Architecture & AttentionPlain-language edition
Why Does It All Feel Like Too Much?
Something changed about how much language there is, and nothing changed about how much of it a person can take in. That gap is not a personal failing and it is not going to close by itself. It has also happened twice before, and both times it ended badly.
There is more language in the world than anyone can attend to. The obvious fix, an AI that reads it all for you and tells you what is true, does not work: the judgement that would take requires having been wrong at your own expense, and nothing a model does costs it anything. So the judging stays with people, and almost nothing we currently inspect can tell whether a person actually judged or simply agreed.
- Why Does It All Feel Like Too Much?
- Why Can't You Just Be Told?
- What Holds a Claim Up?
- How Would You Know?
- Who Pays If It's Wrong?
- What Would You Build Instead?
Six short pieces, each drawn from a working paper. Each one stands on its own; this is the first. They are ordered so that every piece ends by raising the question the next one is called. The full map is in Six Questions, In Order.
What changed, and what didn't
It is not that you are bad at this
You open your phone and there is more worth reading than you can read, more worth checking than you can check, and more arriving while you decide. Most people experience that as a shortcoming of their own: they should be more disciplined, better at filtering, less distractible.
It is not a shortcoming. It is arithmetic, and the arithmetic changed recently and sharply.
This piece is about what changed, why the thing that did not change was never going to, and what happened the last two times a language technology outran the societies using it.
Three things got cheap
For almost all of human history, putting words in front of people was expensive. It cost time, labour, skill, and materials. Three times, something made it much cheaper, and each time the cheapening was of a different thing.
The third one is new in kind, not just in degree. Printing and the internet made it cheaper to move words that somebody had written. The third makes the writing itself nearly free.
The part that did not change
Human attention is not a preference or a habit. It runs on equipment that was shaped for a world in which language was slow, deliberate, and came from somebody in front of you.
That equipment has known behaviours, and they are not character flaws:
- 01Emotionally charged things jump the queue. The fast, alarm-driven part of the brain responds before the slow, reflective part has been consulted.
- 02Quick payoffs get reinforced. Systems built to hold attention are built on exactly the loop that rewards reflex over reflection.
- 03Careful thinking gets tired. The reflective part is slow and expensive to run. When it is worn out, people fall back on rules of thumb and simple stories.
- 04Switching costs more than it looks. Constant interruption erodes the working memory that sustained reasoning depends on.
- 05A narrow view feels like a clear one. When what reaches you has been selected for you, the result is not clarity. It only resembles it.
The paper's own framing, which this sheet keeps: these are biological constraints, not moral failings, and the architecture worked well in the environment it evolved for.
The bottleneck moved
For most of history the hard part was producing language. Writing it, copying it, getting it anywhere took real effort, and that effort was a kind of filter. Not a good filter, but a real one: it limited how fast anything could spread, including the things that should not have.
That filter is gone, and the hard part is now on the other side. The bottleneck is attention.
Which matters, because when attention is the scarce thing, a predictable sort of material wins. Not the best material. The material best adapted to a shortage of attention.
The paper's own summary of the volume involved is deliberately coarse: humanity now produces, on any conservative reckoning, tens of trillions of tokens of new text every day, where a token is roughly the size of a short word or a piece of one. The paper says plainly that this is an order-of-magnitude estimate rather than a measurement, and that is how it is used here.
It has happened before
A pattern with six steps
The paper's strongest claim is not about any one technology. It is that when a language technology outruns the society using it, the same sequence runs, and it has run to the end at least twice.
1486, and a book
An inquisitor named Heinrich Kramer published a manual for identifying and prosecuting witches. Before the printing press, a text like that would have circulated among a small circle of clergy and stayed there.
The press had arrived a few decades earlier. Between 1487 and 1520 the book went through 28 editions and probably put fifteen to thirty thousand copies into circulation, which was a very large number for the time. It became an authority in Catholic and Protestant regions alike.
What had been local superstition became something closer to a shared European account of what was real, and that account was then used to justify around 110,000 trials and, by the paper's figures, forty to sixty thousand executions, mostly of women, over the two centuries that followed.
It did not create misogyny, and it did not create fear of the occult. Both were already there. What the press did was amplify them to a degree that earlier societies had no way to sustain.
On the numbers: the paper cites Kramer's text directly but does not attach a source to the trial and execution counts. They are in line with the general historical literature, and they are the figures on this sheet most likely to be challenged first.
2014, and a phone
Myanmar in 2014 had almost no internet. Within a few years the price of a SIM card collapsed and tens of millions of people came online at once. For a great many of them, one platform was not a part of the internet. It was the internet: the source of news, of identity, of community.
Into that arrived a coordinated campaign casting the Rohingya, a Muslim minority, as an invading and existential threat. Feeds ranked by engagement carried the most inflammatory version of any claim furthest. A new shared account of reality formed, in which a minority moved from being other people to being enemies.
What followed was a campaign of ethnic cleansing. The United Nations fact-finding mission reported in 2018 that nearly 725,000 Rohingya had fled to Bangladesh, and cited an estimate of up to ten thousand deaths, which it judged to be a conservative one.
It did not create the prejudice. It amplified it, speeding up how fast a story formed and how fast feeling spread beyond what Myanmar's institutions, or its citizens, could resist.
The UN mission's own language is cautious in the same direction. Its 2018 report calls the platform an instrument useful to those seeking to spread hate, and says the extent to which posts led to real-world violence must still be independently examined. Report of the independent international fact-finding mission on Myanmar, A/HRC/39/64, paragraph 74.
The same six steps, twice
Five centuries apart, with nothing in common technologically, the two run the same way.
What is different about now
Both previous cases involved language that a person still had to write. A book took an author. A disinformation campaign took operators.
The step that has now collapsed is the writing itself. One person with a generative model can produce more text in a day than a medieval monastery produced in a year, and it can be tailored to the reader, tuned for feeling, and delivered as fluent, credible prose.
Nothing on the receiving side has moved at all. Which means the gap in the six steps (the one at step four, where institutions and people cannot keep up with the surge) is wider now than in either historical case, and widening.
Why a better answer machine is not the answer
The trouble with an oracle
The obvious fix is a machine that tells you what is true. If the problem is too much language and not enough time to sort it, build something that does the sorting.
The paper argues this makes the problem worse, for three reasons, and it is careful to say the argument depends on a particular idea of what such a system would be: one that decides on your behalf. The author notes in the paper that not everyone defines it that way.
There is a fourth problem underneath those three, and it is the one that reframes everything: the shortage is not of intelligence. People reason perfectly well given time, clarity and context. The shortage is of the attention needed to use the intelligence they already have.
A personal note, and why it is in the paper
The paper stops in the middle and the author writes about himself: a period in which his own responses had become reflex rather than choice, and the slow recognition (not a sudden one, he is specific about that) that what needed changing was not his information or his strategy but where his attention was going.
It is there because it is the argument in miniature. People drift into lives shaped by habit rather than intention. Societies drift into information environments shaped by reflex rather than reflection. In both cases the way out is a reorientation of attention, and in neither case is more data the thing that does it.
That is also the step that connects the two halves of the paper. If attention is how a person decides what matters, then losing control of it costs more than knowledge. It costs agency.
Limits, and what to do with it
What would help instead
The answer the paper arrives at is not a better filter but a different relationship: tools that help a person navigate an infinite amount of language rather than tools that generate more of it or decide on their behalf.
In practice that means things like showing where sources actually disagree rather than resolving it quietly, flagging when the range of what is reaching you has narrowed, and noticing when the emotional temperature is high and the analysis is thin. None of these tells you what is true.
A companion piece, What Would You Build Instead?, sets out the design principles and what a system built to them commits to. They are not restated here, because they belong there and saying them twice in slightly different words would make them sound less settled than they are.
The paper itself is careful to present its principles as a floor rather than a closed set: at least these, not only these.
What this does not claim
This is the material in the series where a short version could most easily become something louder and cheaper than the paper. So, explicitly:
- ✕Technology caused these atrocities. The paper says the opposite twice, in its own words, and both statements are printed on the sheets above. The prejudice and the fear were already there. What the technology did was amplify them faster than anybody could respond.
- ✕People are irrational. The paper's explanation is the reverse: not that humans reason badly, but that humans are unusually responsive to narrative, especially when it is abundant, emotionally charged, and framed as a matter of survival.
- ✕Books, or platforms, are dangerous. Stated and rejected in the paper. The claim is about what happens when a language technology outruns the safeguards around it, which is a claim about timing and capacity rather than about the medium.
- ✕Fact-checking and media literacy will fix this. Called necessary but not sufficient. Literacy is a skill held by individuals; it does not scale against the volume. The paper's framing: the skill is one thing, the infrastructure that lets the skill be used is another.
- ✕AI is the problem. The closing position is that the same technologies that produced this contain the means to address it, which is a conditional kind of optimism rather than a comfortable one.
Three questions worth asking
- Which of the six steps are we on?The first three have clearly run. The argument of the paper is about the fourth, which is the last one where anything can still be done cheaply.
- Is this tool widening what reaches me, or narrowing it?A narrowed view feels like a clear view from the inside. That is what makes it worth asking deliberately rather than waiting to notice.
- What did I give my attention to today, and did I choose it?Not a productivity question. If attention is the mechanism by which a person decides what matters, then it is the whole of the argument applied to one day.
Where the title comes from
The breakthrough that produced modern language models came from a paper whose title said that attention was all you needed, attention there being a technical mechanism inside the model, not the human kind.
Attention is all we have.
The paper's inversion of that title, and the name of the earlier essay this one grows out of
The claim underneath it is that attention is not one cognitive faculty among several. It is the thing by which a person builds an inner life and by which a society builds a shared one: how individuals form an identity, how groups reach agreement, how a civilisation holds together at all.
Which is why the paper's sharpest line about the present is not about danger. It is that all of this accelerates everything except understanding.
Where this comes from
The argument here is the one made in a working paper called The Attention Crisis. What the paper adds is the sources, the objections, and the estimates with their workings shown, so that a reader who wants to disagree can do it in a specific place.
It is part of a series on what judgement is, why it has to be carried by people who bear the consequences of being wrong, and what that means for building with machines that bear none.