Jensen Huang Says Your Questions Are Worth More Than Your Keystrokes

Jensen Huang has made a two-part claim, and the second half deserves a place on your desk. In a July conversation with Harrison Chase, the founder of LangChain, shown on NVIDIA's official channel, he said that coding is essentially typing and that typing is becoming a cheap commodity. The follow-up is the part that matters here: what is really valuable is not the answer, but the questions you can ask. Read the two halves as one statement about supply. Answers are about to be produced in effectively unbounded quantity. Questions are not. That asymmetry is the entire argument, and once it is stated, the professional question becomes what to do about it.

The economic logic is simple and worth spelling out. When something valuable becomes cheaply mass-produced, the premium does not disappear; it moves to the complement, the thing that decides which unit of the cheap thing gets produced at all. A machine will produce a good answer, a mediocre answer, or a wrong answer at almost the same cost, and it will produce all three whether you ask well or poorly. The difference between an afternoon that ends in a usable deliverable and one that ends in a plausible-sounding mess is entirely upstream, in the framing of the ask. Your keystrokes are downstream of that. Your questions are upstream of it, and upstream is where the margin now sits.

His public record across the last year runs in the same direction. At the October Berlin GTC keynote he declared the end of the traditional PC and introduced RTX Spark as your personal agent, a digital companion you hire to work for you rather than software you use. At the same event he steered a robot duck with voice commands, making it lie down, roll over, and nap, a small demo with a pointed moral about who gives the orders now. And in the CMU graduation speech he offered the personal version of the argument: AI will not replace you, but someone who uses AI better than you may.

What does the answer-is-now-a-commodity framing change for an ordinary working day? It changes what deserves your attention. Write down this evening's three most important tasks and mark, next to each one, the fraction that is producing text and the fraction that is deciding what the text should say. For most knowledge workers the ratio is inverted relative to where the value now sits. Nothing about this is a verdict on your typing; it is a statement about the market the typing lives in. The scarce skill is interrogation, and interrogation, unlike finger speed, is something you can deliberately train, which makes this the rare performance claim that is also a plan.

Keystroke count was never printed on a review form, but it sat under every metric that was. Lines of code, tickets closed, documents produced, emails answered: all of it is a proxy for volume of text output, and volume of text output used to be the honest measure of work done. That made sense for exactly one reason. When the machine produced nothing, the fingers were the binding constraint of the whole workflow. A person who typed quickly could pull more value out of an hour than a person who typed slowly, so the economy priced the typing accordingly. The metric was not wrong; it was accurate for a world that no longer exists.

That constraint has now dissolved on both ends. AI can produce unbounded code, drafts, summaries, and transcripts. Even the listening side is no longer the bottleneck: NVIDIA open-sourced Nemotron 3.5 Transcribe in late October, with 40 languages, speaker diarization for up to 8 speakers, an hour of audio in a single pass, and the open Granary 2 dataset behind it. The raw capture of sound and text has become cheap enough to give away. Measuring a colleague by keystrokes today is like measuring a mill by how many times the wheel was turned by hand, a measure of effort that no longer scales with value.

So what does the new scorecard look like? Count asks, not keystrokes. A useful first metric is how many well-formed questions you produce in a day, and how many of them aim at work that another hour of typing would otherwise have consumed. A second metric is the size of the gap between your first ask and your accepted output: each round-trip you need to fix a vague request is rework, logged in the ask economy the same way rework was logged in the typing economy. A third is negative: the count of keystrokes you did not type because a machine executed instead.

Here is a forty-minute exercise that makes the shift concrete. Take your last three deliverables. For each one, estimate the hours spent typing it and the hours spent deciding, framing, and reviewing it, then label the document typed or directed. The ratio is a snapshot of exactly how much of your work still runs through the finger economy. The point is not to feel bad about the typed half; it is to notice that the directed half is where the claim you can actually defend lives, and that the typed half now competes with a machine that produces the same text for free.

The most common mistake in the prompt era is treating the AI like a vending machine: you deposit words, the machine dispenses output, and the quality question is supposedly about the wording of the deposit. That model makes a single best-effort prompt sound like the whole craft, which misplaces the craft entirely. Interrogation is the better frame. The first answer from any model is a draft answer, and the value is in what you do after you read it. That is the same move an experienced colleague makes: read, notice the gap, ask the narrower question. The follow-up is the unit of value, not the prompt.

A hygienic conversation needs three ingredients before the first exchange. First, the outcome: a sentence that describes what done looks like, not what the machine should write. Second, the constraints: what cannot be said, who the work is for, what format it has to land in, which details must survive. Third, the criteria: how you will judge the result when it arrives. Notice that all three are things you can say to a colleague in a meeting, and that is precisely the test. Commands to a tool hide these; questions to a person make them explicit. If your request would embarrass you as an instruction to a smart co-worker, it will not work better on a machine.

There is a hardware reason the hygiene has to happen before the machine sees the request. Voice models that transcribe well do exactly what transcription is supposed to do, which is preserve the mess. NVIDIA's Nemotron 3.5 Transcribe is the strongest recent example: it is a serious model, it runs an offline pipeline through the Riva speech library and NIM microservices on a single GPU, and it will still hand back a verbatim transcript full of ums, you knows, false starts, and restarts. NVIDIA built the ears and the throat. The layer between hearing and understanding what you actually want to say is still empty. That empty layer is where hygiene lives.

The habit that makes this practical is a tiny script, and it fits on a sticky note. State your intent in one sentence. Name the output. List the constraints. Then ask the machine what it is missing. Read the answer as a draft, not as a result. Then speak the next question out loud before you reach for the keyboard. The last step is the one most people skip, and it is the one that turns a prompt into a conversation. Keep the loop spoken and the round-trips shrink, which is the only outcome measure that matters in the ask economy.

A to-do list is a keystroke-era artifact. It is a stack of imperative commands written to yourself, and each command assumes you are the one who will execute it. Write the summary. Fix the error. Draft the note. In an economy where the machine executes, every item on that list is misaddressed: the value of the item was never the verb, it was the decision about what the verb was for. Before lunch today, rewrite the list. Every actionable item becomes a question the machine can answer, and you become the person who decided the question was worth asking in the first place.

The rewrite has a useful side effect: it separates your list into two honest piles. If an item will not fit into a question, it is probably a human decision, a conversation you need to have, or a judgment only you can make, and it should stay on your list in exactly that shape. If an item becomes a question easily, then its value has already moved from execution to choosing, and your fingers were never the right instrument for it. The two piles are a clearer weekly plan than any priority system you have used, because they sort by where the value actually sits.

A question is also a better handoff artifact than a command. Hand a command to a colleague and you pass a task; hand them a question and you pass the intent, the frame, and the standard in one package. That package survives you being out of the office for a week, because it does not depend on your memory of why the task existed in the first place. Writing questions is also more honest, because a question forces you to state what would count as answered, which is the acceptance criterion by another name, and criteria are the thing people skip when they are in a hurry.

Do the rewrite before lunch for a timing reason rather than a superstition. The morning is when your framing is sharpest, and framing is now the scarce resource, so the highest-value hour of the day should buy direction rather than transcription. Spend ten minutes converting the list, hand the machine the first question, and let the afternoon run on delegations instead of drafts. If the list contains a heavy item, split it into two questions, one about the destination and one about the obstacles, because a single complex question is simply a vague instruction wearing better clothing. By lunchtime you will have moved from being the author of tasks to being the shaper of asks, which is a much better position from which to spend the afternoon.

There is a reason the best questions in your day are the ones you never ask. They arrive during the walk between meetings, on the hang-up of a call, in the pause after a hard paragraph, and by the time you are seated in front of a keyboard they have been replaced by three more urgent typed tasks. Thinking moves faster than hands can follow, the same way the machine runs ahead of what you type. The value Jensen Huang says is parked in questions is lost to friction, not to an absence of thought. The thought was there; the capture channel was not.

The interrogation loop that prompt hygiene describes becomes much cheaper to run when the ask can leave your mouth at the speed it leaves your head. You read a draft, you think of the gap, you say the tighter question, and the tighter question is what goes back to the machine. That is why voice, not typing, is the natural vehicle of the ask era: it removes the standing delay between thought and inquiry, and it makes the follow-up as cheap as the original ask. The loop closes at the speed of the conversation instead of the speed of the keyboard.

But transcription alone is not inquiry. If you dictate a question and the tool returns a verbatim transcript, you have still handed the machine your ums and restarts, and your sharpest ask is buried under a paragraph of conversational noise. This is the shape of the problem on the desktop today: NVIDIA supplies the ears and the throat, and the layer in between is where Voiskey works, using expression intelligence to understand the real intent, strip the fillers and missteps, and deliver text that is ready to send. It runs as a system-level voice input on Mac and Windows with 100+ languages, keeps audio local by default with zero server retention, and holds SOC 2 Type II attestation.

The drill is small. Keep an ask notepad that accepts your voice, and capture every half-thought you would normally drop: the objection you did not voice in the meeting, the question the report left open, the idea that got interrupted by the next announcement. Once a day, review the capture and ask of each line: if this had become an agent prompt this morning, what would it have saved? You will find that a few percent of your half-thoughts are worth hours of someone else's execution. Retraining the default from typing to asking is exactly the kind of small, repeatable practice the new scarcity rewards, and it works because the capture channel finally matches the speed of the thought instead of the speed of the hand.