7 min read
M04

Machines · No. 04

The Last Skill

When machines learn anything faster than you, the question shifts from what you know to what you can be trusted with.

· · 7 min · 1,650 words

In short

  • When answers become cheap, the value moves to the question and to the decision about which answer to act on.
  • Judgment is built by making real calls and living with the results, which is why it cannot be downloaded.
  • Machines can produce work but cannot be held accountable for it, so owning outcomes becomes a skill in itself.
  • A skill is something you can do; leverage is how many times your decision gets multiplied, and AI is the cheapest leverage in history.

In 1997, Garry Kasparov lost a chess match to a computer, and a lot of people assumed that was the end of the story. Humans had been beaten at the game we used as shorthand for intelligence. Pack it up.

What happened next is more interesting. Kasparov started promoting a new kind of chess where people could play with a computer at their side. In an open tournament of this "freestyle" chess in 2005, the winners were not grandmasters and not the teams with the strongest hardware. They were two amateurs running a few ordinary computers, who were unusually good at one thing: knowing when to trust which machine, what to ask it, and when to overrule it.

Kasparov later wrote that a weak human with a machine and a better process could beat a strong human with a machine and a worse one. Sit with that for a second, because it is the most useful sentence about careers in the age of AI that I know.

The machine was not the edge. Everyone had the machine. The edge was the person deciding what to do with it.

The skill that keeps getting cheaper

Every generation has a skill that the next technology turns from rare to free.

Scribes spent years learning to copy text by hand, and then the printing press arrived. Clerks were prized for fast, accurate arithmetic, and then the calculator fit in your pocket. For most of the last thirty years, knowing where to find information was a career advantage, and then search engines made everyone a librarian.

Now the thing getting cheaper is something we thought was safe: producing competent work. A first draft, a working function, a summary of a long report, a decent logo, a passable legal letter. Tasks that took a trained person hours can now be done in seconds, at a quality that is often good enough.

If your value is the production of competent first drafts, you are standing on ice that is melting under you. That is uncomfortable to hear at twenty, when most of school was designed to train exactly that.

But notice what did not get cheaper when the calculator arrived. Knowing which numbers to calculate. Noticing when the answer looks wrong. Deciding what to do once you have it. The calculator removed the arithmetic and left the thinking.

AI is doing the same thing, just to a much larger slice of work.

Asking is the new answering

When answers are scarce, the person with answers wins. When answers are abundant, the person with the right question wins.

This sounds like a fortune cookie, so make it concrete. Give two people the same AI model and the same afternoon. One asks it to "write a marketing plan for my app." The other asks: "My app has strong day-one retention but almost nobody comes back after a week. Here are the numbers. What are three explanations, and what is the cheapest experiment to tell them apart?"

Same tool. Completely different afternoon.

The second person did most of the work before typing. They knew what problem actually mattered, they knew what evidence they had, and they knew what kind of answer would be useful. The machine filled in the rest.

A good question is a decision about what matters, written down before anyone answers it.

This is why the best users of AI are rarely the most technical. They are the ones who understand a domain well enough to frame it. A doctor who knows which symptom is odd. An engineer who knows which part of the system is fragile. A founder who knows which customer complaint is actually about something else.

You cannot frame a problem you do not understand. So the strange result is that deep knowledge of a field matters more now, not less. It just stops being the output and becomes the input.

Judgment is a muscle, not a file

If questions are the input, judgment is everything after. It is the act of looking at five plausible options and picking one, with incomplete information, under some time pressure, knowing you might be wrong.

Machines are getting very good at generating options. They are much less useful at the moment of choosing, because choosing depends on things that are rarely written down: what your customers will tolerate, what your team can actually execute, which risk you can survive and which one would end you.

Judgment has an annoying property. You cannot read your way into it. You build it the way you build a sense of balance, by doing the thing, falling, and adjusting. Every real decision you make and then live with adds a small weight to an internal scale. Over years, that scale gets accurate.

This is why I think the worst career advice right now is "let the AI do it so you can skip the boring parts." Some boring parts are where judgment gets formed. A junior engineer who has never debugged anything by hand will struggle to tell when the generated code is subtly broken. A writer who has never wrestled a messy paragraph into shape will not recognise when a fluent one says nothing.

Use the tools. Absolutely. But protect a few places where you still do the work yourself, slowly, so that you have the instincts to check the work you did not do.

Someone has to sign

There is a line from an old IBM training document that has been passed around a lot in the last few years. Roughly: a computer can never be held accountable, so a computer should never make a management decision.

You can argue about the second half. The first half is simply true, and it is becoming one of the most valuable facts in the economy.

When a bridge is built, an engineer signs the drawings. When a patient is treated, a doctor's name is on the chart. When a company ships a product that breaks, someone answers for it. The software that helped them did not sign anything.

As machines produce more of the work, the person who takes responsibility for the result becomes the scarce part. That is a real skill, and it is underrated because it does not feel like a skill. It looks like this:

  • Checking the output when it would be easier to trust it.
  • Saying "this is mine" when it goes wrong, and fixing it.
  • Following a project past the point where it is fun, to the point where it is actually done.
  • Knowing enough about every part of the work to be held responsible for all of it.

Companies will pay a lot for people who close loops. They always have. The difference now is that the loops are bigger, because one person with good tools can be responsible for what used to take a department.

Skills versus leverage

Here is the distinction that ties all of this together.

A skill is something you can do. Leverage is how many times the results of your decisions get multiplied. A carpenter has skill. A carpenter who designs a chair that a factory makes a hundred thousand times has leverage.

Naval Ravikant has a well-known framing of this: the old forms of leverage were labour and capital, other people's time and other people's money, and both required someone's permission. The new forms are code and media, which copy themselves at almost no cost and need nobody's permission.

AI is the newest and strangest member of that family. It is a kind of labour that you do not have to hire, manage or pay much for. A single person with good judgment can now direct the equivalent of a small team of tireless, fast, occasionally wrong assistants.

That changes the maths of learning. In a world without leverage, the best investment was usually getting better at a skill, because your output was limited by your own hands. In a world with abundant leverage, small improvements in judgment get multiplied enormously, while small improvements in raw production get swallowed by machines that produce faster.

So if you are deciding what to learn this year, a useful test is: will this make my decisions better, or will it make my output faster? The second kind is still useful. The first kind compounds.

What I would actually learn

If I had to compress this into a practical list, it would look something like this.

  1. Learn one field deeply. Not to compete with the machine on facts, but so you can frame problems in it and smell when an answer is wrong.
  2. Learn to write clearly. Writing is thinking made visible, and a clear prompt, a clear plan and a clear memo are the same skill.
  3. Make real decisions early and often. Start things. Run small projects where the outcome is yours. Judgment needs reps with consequences.
  4. Learn how the tools work underneath. You do not need to train models, but knowing roughly why they fail makes you much better at using them.
  5. Practise finishing. The world is full of people with drafts. It is short of people with shipped work.

None of these will ever be "done". That is part of the point. They are skills that get more valuable the more capable the machines become, not less.

The person in the chair

Go back to that freestyle chess tournament. The grandmasters in that room had spent their lives learning chess. They knew more than the amateurs, by an enormous margin. And they lost, because the game had quietly changed from knowing the best move to managing the process that finds it.

Something like that is happening across almost every kind of knowledge work now. The machines will keep getting stronger. The amount any one person needs to memorise will keep falling.

But someone still has to sit in the chair, decide what the game is, choose which move to play, and put their name on the result.

That is the last skill. Make sure it is yours.

Further reading

  • Deep Thinking — Garry Kasparov
  • The Almanack of Naval Ravikant — Eric Jorgenson
  • Range — David Epstein

More from the shelf

M05 · Machines · 6 minWhat Are Humans For?In 1930 Keynes promised his grandchildren a fifteen-hour week. The real question he left them was what to do with the rest. M06 · Machines · 7 minThe Wrong QuestionTuring called the question of whether machines can think too meaningless to discuss. We have argued about it ever since. B04 · Builders · 6 minRunning to Stay in PlaceIn a world where everyone is improving, getting better is the price of standing still. Winning means changing the race.

See the whole library →