The few knobs you turn

In teaching machines there is a small distinction that I find illuminates a great deal about running things: between what the machine learns for itself, and the few things a person has to set by hand beforehand.

Take GPT-3. It holds around 175 billion “details” inside, and that entire enormous figure is adjusted by the machine while learning, with nobody typing them in. But above them sits a far smaller group, only a handful to a few dozen, that a person must choose before learning begins: how fast it learns, how many layers deep the network goes. The field calls these hyperparameters. Picture them as the few tuning pegs on an instrument: very few pegs, and turning one wrong puts the whole piece out. What is interesting is that no single setting is right for every piece; there is only the patience to try until you find the combination that suits the situation.

Running an organisation seems to me much the same. Most of the work runs on existing momentum. But a small number of root decisions, how time is allocated, the criteria for choosing people, the rhythm of deciding, quietly shape almost all of the rest.

What are the few foundational knobs that, if you turned them right, would let most of the remaining work fall into place?

To your desk

In a model, only a few knobs decide most of the result. Running a business usually turns on a small number of weighty adjustment points too.

This week, try listing the three or four knobs you genuinely have the right to turn: the meeting rhythm, the order of priorities in a day, how much the team is allowed to decide alone, or the criteria for accepting new work. Pick exactly one, adjust it lightly, then watch for a few days before touching the next.

Change everything at once and we cannot tell what actually made the difference. Adjusting one knob at a time is how we come to understand our own system.