Nvidia's CEO shrugs at kids losing basic math. The research he was shown says the skills aren't sticking

Nvidia CEO Jensen Huang was presented in a New York Times interview with research out of China showing that AI tools help students finish homework faster, but that the skills the work was meant to build were not sticking, with performance on quizzes and tests degrading. Huang agreed the retention loss is probably happening to many kids. He also said he is not worried. "Try to get a kid to do long division right now," he said. "The multiplication table is starting to be forgotten. Doing square roots, my goodness. Basic math is being forgotten. Does it matter? I don't think it does." Pressed on which skills do matter, he suggested we will "discover new ones," and offered himself as a case study: he does not know his own home address, ZIP code or phone number. "I can live with it." The trade, he argued, is some loss of "finer intellectual dexterity" in exchange for becoming better "systems thinkers." Kotaku, which reported the exchange, was scathing, and noted that Huang's company sells the hardware most AI systems run on.

Our thesis: Huang is right about a category of skills and wrong about this one, and the difference is the whole argument. Offloading stored facts is ancient and harmless — nobody laments the death of memorized phone numbers, and his ZIP code anecdote is a fine example of a genuinely obsolete skill. But the research he was shown does not describe fact storage. It describes students who produced correct homework and then failed the test. That is not a person who outsourced a lookup. That is a person who never built the internal model the lookup was supposed to install.

That distinction is where our long-standing concern about evaluation lands. Output has stopped being a reliable signal of underlying capability — the same crack that makes AI leaderboards untrustworthy makes graded homework untrustworthy. Completed assignments now measure tool access, not learning. And arithmetic intuition is not trivia; it is the thing that tells you an answer is off by a factor of ten. Someone who cannot estimate cannot audit the machine. In a world where more decisions are drafted by AI, the ability to sense that a number is wrong becomes more valuable, not less.

The "systems thinker" promise is the weak link. It is an aspiration, not a demonstrated outcome. Nobody has yet shown a curriculum that reliably produces higher-order reasoning in students who skipped the mechanics, and the research on the table points the other way. Huang is entitled to his optimism, and his critics are entitled to note, as Kotaku does, that Nvidia's growth depends on AI adoption being unambiguously good. We would not call that disqualifying — plenty of interested parties happen to be right — but it is a reason to weigh the evidence over the anecdote.

Our read: education is one of the areas where AI could deliver the most, and we still believe it. A patient tutor that adapts to each student has been the field's unfulfilled promise for fifty years, and it is now technically within reach. The near-term reality is uglier, because the default use of these tools is not tutoring, it is homework laundering — and schools kept grading the artifact instead of the learning. The fix is not banning the tools; it is changing what gets measured, toward process, oral defense and in-person assessment. Huang's underlying instinct — that which skills matter will change — is correct and worth taking seriously. But "some skills will become obsolete" is not the same claim as "it is fine that a generation cannot check the machine's work," and the second one is the one his own cited research puts in doubt.