On Alignment

01 Jan 2022

Running anything — a life, a company, a country — is two problems: deciding what to do, and getting it done. For anything worth doing, both involve other people. A good decision needs honest information from the people it touches, and getting it done needs them to pull the same way. They do both only when they want what you want.

Early on, you make every decision yourself, and it works, because there are few decisions and you’re in every room where one gets made. Then there are more rooms than you can be in. Every call that needs your judgement waits until you arrive, and you become the bottleneck on your own company. The way out is for other people to make the calls you would have made, or better ones, and to carry them out without asking you first. That is alignment.

A group is aligned when its members want the same end and trust each other enough to argue hard about the means, then commit to whatever gets decided. Unanimity doesn’t qualify, because a team that agrees about everything is either lying or has stopped thinking. Obedience doesn’t qualify either, because it lasts exactly as long as you’re watching. The argument over means is the part you’re actually buying. A team that can’t argue about how to get somewhere has no way to discover that it’s going the wrong way, and the whole point of letting other people make the calls was to be wrong less often.

If that’s the definition, the ways to build it follow from it. A group can only want the same end if the people in it want it, and that comes down to whom you let in. Those people will disagree about how to get there, so they need enough trust between them that disagreeing is safe. Their disagreements need to converge rather than circle, so they need to be working from the same facts. And each person’s private interests need to keep pointing at the shared goal over time, because interests drift. That gives four levers — whom you pick, the relationships between them, what you teach them, and what you reward them for — and I think it’s the complete list. Anything else you can name turns out to be one of these under another name. Countries have a fifth in force; the rest of us make do with four. I’ll take them in the order a growing team comes to rely on them.

1. Selection. The cheapest alignment is the kind you never have to build, because you picked people who already wanted what you want. So hire for eagerness and competence over polish, and watch whether a candidate lights up at the actual problem or only at the job. Selection is also the least precise of the four levers. You learn what someone really wants over years, and you can only choose from the people who happen to show up. It sets the ceiling on how aligned a team can get, and the other three levers do the work underneath that ceiling. The first of them is trust.

2. Personal relationships. People accept a call they would have argued against when they trust the intent of the person making it. That trust is what makes disagreement safe, so this is the lever the definition rests on. It carries a team of ten almost by itself, and it gives out somewhere around Dunbar’s 150, past which nobody can know everyone. It’s built in the conversations that aren’t about work, and nobody has found a process that produces it, which is why you can’t manufacture more of it when the team outgrows it.

3. Training. Training gives people the same facts and the same model of how the business works, so that when they argue about means the argument goes somewhere. It changes what people believe. What they want was settled when you chose them. The Socratic version works best, which means asking questions rather than giving answers: you forget an answer you were handed and defend one you reached yourself. Clear, specific feedback is training too. Training carries a team of a hundred.

4. Incentives. Past a hundred people you’re relying on incentives whether you admit it or not, so admit it, and make wanting what you want the selfish choice too. Munger’s line — never, ever, think about something else when you should be thinking about the power of incentives — gets quoted as if incentives were the easy lever. Goodhart’s law says that when a measure becomes a target it stops being a good measure, and every incentive you can write down is a measure. Reward lines of code, and you get lines of code. That is why Munger tells you to think about nothing else.

   The rule I’ve landed on is to reward outcomes and nothing that stands in for them: equity in the company, or credit for a result a customer noticed. Any proxy you can name will be gamed, and the outcomes you actually want are slow and noisy to measure, so you have to accept the slowness. In exchange, incentives are the only lever that scales without limit, and the only one that keeps working when you’re not looking, provided you pointed them at the right thing.

All of this aligns people to something, and that something should exist on paper. People can argue with a document, which is the point; nobody can argue with a founder’s mood, so a company aligned to its founder is only complying. Thiel ran PayPal by refusing to talk to you about anything except your single top priority. That is the extreme version. The general claim is smaller: you can delegate a decision only when the person can test it against a goal you both hold. So state the goal, and state how far their authority runs — what they decide alone, and what needs a conversation first. Leave the scope unstated and both sides get trapped: they can’t tell what they’re allowed to decide, and you get decisions you never meant to hand over.

A written thesis also gives you Ries’s Build-Measure-Learn loop for free. You hypothesise, build, measure, and then change the thesis in the open when it turns out to be wrong, so that people re-align to the bet rather than to you.

Growth cuts both ways. Revenue growth hides misalignment, because when the pie is growing nobody fights over slices, and a lot of what people call culture is just growth. You find out which one you had when it stalls. Headcount growth creates misalignment, because every hire dilutes selection a little, and at some point you pass the size where relationships can carry the load, then the size where training can. If you haven’t built the next lever by then, you find out the hard way. So build the levers on purpose, before you need them.

The obvious objection is that this is all just selection: hire well, and the other three levers are noise. There’s something to it, and I would take a well-selected team with no process over the reverse. But selection is slow and imprecise, and it caps out at the people you can actually get. Worse, a team selected purely for wanting what you want is a monoculture, and a monoculture can’t do the arguing the definition depends on. It reaches agreement fast and discovers its mistakes late. The other levers exist to close the gap that selection leaves, and to keep the disagreement alive.

That leaves the arguing itself. The definition only works if the argument actually happens and then actually stops. Bezos’s phrase for this is disagree and commit: say your disagreement out loud, lose the argument if you lose it, and then execute as if it had been your idea. Under-communicate, and the disagreement never gets voiced; the team marches in step off a cliff. Over-communicate, and it never gets dropped; nothing ships. Most alignment failures are known to everyone and said by nobody, so the skill to build is being comfortable with confrontation and equally comfortable with losing. The team you want is aligned enough to move fast and disagreeable enough to stop, and you have to build both of those on purpose.

Related: Wanting, Cadence in business.