Designing what the machines notice¶

The draft that used to take a week arrives in a minute. The queue behind it gets longer rather than shorter, and nothing in the tooling says which item deserves the next minute. Cassie Kozyrkov’s Attention Is All You Have puts that choice on the decision-maker: leadership as designing what machines notice, optimise and repeat.

Reading the 2026 Attention Is All You Have essay requires either a subscription or a request for a free copy.

Attention, in both senses¶

Her title changes one word of Attention Is All You Need, the 2017 paper that introduced the transformer. There, attention is a mechanism: the output is a weighted sum of the inputs, and the weights decide which parts of a sequence get to influence it. The first title is about building the machine. The second is about what stays scarce once it is built. The machines hold the mechanism now, and the choosing is what is left.

Neither sense starts from a list of what is important. Both assign weight rather than find it.

Human attention works the same way, and it has an older name. Lewin’s word for it was the lifespace: every fact, person, goal and obstacle that has standing for someone at a given moment. Behaviour is a joint function of the person and that set. What lies outside it has no purchase, however real it is.

Designing what a system notices then stops being a metaphor. It is designing part of a lifespace, deciding which facts get standing. That is a behavioural intervention rather than a communications exercise, and it is field theory arrived at without the name.

The half the essay has right¶

The instinct underneath the essay is sound. Ashby’s law says that a decision-maker needs enough capacity to deal with the range of things that can affect the situation they are responsible for, and the capacity usually falls short. Two moves close the gap: make the decision-maker able to handle more, or reduce the range of things arriving for attention. Attention is the second, and it is the half that usually goes unbought. What a decision-maker attends to determines which things reach the decision at all. Her own name for it is attention scaffolding, a deliberately designed way of deciding what gets through.

Advisory practice already runs instruments of that kind, in both directions. A confidence register settles how much weight a claim can bear before anyone reads the substance, and one that has drifted misdirects attention at exactly that point. A pre-mortem works the other way, adding to what is visible rather than cutting down what arrives. Its stated purpose is not prediction. It is changing what is in view before the decision gets taken.

Where it comes apart¶

Three things, and the first is the essay’s own citation.

It states Ashby’s law as a control system having to be at least as complex and capable as the system it governs. That is the folk version, and it points the wrong way for this argument. Ashby states a relationship between the range of things that can affect a situation and the range of responses available to the person trying to keep it on course. That version allows for reducing what reaches the decision-maker. The folk version suggests adding more capability everywhere. The essay consequently concludes that meeting tomorrow’s demands requires self-amplification. An argument about attention ends up using Ashby to argue for scale instead.

Second, the person deciding what gets noticed is part of the situation being shaped. Deciding what to notice is itself an act of noticing, and it depends on what that person can already distinguish. The essay writes as though attention can be designed from somewhere outside the system it is changing. There is no such position. That is why routing around the person at one point surfaces the requirement at another. A design that leaves the person out can build things with nowhere to put one.

Third, the hard part is deciding what deserves attention, and people differ in what they can tell apart. Ashby’s own example is a two-armed semaphore: an observer close enough to tell the arms apart sees sixty-four positions, one who cannot sees thirty-six. An attention system can filter only on distinctions the decision-maker already makes. Everything else is not filtered out. It is not seen. It never reaches the queue at all, and a funded blind spot is harder to find than an unfunded one.

    everything that could affect the situation
                  │
                  │   only what the decision-maker can distinguish
                  │   registers as a separate thing
                  ▼
            what registers as distinct
                  │
                  │   of those, some get surfaced, some set aside
                  ▼
            what reaches the decision
                  │
                  │   fewer available responses than distinguishable
                  │   situations reaching the decision, so different
                  │   situations have to share responses
                  ▼
                outcome

What survives¶

Two smaller observations hold. Feedback can sharpen judgement rather than replace it, and misplaced attention can produce more work rather than less. Both keep a person in the picture. “Scaffolding” does not. It suggests something designed once and then stood on. Attention has to be revised as the situation and the decision-maker’s understanding change.

The closing lines, promising an era most people will miss and an opening for those who choose otherwise, are written to a formula rather than to the argument.

The usable version is narrower and holds. Deciding what gets attended to is one of two ways of dealing with a situation that contains more things to notice than a person can handle at once. The person making that allocation is inside the situation, and the allocation can turn out to have been wrong. That is what makes it a decision rather than a property of the system.

Last updated: 9 August 2026