Attention is not a spotlight you shine

Paper 29 · Pødenphant Lund (2026) · Read on Zenodo

We usually imagine attention as a spotlight: a beam we point at whatever we want to think about. Here is a different picture. Attention is not a beam standing over your mind, deciding where to shine. It is the policy that decides which of your internal races get started in the first place, and in what order. What sets that policy is your inner landscape, the structure of what already matters to you. You do not shine attention on the world. The world, filtered through what matters to you, starts the races you call paying attention.

The picture this replaces

Attention has been described in many ways: as a limited pool of resource, as a competition between things on the screen for a winner, as a filter that lets some signals through, as a spotlight you aim. Each of these treats attention as a separate faculty, something that sits above your thinking and points it. There is an earlier question worth asking. Your mind handles uncertainty by running competing little resolutions, and letting one of them finish first. Before any of that, something has to decide which resolutions even get to start. That deciding is what attention really is.

Those competing resolutions are called races. A race is your mind trying to settle something, running candidate answers until one of them wins and commits. The claim is simple: attention is what starts the races, and the thing doing the starting is your inner landscape.

The landscape that starts the races

The landscape is the structure of what already matters to you, built up by everything you have lived and learned. It is what decides, for your particular mind, which bits of the world are worth resolving at all. Pure noise, however much of it there is, does not start a race, because nothing about it pulls against what you already expect. Something that is ambiguous but relevant to you starts a race at once.

Friction Theory describes that landscape as four fields, four kinds of thing your mind treats as worth attending to. Each one starts races in its own way.

• Safety — the threat check. It produces the fast, involuntary grab of attention by sudden change and danger, the kind you cannot help.
• Meaning — the pull toward what is personally and socially significant: your own name across a room, the thing that matters to who you are.
• Ability — the draw toward what you could learn or master, content pitched just at the edge of what you can do.
• Effort — the lean away from what looks costly, the quiet steering toward the cheaper option.

The first two grew up early and run fast and largely outside your control. The last two are slower and more computed. This is the same old split psychology has long drawn between attention that is grabbed from outside and attention you steer from within, but now it has a mechanism underneath it: which field opens the race, and how fast that field works.

Starting a race, not holding one open

One idea does most of the work here. The landscape shapes which races get started. It does not reach into a race already running and steer it mid-flight. That sounds like a technicality, but it changes how three familiar things fit together.

Capture is just a race getting started: the landscape opens one on something that suddenly matters. Holding your attention looks like a separate force that keeps a race open, but it is not. It is the landscape starting the same unfinished race again, and again, because it has not yet been resolved. This is why a distraction breaks the spell so completely: it does not pry your attention loose from something, it simply stops the restarting. A good cliffhanger or an unanswered question works the same way. It leaves a race open that your mind keeps reopening until it gets its answer.

A glance that declines is the case the spotlight picture handles worst. You look at a dense block of text and turn away before you have read a word of it. Nothing was filtered out; you simply ran a cheap, fast guess that read the page as too costly, and that guess lost the competition for your attention before the real content was ever touched. The discomfort you feel is not the page being hard. It is the felt edge of a race you keep declining to run.

How long any of this lasts comes down to whether the thing can be resolved. A race that settles in an instant lets attention go again, which is what boredom and habituation feel like. A race far too big to settle pushes you away. A race that sits in the middle, open and resolvable but not yet resolved, holds you all the way through. That held middle state is what you experience as curiosity, as productive struggle, as flow. And the size of gap that holds best depends on how much you already know: a beginner and an expert are held by very different gaps.

What language models add to the picture

To test whether this is really about the mechanism and not just about brains, I look at large language models, which run the same kind of competition on a much simpler landscape. A model's landscape is not the four human fields; it is whatever pre-training and fine-tuning have laid down. That makes it a clean comparison: keep the competition, change the landscape, and see which parts of human attention survive.

• Curiosity gaps, before and after fine-tuning. This is the strong result. Given a question whose answer is genuinely under-determined, raw base models run straight through the gap and commit to an answer anyway. Fine-tuning installs a small, reliable habit of registering that something is unresolved, and it shows up in every one of five model families tested. Fine-tuning does not give the model a way to hold a race open; it changes what its landscape flags as not yet settled.
• The two sides of friction in real headlines. On a large archive of headline experiments, the friction that comes from a sentence being hard to process lowers how often people click, reliably. The friction that comes from an open curiosity gap pushes the other way, toward clicking, as predicted, though that half of the effect is weaker in the data. Difficulty and curiosity are not the same force, and they pull attention in opposite directions.
• More thinking, not a special loop. Whether the model answered in one pass, thought step by step, or ran in an agent loop, its ability to notice that a question was under-determined tracked how much computation it did, not whether it had a loop to "hold" anything. There is no separate machinery for holding a race open, in models or in us.

Where humans and models part ways

Everything points to one through-line. The competition itself is general; it runs in brains and in models alike. What differs is what starts the races. A human mind has all four field-openers, including the fast involuntary ones: threat grabs you, your own name grabs you, before you have decided anything. A language model has the slow, computed pull toward what it can handle, but it has nothing that grabs it the way danger grabs you. So it starts a narrower set of races than you do.

There is one more human capacity a frozen model lacks. When you attend, your landscape is being run over the input on the spot. But your landscape also reloads and reshapes itself between episodes, slowly, as you live. For a model that second process happens only during training, never while it is running. You keep rebuilding the very landscape that decides what you will notice next. A model, once trained, does not.

What this account claims, and what it does not

I follow one idea through: attention is the landscape-governed starting of races, with capture, holding, and decline falling out as phases of that one mechanism rather than as separate forces. I borrow, rather than re-argue, the four-field account of what your landscape cares about, which lives in the companion papers below. I offer one well-powered result and a set of predictions that invite the decisive experiments. I make no claim about consciousness.

The cite

Read on Zenodo → · Technical version · Dansk version

Related on this site:

• Paper 0 (BFT) — the four-field architecture this account borrows for which things the landscape cares about.
• Paper 1 (Friction Theory) — the substrate-universal race architecture whose competing-routes signal this paper reads as the commit-race.
• Paper 13 (Operational FT) — race-opening and recursive resolution; when an input starts a race in the first place.
• Paper 16 (Physics of Learning) — the capacity-match sweet spot that governs which races hold attention all the way to resolution.