NVIDIA's Jensen Huang: Stop Typing More, Start Deciding What to Type

Something strange happens when production becomes free: people who produced little start producing the same amount, with less verification per unit. The machine does not create wisdom by lowering the cost of words; it creates noise, because the marginal word, which used to cost effort and therefore passed a filter, now costs nothing and therefore filters nothing. Jensen Huang's July remark, that coding is essentially typing and typing is becoming a cheap commodity, was an economic statement, and the consequence is already here: when the supply of text is infinite and free, the scarce input is not text but the judgment about which text deserves to exist.

Every inbox and review queue shows the trap already set. More drafts, more tickets, more summarized threads, each one plausible, each one requiring exactly the same human attention as the version that mattered. Output grew faster than the attention that gates it, and the shifted ratio changes the experience of work: writing got easier, reading got harder, and teams that measure themselves by output can no longer tell whether they are moving faster or merely printing more copies of the same distance. Volume was a decent proxy for progress while volume was capped by effort. It is a broken metric now that the cap is gone. The person who still calls output progress is measuring input, and the machine owns that metric now.

The trap has a second, subtler set of jaws: confidence inflation. A person who types slowly has heard their own reasoning twice; a person who only sends the machine's version has not. The same forces that made production free made the production indistinguishable from thought, so the queue fills with assertions, and asserting is cheaper than verifying forever. Teams that notice the trap do not produce less; they gate differently, and they spend a much smaller share of the day on deciding what to type at all, which is the argument this entire article is testing. The testing is cheap: one week of logging, and no tools beyond the ones you already have.

So the reframe is deliberate rather than contrarian: the goal is not to type less for its own sake, and it is certainly not to type less thoughtfully. The goal is to pay attention to the minute before the words, the minute where the choice happens, because that is the only minute the machine cannot perform for you. Stop typing more is a compass, not a rule; its true heading is start deciding what deserves to be typed, and the rest of this article is about how to find that heading in the middle of a normal, noisy workday. The compass has three questions and zero settings, which is why it survives the noise instead of becoming more of it.

Most typing is not thinking in progress; it is rehearsal. By the time the fingers start, the decision usually exists, sometimes for days, and the keystrokes are not arriving at an answer, they are restaging the moment the answer arrived. Watch it happen and it is almost funny: the second draft of a paragraph is identical in content to the first because nothing between the two was a decision, only a repetition. This matters because it changes where the scarce minutes go. The minute you decide is the minute worth protecting; the minutes you type the decision are, in a market where the machine types for free, minutes spent on the commodity half.

Watch it happen in real time, on any machine. The email you rewrite four times says the same thing four times, and what changed between versions was comfort, not content. The ticket that spends a full day in typing never gains a new piece of information; it gains a more detailed restatement of a decision that was already made at the whiteboard. And the meeting about the document is usually longer than the document, because the deciding happened in the meeting and the document was merely the recording of it. The typing was never the hard part, and saying that is not an insult to typing; it is an audit of where the work actually was. And like every good audit, it ends with a recommendation, which is the routine in the next section.

The audit is uncomfortable the first time because it indicts the most respectable part of the office: effort. A day full of typing feels industrious, and a day with two decisions in it feels thin, so the rehearsal habit is reinforced not by laziness but by the theater of productivity. That theater is now, mercifully, measurable. A keystroke log, a document version history, or a simple timebox shows you the ratio, and the ratio is the diagnosis: how much of the session rehearsed, how much recorded, how much actually decided. None of the three needs to be defended, but only one of them is yours alone to do.

The rehearsal share is also the best trade candidate in the whole workflow. Every minute of restatement can be handed to a machine that will restate more fluently in seconds, and every minute of deciding cannot be handed to anything. When those two facts are finally put in the same sentence, the sentence becomes the study plan for the rest of this article, which is why the next section is not about writing better but about the three questions that decide whether the session should produce anything at all. Or rather, the three questions that decide whether the session should produce, which is a different and much more useful question than how to produce it faster.

Run three questions before you produce anything, and the shape of the session changes immediately. Question one: what decision does this artifact allow someone to make? If the answer is none, the artifact is decoration, and decoration at any speed is still decoration. Question two: what did I decide in the last ten minutes that changed what I am about to write, and if nothing changed, why am I typing instead of sending? Question three: if the machine produced this in thirty seconds, is the typing the work or the wrapper around the work? Three questions, one minute, and they act as a filter rather than a ceremony.

Apply them to an ordinary morning and the audit is nearly absurd. Inbox open, forty minutes of typing, and the split: the reply that merely confirms a date, no decision, eight seconds of speech; the draft that restates the decision from the standup, no decision, one line of context and the machine does the rest; the single email where a real trade-off finally gets settled, and that is the ten minutes worth keeping. Subtract the rehearsal from the session and the day gains time and quality together, because the same minutes land on the part that builds value instead of the part that records it.

The routine earns its keep through a side effect nobody mentions: it makes omission loud. During week one, question two is answered with nothing changed almost every time, which is precisely the confession needed, because a session with no new decision is a session that should not have started. Week two, the questions start causing meetings to be shortened and emails to be sent unpolished, both of which look like losses and are gains. Week three, the routine folds into the moment before the typing cursor blinks, three seconds, no ceremony, and the quality of the day rests on the three seconds that used to not exist.

Two guardrails so the routine does not become a guilt machine. First, rehearsal is not a crime: sometimes the point of a session is to hear the shape of a thought, and hearing it is legitimate work, so the questions are there to make that explicit rather than to forbid it. Second, the routine does not replace the decision, it locates it: the point of the questions is to find the minute where the real work lives so you can spend the real hours there. Guardrails in place, the routine becomes the cheapest productivity tool in the stack, and the next section shows what the choosers who use it do.

Great typers, in the old economy, optimized the downstream: faster input, faster formatting, faster turnaround. Great choosers optimize the other end, and their observable habits are almost embarrassingly plain, which is exactly why they are learnable. They say no in a minute. They kill options by criteria rather than by mood. And they keep a short written list of what they are not doing this week, because not choosing is also a choice, and an unrecorded one is a choice made by the default. None of these habits is talent; all of them are pipelines, and pipelines can be rebuilt.

They also decide against their own confidence: choosing is not consensus-free, but a chooser knows the difference between the decision and the reasoning, and can restate the decision in one sentence a critic could attack. That restatement is what separates a chooser from a decider who is merely loud. When the machine can produce the execution, the person who can defend the choice owns the work, and the person who only produced the words owns a byline and a revision history. The typing market now pays for the defense, which is a new sentence in the career math, and a more useful one than any typing test ever was.

Watch a chooser in a review meeting and the pattern is one move, repeated: they do not praise the draft, they replace the criterion. Same in the chooser's queue hygiene: decisions get a deadline and a named owner, because a decision with no owner and no date is a wish. And above everything, a chooser rarely asks for more time by typing more; they ask for more time by cutting scope, the most expensive-looking and cheapest move available, and the one the machine cannot make for them, because the machine's default is always to add scope. The choosers will not be automated away; the typers already have been.

The final habit is the quietest and the most transferable: choosers schedule the choosing. A decision arrives, and the reflex is to schedule it next to the typing, at the end of the day, when attention is spent. The chooser inverts the day, putting the ten decision-minutes first, because they have internalized the premise this article has been circling: the decision is the work, and the typing is the billable translation of it. Schedule the translation anywhere; protect the choosing. That single inversion is the whole difference between the people who feel the AI era as a windfall and the people who feel it as a race.

The chokepoint between deciding and committing is where most decisions quietly die, and the chokepoint is still a keyboard. A decision arrives in a meeting, complete with its reasoning, its caveat, and its confidence, and then it is transferred into a document by one person later, alone, with the nuance already gone, or worse, transferred into no document at all and reconstructed from memory next week. The decision and its capture happen at different speeds, and the mismatch is the loss. The fix is structural rather than disciplinary: capture the decision at the speed the decision occurs. Capture, not polish: the polish is later work and the machine is happy to do it, but no machine can recover a decision once the meeting has moved on.

That is the case for speech as the capture surface. Speaking is the only input that matches thinking speed, and the version you speak is the version that contains the constraint you would have edited out while typing, because editing while typing is exactly where the caveat is lost. The historical objection, that speech produces audio rather than artifacts, is precisely the problem expression intelligence solves: the layer understands the intent, removes the fillers and missteps that carried no meaning, and returns text you would actually send, so the spoken decision becomes the committed artifact rather than a recording of one. Say the constraint before you say the plan; the constraint is what survives the polish.

Voiskey runs this at the system level on Mac and Windows and supports more than 100 languages, so wherever the decision happens, in a meeting room or across a table in another language, the capture surface is the same. It stays private by design: audio is handled locally by default, with zero server retention and SOC 2 Type II attestation, which matters here because the half-formed version of a real decision is exactly the artifact you do not want leaving the room before it is finished. The commit can wait for polish; the capture should not wait for the commit.

Close with the habit, because a habit is the only deliverable that survives the article. The next time a decision forms, say it out loud in one sentence, the same sentence you would write an hour later, and then commit it the same hour. That single swap, speak, polish, type, edit, keeps the deciding and the typing from pretending to be the same job. Jensen Huang said typing is a cheap commodity; the unstated corollary is richer: choosing is not, and the chooser who captures decisions at the speed they happen has stopped paying commodity prices for the minute that matters.