How Do We Think About Hiring in the AI Age | Fred Liang<br>For about a hundred years, growing a company's capability meant one thing: adding people. More to do, more hands to do it. Headcount was the proxy for ambition. A bigger team was, almost by definition, a stronger one.

I think that equation is now broken. I think most organizations haven't noticed yet.

The honest version of how we think about hiring is uncomfortable to say out loud: a hire is a liability until proven otherwise. Not the person. A person is never a liability. The act of adding one to a team is. By default it is a negative, because it taxes the one thing that actually limits a modern team. And that thing is no longer labor.

It's context bandwidth.

The constraint moved

For most of the history of work, the bottleneck was execution. There was always more that needed doing than there were people to do it, so the scarce resource was hands, and you bought capability by buying hands.

That is the part that broke. Execution is becoming abundant. The machine can now take a goal instead of a keystroke, carry it forward, test it, revise it, and try again. When execution gets cheap, the bottleneck moves somewhere else, and where it moves is into the team's shared context: the live, mutual understanding of what we are doing and why.

Call it context bandwidth. It is the amount of shared understanding a team can hold and exchange at once. In a small team, the daily rate of that exchange runs at a level that would have looked insane a few years ago, and it is the actual engine of the work. What you want is overlap of context, not control of people. Everyone roughly holding the same picture, updated constantly, so that decisions don't have to route through anyone.

Here is why this makes headcount a liability. Context bandwidth does not scale with headcount. It scales against it. Each person you add multiplies the surface that has to stay in sync. Coordination cost is superlinear, and shared context is the thing it eats first. So past a very small number, every additional person makes the team slower at the precise thing that made it fast.

Which means small is not a stage you grow out of. It is the design.

Execution is becoming free

It's worth being precise about why execution stopped being the scarce thing, because the wrong analogy leads to the wrong conclusion.

The spinning jenny multiplied a worker's hands. It made an existing skill go further. The water-frame was different. It substituted for the skill itself. AI is the water-frame for execution. It is not making skilled operators faster at operating. It is absorbing the skilled operation, the part we used to pay a premium for, and pushing the human up a level: from working the machine to designing and maintaining the line that does the work.

The goal, increasingly, is to get the human out of the loop on execution entirely. Not because people are the problem, but because a person standing in a loop adjusting one fluctuation at a time is no longer the right place to spend a person.

So you need fewer people, and far more cracked ones

The lazy reading of all this is that you need fewer skilled people. That is exactly backwards.

Anyone can run a line. Almost no one can design one. The work that's left is the hardest work there is, and only the most capable people can do it: specifying what should exist, building the pipeline, encoding judgment into it so the system produces the right thing and not merely a working thing. The need for ordinary execution collapses. The need for genuinely cracked people goes up.

And here is the part that should worry everyone, because it is a trap that closes slowly. The way you used to mint a cracked person was by having them do the execution. You ground through the manual work, and somewhere in those years you developed taste, intuition, and the ability to tell a good system from a bad one. That apprenticeship is exactly what's being automated away. We are removing the training ground that produced the only people who can do the new work.

So the supply of cracked people does not just stay flat. It falls off a cliff. Scarcity is migrating away from hands, which are now abundant, toward the small number of people who can design the machine. Those people are about to become the rarest resource in the economy.

The question we actually ask

Once you see hiring this way, the question you ask about a candidate changes completely.

The old question was: what can this person execute? It assumed execution was the scarce good, so you were buying throughput.

The new question is: what will this person do to the team's context bandwidth? Will they raise the shared context or drain it?

This matters more now than it ever could before, because the variance has exploded. The gap between the best operator and the median used to be 2x, maybe 10x. Now it can be 100x. Or, just as easily, negative. There is a real archetype now of the person who ships enormous volume with no...