skills evaluation in a domain where AI assistance is already ubiquitous and undetectable.

Practical Application

Evaluate junior hires and new grads on completed, self-directed projects over credentials

Following the Academy's assessment philosophy, when evaluating early-career candidates, weight evidence of something they actually built and finished (even a small, imperfect project) more heavily than credentials or test performance, since Horowitz and Biyani argue the tools now available make "can answer the question correctly" a weak signal compared to "can direct these same tools toward an ambitious, open-ended outcome."

Force skill acquisition through real project necessity rather than training modules

If you're onboarding junior team members, consider whether assigning a real, slightly-too-ambitious project with genuine stakeholders and constraints (the way Biyani describes a hardware project forcing a student to learn cold outreach) develops a specific missing skill faster than a corresponding training module or lecture would. The gap being filled should come from what the project genuinely requires, not from a generic curriculum.

Debrief failed initiatives for the specific sub-problem uncovered, not just the outcome

When a project or initiative on your team fails to hit its original goal, follow Horowitz's framing explicitly: ask what specific, previously unknown difficulty the attempt revealed, and whether that difficulty points to a narrower, adjacent problem worth pursuing next. Treat "we learned exactly why this is hard" as the actual deliverable of the failed attempt, distinct from and sometimes more valuable than the original goal itself.

Bottom Line

Horowitz and Biyani's core bet is that the skills the AI era actually rewards, judgment, people skills, and the ability to direct powerful tools toward ambitious, open-ended problems, can only be built through real projects with real stakes and a complete surrounding ecosystem (community, mentorship, employer access), not through better lecture content or tighter test design, and that most previous attempts at reinventing education failed by building only a partial substitute for that full bundle.