So What Is Valuable? Jensen Huang's Answer: The Way You Ask

Every few decades an input gets mass-produced, and the first casualty is always the price of the output that input made. Answers are the current case. Code, drafts, summaries, translations, boilerplate memos, all of it is now produced at a marginal cost close to zero, and the shape of Jensen Huang's July remark to Harrison Chase is economic rather than rhetorical: typing is a cheap commodity, and what is really valuable is not the answer but the questions you can ask. The second half of that sentence is the one worth printing, because it states where the market has moved the premium.

The mechanism matters because it predicts what happens next. When supply of an output becomes unbounded, the premium does not vanish, it migrates to the input that decides which output gets produced. Grain became cheap and the scarce thing became land. Answers become cheap and the scarce thing becomes the question, the thing that selects one answer out of an almost infinite supply. The person who can ask the question that picks the right answer owns a form of leverage that used to belong to whoever could produce the answer faster, and the leverage now compounds because every model in the world is waiting to serve the same ask.

That is uncomfortable news for people whose professional identity is built on leading with answers. The analyst, the consultant, the domain expert who has been paid for years for the answer itself has watched the answer get cheaper by the month, and watched a junior with a model produce a passable version at meeting speed. The part of their craft that survives is the part they rarely put in the deck: knowing which question the client is actually asking, what the frame of the decision really is, and which constraints would make the standard answer wrong in this particular case.

The inflation runs the other way. There is no market glut in well-formed questions, and there is no sign of one, because the question is upstream of anything the machine can do, and because a good question is the rarest object in the exchange: it is the only thing the other side of the desk cannot substitute. When the recommendation for the next six months of your own career reads like this, the practical implication is short. Start treating the way you ask as a craft with practice hours, examples, and a review loop, the way you would treat any other skill you intend to be paid for.

A question is never just a question. Every ask carries a frame, and the frame is the part that does the most deciding: why the question matters, who will act on the answer, and what decision the answer will feed. Two people can ask the same sentence and want nearly opposite work, because the frame differs, and a machine given the sentence without the frame answers the sentence exactly as written. The difference between the output you wanted and the output you got is almost always a difference of frame that was never stated, which is a fixable difference once you know to look for it.

Alongside the frame sit scope and constraints, and both are invisible until they are missing. Scope says what is inside the question and what is genuinely out, which is the difference between an answer you can use and the same answer with three wrong assumptions welded on. Constraints are the hard edges: the deadline, the audience, the privacy rule, the tone that cannot be struck, the fact that must be checked against the source, the format the next step requires. A trained asker names all three before the first exchange, because each one is cheaper to state than it is to rediscover in review.

Run this diagnostic on your last ask. Take the request you last sent to an AI, to a colleague, or to yourself, and write down its frame, its scope, and its constraints explicitly. Most people discover that two of the three were implicit, which means the other side of the exchange had to guess, and every guessed constraint is a latent risk. Not the risk that the answer is slightly off, but the risk that the answer is confidently wrong in exactly the way you did not specify, and that it looks right until the moment it costs something, usually at the worst possible point in the schedule.

The craft part is to make the implicit explicit on purpose, and the sentence form helps, so the habit is worth memorizing in a shape you can say in thirty seconds: I am asking because X, the answer will feed Y, inside this scope, excluding that, and it cannot violate the three constraints I just named. Said out loud, that sentence converts a question into a directive-quality artifact, and it takes less time than the first draft of the question took. The way you ask is the way you set the terms of the work, before the work begins, and the terms are the only part of the exchange the other side of the desk does not already have.

The prompt era created a folk hero: the one-shot asker, the person who composes a long, perfectly specified prompt and receives the ideal answer in the first round. The hero is mostly a myth, and the myth does actual damage because it teaches people to optimize the first message and to treat a mediocre first answer as the answer. The model's first output is a draft, and drafts are repaired, and the repair is where the craft lives. Any expert who has watched a newcomer get better results with the same tool has watched the follow-up skill at work without noticing it.

The craft of asking is mostly the follow-up, and the follow-up is invisible precisely because it is a conversation. The moves are small and worth naming: what is missing here; which of these claims would you check first; what would make this answer wrong; enumerate the three things I might be misreading; give me the reasoning, not just the result. Whoever is good at these moves appears, from outside, to have access to a better model, because they surface a better answer from the same one. The model is the same; the interrogation is not, and the interrogation is the part that is actually yours.

The cost of skipping the loop is measurable. An ask that needs three rounds to converge costs three times the one that converges in one, and the difference between a one-round exchange and a three-round exchange is rarely the model and almost always the follow-up, because the follow-up is where the missing frame gets supplied. In a world where the ask is the skill, the round-trip is the unit of the economy, and the invisible skill of the follow-up is what determines the average round-trip. It is also the skill nobody can automate for you, because it requires you to know what you actually wanted.

To train it, force the habit: never accept the first answer as final, and by accept, mean read it with the specific intention of finding what was not asked. That forcing move is the whole drill, and one round of it changes the quality of the work even when the follow-up ends up being mild, because closing the gap is the value. The aim is not skepticism for its own sake; it is the discipline of interrogating until the answer addresses the question you were trying to ask, which is the standard every expert applies to their own work and almost nobody applies to a machine's.

Commands and questions look alike in a prompt box, and they behave very differently. A command is the writer's output pre-decided: write a summary, do the analysis, draft the memo. The sentence is a delivery instruction with the thinking already done, and the model is being used as a secretary. A question is an admission that the thinking is not done: what are the three decision-relevant differences between these two documents, and where would this analysis be most likely to go wrong? The difference is not formality; it is whether you are depositing decisions or asking for them, and it is the difference between a tool that types for you and a partner that thinks with you.

The diagnostic has a sharp tell: does your ask anticipate the answer? If you have already decided what the output should look like, you do not need an AI, you need an editor, and you should produce the text yourself or use a tool that polishes what you said. If you genuinely do not know, and you want the reasoning alongside the result so you can judge it rather than accept it, you are asking, and the exchange is worth having. A prompt box is the only place in professional life where the two modes look identical, which is why so many people never notice which one they use.

Run the exercise on your last five asks. Rewrite each one as the question you would ask a smart colleague, keeping your intent but dropping the pre-decided output, and count how many of the five changed shape. A ratio like four out of five reveals that most of your prompt usage is transcription, not inquiry, which is exactly the economic risk of this era: you are hand-executing a machine that would answer a question at the same cost. The rewrite takes ten minutes, and the ratio is a number you will want to check again next month, because it is the honest measure of whether the shift has reached your desk.

The diagnostic is a snapshot, not a verdict. Nobody asks all the time, and a large fraction of requests are honestly commands, which is fine, and in fact is the correct shape for a lot of routine work. The goal of the exercise is awareness of the mix, because the mix is a choice about where your effort goes. If a week of asks reads like a week of typing, then the ask economy cannot deliver the benefit, because the benefit only appears where questions actually start: at the point where you do not yet know the answer and say so out loud, without pretending you already do.

Actual thinking arrives messy. It starts mid-sentence, restarts, digresses, and corrects itself as it goes, and the well-formed question is usually the version you produce after talking your way to it. That is why the highest-quality asks in an individual's repertoire are the ones they never submit: the mess was never captured, and by the time they sat down to type, the version had already expired, which is the single most common way good questions die in professional life. The gap between thought and ask is not a discipline problem; it is a capture problem, and capture problems have a fix.

Speech is the capture channel, and the capture half is solved. NVIDIA's Nemotron 3.5 Transcribe, open-sourced in late October, handles 40 languages, diarizes up to eight speakers, and ingests an hour of audio in a single pass, and the Riva library with NIM microservices runs the whole pipeline offline on one GPU. TechCrunch described the moment as voice AI in vogue, and the description fits. But capture is not comprehension. The transcript preserves the ums, the restarts, and the abandoned sentence exactly the way the speaker abandoned them, which means the mess is now faithful rather than forgotten, and faithfulness without understanding is just a better recording of a problem.

The missing layer is the understanding one, and this is where expression intelligence works: take the messy utterance, work out the intent that was actually behind it, remove the fillers and the missteps, and hand back a well-formed question, ready either to send or to feed an agent. Voiskey implements it as a system-level voice input for Mac and Windows, supports 100+ languages, keeps the audio local by default with zero server retention, and carries SOC 2 Type II attestation, so the channel you use for your most honest thinking is not also the channel that leaks it. The polishing is the point, not the typing.

The drill for the week is one small conversion. Tonight, capture a half-thought as a voice note, the kind you would normally discard: the objection you did not raise, the question the report left open. Tomorrow morning, turn that note into a well-formed question with a frame, a scope, and constraints, and direct it at a real task. You will see the distance between the raw thought and the finished ask, and then you will understand why the way you ask, rather than the way you type, is the thing worth practicing. The thought was always there; the capture and the polish are what make it work.