Jensen Huang on AI: What Sets You Apart Is Not Keys, But Clarity

For twenty years a knowledge worker's productivity was measured in keystrokes per minute. Recruiters tested typing speed, tools optimized autocomplete, and teams judged meetings by how fast the notes appeared. Jensen Huang's remark cuts against that habit: coding is essentially typing, and typing is becoming a cheap commodity, while what has value is what you can ask. In an agent world the machine happily produces 10,000 words for any prompt, so the bottleneck moves to the other end of the pipeline: the quality of the request. Typing speed is now table stakes. The real bar is whether you can state what you want precisely enough that the first output is the last iteration you need.

Consider what a modern assistant does with a vague ask. 'Improve this report' returns a report-shaped document full of confident generalities. 'Rewrite the opening so it leads with the revenue figure the CFO asked about, keep it under 120 words, and cut every adjective' returns something usable. The difference is not vocabulary, it is specification. When generation costs almost nothing, the cost of a miss is the review loop: read, reject, restate, retry. Three rounds of that cost more than the typing you saved. Clarity is the new efficiency metric because it multiplies through every task that follows, and it compounds in the other direction when the ask itself is sharp.

The hard truth is that clarity is a different muscle from speed. Most people type fast and think messily; keystrokes export whatever the brain streams, which is why polished writers rewrite by hand. Voice-first work exposes the muscle honestly. When you dictate, the hesitation, half-sentences, and backtracking that typing quietly filters out appear in plain sight. You see how often you know the shape of the argument but not the target reader. You discover how rarely you name the constraint that actually decides the answer. That visibility is the raw material for the fix, and learning to state a clear ask is precisely what makes voice output useful rather than merely fast.

The new performance bar has a simple test. Could someone who has never met you produce exactly the thing you want from your written request alone, with no follow-up? Fast typing lets you re-explain privately; clear intent makes the first explanation the last. This is the standard Jensen's comment implies, but you can measure it without his permission. On any given day, count the round trips: the restatements, the clarifications, the apologies for a vague instruction. The number of round trips is the successor to your words-per-minute figure, and it goes down when the ask gets precise instead of quick.

The phrase 'improve this' has quietly become one of the most expensive strings in the workplace. It looks efficient: four words, no friction, and then the afternoon evaporates. The first model pass returns a plausible rewrite that misses the point. The second pass corrects the direction. The third pass fixes the scope you never mentioned. By the end you have spent more time steering than you would have spent writing the whole thing from scratch. The keyboard was never the expense; the ambiguity was. With agents the same dynamic runs faster and more confidently, because the machine never hesitates, never asks for clarification, and never flags that the brief contains nothing to work with.

The mechanism deserves a name because it explains why the waste compounds. A vague request contains no rejection criteria, so the first output cannot be wrong outright; it can only be wrong in a direction you did not intend, with no way to say why. Every iteration is a fresh guess instead of a correction. In an economy where execution is free, the cost of guessing moves entirely to the human side: checking, re-explaining, and re-queuing. Engineers have known this from bug reports for decades; the people who file those reports are learning it now that they use AI assistants for their own drafting. A good bug report is a spec. A good prompt is a bug report for a deliverable.

The price is not just time, it is credibility and atrophy. Each 'improve this' that comes back off-target teaches the people watching that your briefs are not decisive, and it trains the model to hedge. There is also an invisible tax: the more you lean on an agent that guesses for you, the less you practice stating what you want, and specifying atrophies at exactly the moment it is becoming your most valuable asset. That is the trap in Jensen's remark. The cheap commodity is the typing, which you are outsourcing. The expensive asset, knowing what you want, is the very thing vagueness stops you from exercising.

The fix does not require eloquence; it requires pre-work. Before you open the tool, answer three questions in one sentence each: who is this deliverable for, what decision does it support, and what must it not contain? Three answers, thirty seconds, and most 'improve this' requests turn into three-sentence briefs. If you cannot answer, the problem is not the tool but your understanding, and no prompt engineering will rescue it. This is why the advice to say it out loud first works: forced articulation surfaces the missing pieces. The vague request burns the afternoon because it defers the thinking; the clear request does the thinking before anything generates.

Treat a request as a spec and the whole conversation changes. A spec names the audience, the CFO rather than the engineering team, the constraints, under 150 words, no jargon, lead with risk, and the acceptance criteria, the summary states a go or no-go. A vibe is the opposite: a felt sense of rightness you cannot articulate and the machine cannot share. Vague requests are a vibe in spec's clothing. They feel clear to the author because the author has the context in their head. You can feel the difference the moment someone asks you to review their work: a spec makes the review quick, a vibe makes it endless. The failure only shows up on the other side, where the context is missing.

The audience is where most spec failures happen. People write for the reader they wish they had instead of the one they have: an engineer writes for the CEO, a salesperson writes for the product team, a support agent writes for compliance. The machine produces exactly what you ask for, so a request addressed to the wrong audience is a perfect output for the wrong reader. Naming the audience out loud, to a colleague or to a voice tool, is the cheapest fix in the entire discipline. Once the reader is fixed, word choice, depth, and evidence line up almost automatically, because most tone problems are really targeting problems.

Constraints are what actually steer a generation. 'Make it better' has no answer because better is unbounded. 'Cut the opening and keep the numbers' has a boundary. No more than three examples, do not mention competitors, end with the recommendation first: each one deletes an entire region of possible outputs. Constraints are also where the human earns their wage, because they encode judgment, the decision about what this specific situation requires, the thing the machine cannot infer from a shared corpus. When Jensen says the scarce resource is knowing what you want, he is describing constraint production. The people who keep their pricing power will be the ones who enumerate boundary conditions instead of describing moods.

Acceptance criteria turn vague feedback into votes. Instead of 'I like it more now,' criteria say: passes if it names the three findings, the recommendation, and the open risk; fails if it exceeds 200 words or uses phrases such as 'leverage.' With criteria, review becomes a checklist instead of a taste test, and the model can self-check as it goes. The habit works in speech too. Finish your dictation by stating what success looks like, then hold the output against it. If no one can check the result, nobody can verify that it is good, and you are back to the vibe economy, where the only arbiter is your mood.

Speaking has a bad reputation in knowledge work because of the static: you know, um, like, sort of, I mean, basically, plus the false starts where a sentence dies mid-air. People who say they think better when they talk are usually told to write it down instead, as if the stumbles were the problem. But the stumbles are normal speech, present in every executive and every domain expert, and they are not signs of thinking failures. They are the visible residue of thinking in progress. What looks like noise is usually the shape of the reasoning, and the shape is worth keeping. The difficulty is only that the static leaks into anything you actually send.

The leak is most visible at the boundary between thinking and sending. In a meeting, listeners hear the correction and the restatement together and resolve the intent for free. In typed text the correction never appears, because the author deletes it and only the clean final thought survives. In a verbatim transcript both versions are preserved and the message carries the hesitation. That is why raw transcripts read as foggy even when every word is correct: the intent is entangled with the search that produced it, and a reader who does not know you cannot detangle it. The clarity you had in your head is not in the file.

This is exactly the layer expression intelligence targets: understand the intent underneath the stumbles and produce the thought the speaker was reaching for. The gain is not politeness. It is that the content arrives without the noise, which is what makes voice viable as a medium for real output rather than for rough notes. Think of it as removing the broadcast static so the signal reaches the receiver. The signal itself, the point you were making, is preserved and sharpened because the digressions no longer dilute it. Fillers are leaks, and clearing them is not decoration; it is the difference between sending a thought and sending the search for it.

The speaker-side discipline is small and learnable. Before you dictate, state the outcome in one clause. While you talk, do not apologize for the corrections, because the repair is the machine's job; you are thinking out loud, not performing. When you review, look for the sentence you repeated or the point you circled, because the circled point is the intent; promote it to the top or cut the path to it. Every 'you know' marks an assumption of shared context, and context assumptions are the true cost of vague communication in a workflow where the listener is an agent rather than a colleague.

A dedicated polish layer changes who does the cleaning. The standard voice flow today is you speak, the tool transcribes, you edit: meaning you still do the work and the transcript is a to-do list. The expression-intelligence flow inserts a step between intent and output. The system recognizes what you meant, drops the fillers and the retries, and hands back text that is ready to send. Your edits shrink from restructuring a mess to adjusting a word. At that point speaking becomes a delivery mechanism rather than a draft generator, and the draft you could never write on a deadline becomes the message you actually send.

The ordering matters: polish before the agent, not after. When you hand a rambling instruction to a coding agent or a research assistant, the agent happily makes a guess, and the guess is typically the average of every half-thought in your brief. Polish the request first, one audience, two constraints, three acceptance criteria, five lines, and the agent's output arrives near the target on the first pass. Jensen's framing of the machine as something you hire rather than use makes the same point from the other side: you hire an employee for the brief, not for the spelling, and you would not brief someone in guesswork.

The practical argument is reach. A dictation layer that installs at the system level on Mac and Windows can sit inside Gmail, Outlook, documents, a chat window, or the prompt box of a coding assistant, so the polish layer is not one more tab; it is the front door of every text field you own. The contract with it is honesty: you say what you mean, including the half-formed version, and it returns what you were trying to send. That exchange is worth far more than faster transcription, because sendability, not accuracy, is the actual test of a voice tool.

And then the ask lands: clear intent plus precise expression is what survives when typing itself is free. The workflow is speak, polish, type, edit, and the edit pass shrinks to taste, the part you actually enjoy, because the mechanicals are the machine's job. Jensen's claim, that what sets you apart is not the keys but the clarity, was never about being faster at a keyboard. It is about producing an input the machine can run with on the first try. The polish layer is the piece that turns the raw sound of your thinking into a precisely specified brief, and the brief is the entire game in an agentic workplace.