In early 2025, Nvidia CEO Jensen Huang made a remark that landed like a grenade in academic common rooms everywhere: that it essentially doesn’t matter what kids study anymore, because AI will handle so much of the technical execution that the field of study has become secondary. It’s a provocative claim. And I think he’s more right than the education establishment wants to admit — but also missing something critical that makes his framing dangerous if taken at face value.
The Question No One Is Asking
The predictable reaction to Huang’s statement has been defensive. Educators push back: of course domain expertise matters. Doctors need to understand biology. Engineers need physics. Lawyers need to understand statute. And they’re not wrong. But they’re answering a question Huang wasn’t really asking.
What Huang is actually pointing at — and what almost nobody in the education debate is confronting directly — is this: when AI can write the code, synthesise the literature, draft the legal brief, and generate the data model, what exactly is the human’s job? That’s the question no one is asking with the seriousness it deserves. The conversation keeps circling around “which skills are safe” and “which jobs will survive,” when the real inflection point is about what kind of thinking AI cannot yet replicate at production quality.
We’re not debating course catalogues. We’re debating what human cognition is for in a world where a well-prompted large language model can pass professional licensing exams, generate peer-review-worthy literature reviews, and produce working software from a plain-English description. That’s a fundamentally different civilisational question, and Jensen Huang — perhaps more than any other tech executive alive — has the vantage point to see it clearly. His company’s chips are, quite literally, powering the transformation he’s describing. As we’ve covered in our analysis of Nvidia’s Vera CPU and the $200B agentic AI market, the infrastructure buildout is not slowing — it’s accelerating. The shift Huang describes isn’t theoretical; it’s already in production.
Why It Matters
Before the current AI wave, the educational bargain was relatively stable: study a technical or professional field, accumulate domain knowledge that’s expensive to replicate, and that expertise becomes your career moat. A radiologist knew how to read a scan. A software engineer knew how to write production-grade code. A financial analyst knew how to build a discounted cash-flow model. These skills took years to acquire and were genuinely scarce.
What changed — specifically around 2022–2023, with the public release of capable large language models and their rapid capability gains — is that the execution layer of most knowledge work became automatable at a pace that outstripped any reasonable curriculum reform cycle. Universities plan in five-year horizons. AI capability is advancing on a six-to-twelve-month cadence. That mismatch is now a structural crisis, not a future risk.
This is the inflection point Huang is naming. And it’s real. The question isn’t whether a biology student needs to understand cellular mechanisms — of course they do. It’s whether spending four years memorising the execution details of tasks that a well-configured AI agent will perform for them is the right allocation of human developmental time. I’d argue it increasingly isn’t.
Here’s the synthesis that the “just learn to prompt” crowd and the “nothing has really changed” traditionalists both miss: the value of domain education is shifting from knowing how to do things to knowing enough to evaluate whether the AI did it correctly. That’s a subtly but profoundly different cognitive posture — it requires conceptual depth without necessarily requiring procedural mastery. A medical student who deeply understands pathophysiology but doesn’t memorise every drug dosage is arguably better prepared for an AI-augmented clinical environment than one who has drilled both equally. The curriculum implication of that shift is enormous, and almost no institution is designing for it yet.
My Answer
I believe Huang is right in one specific, important sense: the choice of major matters far less than it used to as a predictor of career outcomes. What matters enormously — and what almost no one in the mainstream education debate is centring — is the cultivation of judgment, taste, and the ability to ask the right questions. These are the skills that sit above the execution layer that AI is colonising.
The students who will thrive in the next decade are not necessarily the ones who studied computer science or AI. They’re the ones who studied anything deeply enough to develop strong intuitions about what “good” looks like in a domain — and who learned to communicate those intuitions clearly enough to direct intelligent tools. A history student who can formulate a precise, well-scoped research question and critically evaluate a sourced argument is, in my view, better positioned than a computer science student who can write code but has never been asked to defend an interpretation against a sceptic.
This reframing has practical consequences. It suggests that the humanities and social sciences — perennially under threat from STEM-focused curriculum pressure — may actually be undervalued in the AI era, not because AI can’t do STEM tasks, but because AI is especially good at them. As we’ve explored in our piece on data nihilism and AI’s appetite for human-generated knowledge, the most scarce input to AI systems is high-quality human judgment about what matters and why. That’s not a data engineering problem. It’s a liberal arts problem.
The Counter-Argument
The steel-man version of the opposing view goes something like this: Huang’s claim is the luxury belief of someone who runs a hardware company and benefits when people stop worrying about AI displacement. If everyone believes domain expertise doesn’t matter, universities hollow out their rigorous programmes, we produce a generation of generalists with shallow knowledge, and then — when AI makes a consequential error in medicine, law, or engineering — there are no true domain experts left to catch it.
This is a serious argument. And it contains real truth. There is a legitimate risk that Huang’s framing, popularized without nuance, becomes a permission slip for intellectual laziness. “AI will handle it” is a dangerous mental model if it means students never develop the deep domain intuition needed to supervise AI outputs. The problem of AI-generated content requiring human cleanup is already well-documented, and it scales badly without domain-literate humans in the loop.
But — and this is where I part company with the traditionalist reflex — the solution is not to preserve pre-AI curricula unchanged. It’s to redesign education around the specific cognitive tasks that AI cannot yet perform reliably: forming original hypotheses, exercising ethical judgment in ambiguous situations, building trust through interpersonal accountability, and maintaining the conceptual map of a domain even when the implementation details are outsourced. The choice is not “deep expertise vs. generalist prompting.” It’s “what kind of depth matters now?”
What the AI Era Education Story Is Missing
Huang’s provocation has generated enormous reaction, but even the most thoughtful responses tend to leave several critical dimensions unexamined:
- The equity dimension is nearly invisible. The “major doesn’t matter, develop judgment” argument implicitly assumes access to the kind of rich, intellectually diverse educational environment where judgment gets trained. For students at under-resourced institutions — or in countries where education is narrowly vocational by necessity — the advice to “study what you love and let AI handle the rest” is not actionable. This conversation needs to grapple with who gets to benefit from AI’s execution layer and who gets displaced by it. The work on AI’s uneven economic effects on blue-collar versus white-collar workers is directly relevant here and largely absent from the education debate.
- The verification problem is underweighted. Even if we accept that AI handles execution, humans still need enough domain knowledge to verify outputs — especially in high-stakes fields. How much is “enough”? Nobody is seriously designing curricula around a minimum viable domain depth for AI supervision. That’s a research and policy gap that urgently needs filling.
- Institutional incentives are ignored. Universities are assessed on graduate employment rates, research output, and enrolment numbers — none of which currently reward the kind of curriculum restructuring Huang’s thesis implies. Without changing the incentive structure, the advice to rethink education remains aspirational. The governance question — who decides what gets taught, and by what criteria — is conspicuously absent from the public conversation.
What Changes If I’m Right
If the shift Huang is describing is real — and I believe it is — then the second-order effects on research, academia, and industry are significant. Graduate programmes will face pressure to justify their existence as credential-granting institutions if the credential no longer signals execution competence. The research literature itself will need new norms for distinguishing AI-assisted synthesis from original insight. Peer review, already under strain, will need methodological standards for AI-generated content that the community has barely begun to develop.
Beyond academia, companies will redesign hiring to test for judgment and domain taste rather than technical execution — a shift that advantages people who can articulate why something is correct or beautiful or well-reasoned, not just produce it on demand. This connects directly to the broader conversation about AI alignment — because the humans directing AI systems at every level, from enterprise tools to frontier research, need strong enough judgment to recognize misalignment when they see it. That’s a human capacity, and it has to be developed somewhere. It might as well be in school.
Signals to Watch
University enrolment shifts by discipline. If Huang’s framing takes hold, watch for accelerating enrolment declines in traditionally “safe” technical majors — not just humanities — as students struggle to identify which degrees offer durable value. A reversal (unexpected growth in philosophy, rhetoric, or interdisciplinary programmes) would be a leading indicator that the market is responding to the judgment-over-execution thesis.
Employer hiring criteria in job postings. A meaningful shift away from degree and skill-stack requirements toward problem-framing, communication, or critical-thinking assessments in tech and knowledge-work job postings would signal that the labour market is internalizing the AI-era education thesis faster than universities are.
AI error rate in high-stakes professional domains. If AI-generated outputs in medicine, law, or engineering produce a cluster of high-profile failures attributable to inadequate human supervision, expect rapid regulatory and institutional overcorrection — including calls for stricter domain-expertise requirements that push back against the Huang thesis.
Curriculum reform announcements from leading research universities. Watch for whether top-tier institutions begin restructuring core requirements around AI literacy and evaluative judgment — or whether they simply add an “AI tools” elective and call it a day. The depth of the response will indicate whether institutional leadership is genuinely reckoning with the inflection point or performing it.
Policy moves on AI in credentialing. Regulatory bodies in medicine, law, and engineering that govern professional licensing will eventually need to address what competencies AI changes in their fields. The first major licensing body to revise its exam structure in direct response to AI capability will mark a meaningful moment in this shift.
