Kostas Karolemeas
Writing archive

· AI in education

Jensen Huang Is Confusing Doing the Work with Learning from It

AI can make a task unnecessary without making the learning it provides unnecessary. Jensen Huang's remarks about mathematics expose a distinction that matters in classrooms and engineering teams alike.

AI in educationhuman judgmentsoftware engineeringlearning
Dark teal and copper planes overlap beneath a rising translucent aqua brushstroke and a broad coral plane on cream.

Jensen Huang's remarks about children forgetting mathematics raise a question that reaches well beyond school: when a machine can do the work, what should a person still learn by doing it?

In his September 23 conversation with Ezra Klein, Huang referred to long division, multiplication tables and square roots, then asked: “Does it matter?” When Klein returned the question, Huang answered: “I don't think it does.” He acknowledged that other skills would matter and suggested that new ones would emerge. The exchange begins around 22:52 in the episode transcript.

That is narrower than saying children should learn no mathematics. A fair reading is that Huang thinks technology changes which capabilities deserve our effort. I agree with that starting point.

My objection is to the step that follows. A task can lose its economic value while retaining its educational value. Producing the answer and developing the person are different outcomes, even when the same exercise once delivered both.

Delegating work after acquiring a mental model is different from delegating the work through which that mental model would have developed.

The parallel with Musk has a limit

There is a familiar precedent. In Mark Harris's June 2018 reporting on Ad Astra, the school's co-founder Joshua Dahn attributed its omission of foreign languages to Elon Musk's expectation of immediate computer translation. This was Dahn's account of Musk's reasoning, rather than a direct public quotation from Musk.

The parallel is the assumption that automating an output removes the reason to acquire the underlying skill. Translation supplies communication across languages; computation supplies numerical answers.

But the analogy cannot decide a curriculum. I would still value learning another language for the ability to participate directly in another linguistic world: to follow a joke, choose a register, or notice that a translation has changed the relationship between two people. Translation can be enormously useful while leaving those reasons to learn intact.

Mathematics presents the same question in a different form. What do we want the learner to become capable of, beyond obtaining today's answer?

Understanding needs something to grow from

The tempting compromise is to keep conceptual understanding and discard procedural fluency. It sounds progressive: let machines calculate while people reason.

The difficulty is that these capabilities develop together. The National Research Council's Adding It Up, published in 2001, describes mathematical proficiency as interconnected strands. Understanding supports procedures; using procedures can deepen understanding. The report also distinguishes useful computational fluency from drilling speed at large calculations by hand. It offers no reason to preserve every exercise unchanged.

That is a more demanding position than either defending the old curriculum or declaring its mechanical parts obsolete. We need to identify which forms of practice build the understanding we want, then test whether a replacement builds it as well.

Consider a simple example. A price rises by 20%, then falls by 20%. It finishes below its starting point: 100 becomes 120, then 96. A calculator can supply those numbers. The learner still needs to recognize that the second percentage applies to a different base.

Now change the story. An AI service becomes 20% cheaper per request, but usage rises by 20%. Under those assumptions, total spending becomes 96% of the original amount. The arithmetic is elementary. Recognizing the relationship is the useful capability.

I want students to estimate the direction, explain the changing base, and recognize the same structure in a different setting. Some calculation helps make that structure visible. The appropriate amount and form of practice should follow the learning goal.

Better homework can conceal weaker learning

We already have evidence that assisted performance and acquired capability can diverge.

In a randomized field experiment published in PNAS in June 2025, Hamsa Bastani and colleagues studied nearly a thousand high-school mathematics students in Turkey. Students used either a general GPT-4 interface, a tutor with teacher-informed safeguards, or no AI.

The general interface improved practice grades by 48% relative to the control group. When assistance was removed, that group's exam grades were 17% lower. These are relative changes, not percentage-point differences. The safeguarded tutor largely removed the negative exam effect, without establishing a positive exam effect over the control group. It supplied guidance designed to support the student's own attempts rather than simply hand over solutions.

The scope matters: one school, particular GPT-4 tools, and short-term assessments. This is evidence about those designs, not a verdict on every AI learning product. Paper and methods.

A newer preprint by Sina Rismanchian and colleagues, first posted in May 2026 and revised in July, examines millions of mathematics-learning interactions on ALEKS. Its quasi-experimental analysis reports faster completion on tasks more susceptible to AI assistance alongside weaker performance on proctored retention questions. It infers effects from differences across tasks and conditions; it does not randomly assign each student's ordinary AI use. I read it as additional reason to examine retention separately from completion, with the caution appropriate to that design and a preprint.

Neither study supplies a complete curriculum. Both make a simple success metric inadequate: the work got done faster and the submitted answer improved.

The same question belongs in engineering teams

For a team adopting coding agents, I would ask the same question about the engineer as about the student: what capability remains after the assisted task?

An experienced engineer may delegate an implementation while retaining a clear account of its assumptions, failure modes and acceptable tradeoffs. A beginner can produce a similar-looking artifact without acquiring that account. This is an illustrative distinction, not a claim that every experienced engineer uses AI well or every beginner uses it badly.

The management risk is to observe the artifacts and assume the learning paths are equivalent.

In my article on technical understanding and clear expression, I used the example of an order submission whose response is lost after the order has already been created. An agent can implement a retry. The engineer still needs to explain whether that retry can create a second order, and what evidence would settle the question.

I would make that explanation part of training. Let the engineer use the agent, then change the requirement or introduce a failure. Ask them to predict the behavior, identify the relevant mechanism and check the result. If they cannot do that, the completed feature has told us less about their development than we hoped.

This is a proposal for protecting learning inside productive work. It does not require withholding useful tools until someone has recreated an entire profession unaided. It requires making the learning outcome explicit enough to observe.

Teach delegation and independence together

My preferred response would combine deliberate independent practice with deliberate AI use.

Before assistance, ask for an estimate, an initial approach or a description of what is confusing. During assistance, require the learner to compare explanations, make a choice and justify it. Afterwards, change the numbers or the context and check whether the learner can transfer the idea without being led through it again.

Those activities should be adapted by teachers to the learner and subject. They are a design proposal, not a validated universal teaching protocol. Their purpose is to expose the capability that a polished answer can hide.

AI could make this kind of learning easier to provide: more opportunities to ask questions, request another explanation and practice a specific weakness. The safeguarded-tutor result gives a concrete reason to take design seriously. It gives no license to assume that any conversational interface automatically teaches.

Huang is right to expect the skills we need to change. My concern is with treating the arrival of replacement skills as sufficient reassurance about the ones we stop developing.

Before we remove a learning activity, we should be able to say what capability it built, why that capability is still needed or no longer needed, and how we will know the replacement works.

The educational promise I want from AI is that more people become capable of understanding and acting in the world. A machine completing their assignments is too small an ambition.

Sources