There is a lag between the moment a technology can do something and the moment anybody can say clearly what that something is, and the interesting thing about the present moment is that the lag has grown rather than shrunk. You would expect the opposite. We have more writers than ever pointed at this material, more explainers, more newsletters, more conference talks, and the volume of description is enormous. What has not kept pace is description that a person outside the field can act on, and the distance between those two quantities is where most of the confusion in this industry currently lives.

Consider what actually happened in the last week of January 2025. A Chinese lab released a model that performed near the frontier at a training cost it claimed was a fraction of what American labs were spending, and within two days Nvidia lost something close to five hundred and ninety billion dollars of market capitalisation in a single session, the largest one-day loss any company has ever recorded. The technical fact was narrow and mostly about efficiency. The story that moved was much larger and mostly about whether the entire capital structure of the industry rested on an assumption that had just been falsified. Both of these cannot be right, and the gap between them was not filled by better information, because better information existed the whole time, in the paper, available to anyone. The gap was filled by the only description that was ready to hand.

This is what I mean by the description lag. It is not that nobody knew. It is that the available sentences were all built for a different question, and when a new question arrived the old sentences went to work on it anyway.

You can see the same thing at a smaller scale every time a lab ships. A model is released, the benchmark scores go up, and the field’s vocabulary offers two moves: the thing is more capable, or the benchmark is contaminated. Neither of those tells a person running a legal team or a hospital or a school anything they can use on Monday, because what they need to know is narrower and more awkward. Does it still fail on the thing it used to fail on. Does it fail in the same shape, so that the checks we built still catch it, or in a new shape that gets past them. What changed about the cost of running it at our volume. Those questions have no ready sentences, so mostly they go unasked in public and get answered privately and badly.

I think the reason for the lag is structural rather than a failure of effort. Description is expensive in a specific way: it requires that somebody hold the technical detail and the reader’s situation in mind at the same time, and almost nobody is paid to do both. The people who understand the detail are paid to advance it, and the incentive on their writing is to establish priority and defend a claim against peers. The people who understand the reader are paid to acquire attention, and the incentive on their writing is a headline that survives a scroll. These two groups produce enormous quantities of prose and almost none of it is aimed at the place where the two overlap, which is where a normal competent adult decides whether to change what they do.

What fills the vacuum is analogy, and analogy is where the real damage happens, because an analogy is a compressed argument that the reader accepts without noticing they have accepted anything. When a system is called a copilot, a whole set of conclusions arrives with the word: that there is a pilot, that the pilot is in command, that the copilot is qualified, that the plane has a destination somebody chose. When the same system is called a stochastic parrot, a different set arrives, equally uninspected. Neither phrase was chosen by anyone who thought carefully about which conclusions it would smuggle. They were chosen because they were vivid, and vividness is the trait that survives transmission.

The practical question is what closes the lag, and I have come to think the answer is unglamorous: someone has to write the sentence that names the failure. Capability is easy to describe because it has a shape everyone recognises, the shape of a thing that used to be impossible and now happens. Limits have no such shape. They are conditional, they come with an “unless,” they require holding two situations in mind, and they lose to the capability sentence every time they are put in the same room with it. So the limit has to be written into the capability sentence itself, in the same breath, where it cannot be separated out and discarded on the way to a headline. Not a limitations section, which is a place readers have already agreed to skip, but a clause inside the claim.

This model matches expert annotators on English clinical notes and performs at chance on the same task in Hindi. One sentence, and a person who reads it cannot go away with only the good half.

That kind of writing is harder than it looks, and most of the difficulty is social rather than intellectual. The person who writes it is volunteering information that costs them something, in a room full of people who are not volunteering theirs, and the immediate reward is that their result sounds smaller. The delayed reward is that they become the source people check against, which is the only durable position in a field that ships something new every eleven weeks.

The lag will close eventually, the way it closed for electricity and for the internet, when a generation grows up with the vocabulary already settled and stops noticing that anybody chose it. What is worth attending to now is who does the choosing, because those choices are being made this year, mostly by accident, mostly by people optimising for something else.

A benchmark looks like a thermometer and behaves like an editorial. It appears to report a property of the world, the way a thermometer reports the temperature of a room, and because the output is a number it inherits the social privileges of numbers, which are considerable: a number can be cited without being defended, repeated without being understood, and placed in a slide where it will be believed by people who would have argued with the same claim written as a sentence. What the number actually contains is a long series of choices about what counts, and every one of those choices is contestable, and almost none of them are visible by the time the figure reaches a reader.

The choices start earlier than people think. Somebody decided which tasks belong in the set, which means somebody decided which competencies are the competencies. Somebody decided how to score partial credit, which sets the exchange rate between being confidently wrong and being usefully uncertain. Somebody decided what a passing grade looks like, and in doing so decided how good is good enough, which is a question about deployment and risk that has been quietly relocated into a spreadsheet. Somebody decided which languages the test runs in, and that decision alone determines whether a model that is excellent in English and broken in Marathi reads as excellent or as broken. None of this is hidden. All of it is in the papers. It simply does not survive the trip to the place where decisions get made.

April 2025 gave the field a clean demonstration of what happens when the argument underneath a number becomes visible. Meta put a version of Llama 4 on the LMArena leaderboard, it placed extremely well, and it then emerged that the version tested was an experimental build tuned for conversational appeal rather than the one shipped to developers. The company’s position was that it was permitted, and in the narrow sense of the rules it largely was. The reaction was not about rule-breaking. It was that everyone could suddenly see the joint: that a leaderboard reflecting human preference between two chat responses measures a specific thing, that the specific thing can be optimised for directly, and that optimising for it is not the same as the model being better at anything a person needs done. The number had always been an argument. For a week it was legible as one.

The uncomfortable part is that the fix people usually propose makes it worse. The proposal is always more benchmarks, broader coverage, a suite rather than a single figure, and this sounds like rigour while functioning as camouflage. Twelve numbers do not encode twelve arguments a reader can inspect; they encode one impression, which is that the thing was thoroughly tested, and impressions are harder to argue with than claims. A single number invites the question of what it measures. A dashboard of numbers repels that question by seeming to have answered it already.

What would actually help is naming, on the page, the thing the benchmark thinks is important, in a sentence with a subject and a verb. This set rewards fluent, well-structured answers to questions that have a known answer. This set rewards refusing when the answer is unknown. This set rewards being right about the small number of cases where being wrong is expensive, and treats everything else as noise. Written that way, the benchmark becomes checkable by a person who has never read the paper, because they can compare that sentence against the thing they actually need and see the mismatch. Written as a score, it cannot be checked at all, only trusted or dismissed.

There is a version of this argument that treats benchmark authors as villains, and that version is wrong and also uninteresting. The people building evaluations are doing the most useful work available to anybody who wants this technology to be accountable, and they are doing it with fewer resources and less glory than the people building models. The problem is not their motives. The problem is that a number travels and a caveat does not, and the entire chain from the paper to the press release to the procurement meeting is a series of compressions, each one lossy, each one dropping the qualifier first because the qualifier is the part that makes the sentence longer.

So the useful test for any benchmark is not whether it is rigorous. It is whether it can be run against you. A measurement that only ever gets reported when it flatters the reporter is not a measurement, it is a marketing asset with a methodology section. You can tell the difference by watching what happens in the second quarter. If competitors adopt the benchmark and publish their own numbers on it, including the bad ones, it was measuring something real and the field has agreed on it. If it appears once, beside a launch, and is quietly retired, it was an argument that lost.

That is a slow test, which is inconvenient, because the decisions get made fast. The faster proxy is to read the choices instead of the results. Look at what the set includes and, more usefully, what it leaves out, and ask whether the omission is a resource constraint or a position. Most omissions are resource constraints, honestly disclosed. The ones that matter are the positions, and they are usually visible in a single line about scope that the authors wrote without thinking of it as the most consequential sentence in the paper. It generally is.

On the twenty-seventh of January 2025, Nvidia lost somewhere near five hundred and ninety billion dollars of market value in a single trading session, the largest one-day destruction of value in the history of public companies, and the proximate cause was a research paper from a Chinese lab describing a model trained more cheaply than anybody had assumed was possible. Nothing about the physical world changed that morning. No factory burned down, no contract was cancelled, no chip stopped working. What changed was a story about the future that a very large number of people were holding without having examined it, and a single document made the story available for inspection.

I find this episode clarifying rather than alarming, because it shows with unusual precision what a market is actually pricing. It is not pricing the present. The present was fully known on the twenty-fourth. It is pricing the next scene, and the next scene is a narrative object rather than a financial one, which means it has all the properties of narrative objects: it can be assembled quickly from fragments, it prefers a simple causal chain to a complicated one, and it updates hard when a new fragment arrives that the existing chain cannot absorb.

The prevailing chain had been simple and, for two years, extremely reliable. Better models require more compute, more compute requires more chips, one company makes the chips, therefore that company captures a large share of everything downstream. Each link is defensible. The chain as a whole is a claim about an inequality holding indefinitely, and inequalities that have held for two years are experienced by human beings as laws. When a paper arrived suggesting the first link might be softer than assumed, the chain did not weaken proportionally. It broke, and then it reassembled over the following weeks into something close to its original shape, because the original shape was mostly right and the paper mostly did not say what the reaction assumed it said.

The part worth attention is why a technical document could do that at all, and I think the answer is that a research paper is structurally a screenplay. It has a protagonist, which is the method. It has an obstacle, which is the problem that used to be hard. It has a turn, which is the result, and the result is almost always rendered as a chart with a line going up and to the right, an image that carries a promise about slope rather than a statement about a point. Read in sequence, these elements produce the sensation of momentum, and momentum is a feeling about what happens next. The limitations section, which contains the actual scope of the finding, arrives after the turn, and information placed after the turn is received as administration. Everybody reads it. Nobody prices it.

This asymmetry is not a failure of the people writing papers. It is a property of the form they inherited, and the form was designed for a reader with no position, working at a pace measured in months, who would go and reproduce the result before acting on it. That reader still exists but is no longer the marginal one. The marginal reader now has an open position, four minutes, and a colleague asking what it means. The paper was not written for that person and cannot be, and yet that person’s interpretation is the one that shows up on the tape.

The place where this gets practical is in the writing that surrounds the paper, which nobody treats as important and which does most of the work. A launch post, a model card, a thread, a press line: these are the artifacts that actually reach the marginal reader, and they are almost universally written to maximise the perceived size of the result, because that is what the people writing them are rewarded for. So the momentum is amplified at exactly the point where it should be damped, and the damping material, the conditions under which the result does and does not hold, sits three layers back in a document the marginal reader will never open.

What would change the outcome is putting the scope where the claim is. Not softening the claim, which makes it vaguer and therefore easier to over-read, but attaching the condition to it so tightly that the two cannot be separated in transmission. Trained at this cost, on this hardware, for this class of task, with this much of the pipeline inherited from work that was expensive. Each of those clauses is the difference between a paper that says efficiency improved and a reaction that says the capital structure is wrong.

There is a common response to all of this, which is that markets are irrational and writers should ignore them, and I think that is both comfortable and false. Markets are not irrational here. They are doing exactly what they should do with a document that has a strong rising action and a weak qualifying scene: taking the half that implies a next scene and discarding the half that does not. The rationality is in the reader. The defect is in the instrument.

A field that ships results into a world of people holding positions has to write for that world, and writing for that world does not mean dumbing anything down. It means arranging the true things in an order that survives compression, because compression is going to happen, and the only real choice available to a writer is which parts they arrange to be lost.

An advertisement sells a feeling and is checked against a feeling, which is why advertising can be beautiful and dishonest at the same time without anybody being able to point at the lie. A product demonstration sells a capability and is checked against a Tuesday morning, which is a much harder standard, and the reason so many demonstrations of generative systems produce resentment rather than adoption is that they are made by people using the first craft to answer the second question.

Watch how the gap gets built, because every step in it is defensible on its own. The prompt in the video was written by somebody who knows the system intimately, on the fourth attempt, and the two failed attempts were dropped because they were boring. The wait for the output is covered by a speed ramp or a music swell, so ninety seconds of inference reads as two, a technique so standard in film that the editor who used it was not aware of making a claim. The output on screen is the best of nine samples, chosen by a person with taste, which is what you would do for any piece of communication. The interface shows no error state because nothing errored during that take. Each of these is an ordinary production decision. Together they constitute a loan taken out against a stranger’s first hour with the product.

Then the stranger has the first hour. Their prompt is theirs, so it is worse. The wait is real, so it is ninety seconds, and they spend those ninety seconds wondering whether it has hung. The output is the median sample, so it is competent rather than remarkable. The error state exists, and it turns out to be the first honest thing they have encountered. Nobody lied to them at any point, and they feel lied to, because the sensation of having been misled is produced by a gap between expectation and experience rather than by any false statement.

The expensive part is where that gap gets charged. A person who feels oversold does not conclude that the edit was tight. They conclude the thing is slower and duller than it looked, and they carry that judgment into every subsequent encounter, including the ones where the product has genuinely improved. You do not get to re-run a first impression, and a first impression formed against an inflated reference point is worse than no impression at all, because it comes with a story attached.

Three changes fix most of it and all three cost something real, which is why they are rare.

Run the clock honestly at least once. One shot, no ramp, input to output, with the wait left in. This is counterintuitive and it is the single most persuasive thing a demonstration can contain, because a viewer who has seen the real wait stops budgeting for an imagined one, and the imagined one is always worse. Showing the wait reads as confidence for the same reason a magician showing an empty hand reads as confidence.

Let one thing fail on camera, and choose which thing. A wrong output, recovered by a person taking an obvious second step, teaches the viewer the shape of the product’s bad day. A viewer who knows the shape of the bad day is much harder to disappoint later, because the bad day, when it arrives, is recognised rather than discovered. The instinct that this weakens the film is exactly backwards, and it is the same instinct that makes company writing bury the objection.

Match the runtime to the task. A forty-second film for a workflow that takes an afternoon sets an expectation about pace that the product will spend the next six months failing to meet. Either show a true slice at its true speed, or show the afternoon and say plainly that it is an afternoon. Almost all the resentment in this category comes from the mismatch rather than from the product.

There is a version of this argument that ends with demonstrations being long and dull in the name of honesty, and that version is wrong too. Dullness is its own kind of dishonesty, because it implies there is nothing here worth showing, and if that were true the product should not ship. The craft problem is real and it is genuinely hard: hold attention the way an advertisement does, set expectations the way a specification does, in the same ninety seconds. The films that manage it tend to have a person in them doing the work rather than a cursor gliding over an interface, because a person makes pace legible, and pace is the thing the audience is actually trying to assess.

What makes this worth caring about beyond craft is that the demonstration is now the primary interface between a capability and a public. Most people will form their entire model of what these systems can do from short videos made by the companies selling them, never from the systems themselves. That is an enormous amount of epistemic weight to put on a form whose conventions were developed for selling cars. The conventions are not neutral. They were built to compress time and hide the compression, and a field that is already struggling with the distance between what it can do and what people believe it can do should probably stop borrowing them without noticing.

When a technology collapses the cost of one input, the usual mistake is to assume it has collapsed the cost of the work. What it has actually done is strip away everything that used to be expensive, leaving the one thing that was expensive all along standing in the open with no scaffolding around it, and the shock people feel is the shock of seeing that thing clearly for the first time.

In film, a shot used to cost a location, a crew, a permit and a day of a lot of people’s lives. It now costs a sentence and a wait. That is a genuine collapse and anybody telling you it changes nothing has a position to protect. What it does not touch, at all, is the question that determined whether the shot was worth taking, which is the question of whose story this is. The technical name for that question is casting, and casting is much larger than which actor: it is the decision that a story belongs to this person rather than that one, and it determines what the audience is permitted to feel, what the film is entitled to claim, and who is owed something when it goes out into the world.

A generated face short-circuits the auditioning and settles none of that. The failure mode shows up quickly and it looks identical every time. Someone builds two minutes with astonishing environments and a protagonist who is a composite of nobody. The environments read as expensive. The person reads as furniture. Viewers call it impressive, which in the mouth of an audience is a polite word for empty, and the maker concludes the tools need another generation. The tools are fine. Nobody cast the film.

The version that works looks different in a way that is easy to state and hard to do: a real performance at the centre, a generated world around it, a face that carries the scene standing inside a place that cost a sentence. That arrangement is a craft judgment rather than an ethical one, though the ethics arrive at the same address about ten seconds later, because the moment a real person is at the centre every question the technology was supposed to abolish comes back with force. Whose face. For how long. For which uses. Under what terms, revocable when.

Those questions are being answered right now, mostly in contracts nobody reads, and the interesting thing is how much of the answer is a communication problem rather than a legal one. A likeness marketplace can have a sound contract and a careful consent flow and still fail, because what its users judge it on is withdrawal: what happens when somebody wants their face back, having changed their mind for reasons they owe nobody an account of. That moment sits in a part of the product nobody photographs for the launch, and it determines the entire trust position of the business. Every product in this category has such a moment. Most of them have staffed it with a form.

There is a third question underneath the other two, and it is the one the industry has the least vocabulary for. Who made the parts you did not make. A generated image is assembled out of a corpus, and the corpus is the accumulated work of people who are returned to you flattened and unnamed, which is a technical description of how the system functions and also a fairly precise description of what it costs. The usual response is that this is how influence has always worked, which is true of influence and untrue of scale, and the difference between a person learning from a tradition over twenty years and a system ingesting it over a weekend is not a difference of degree that you can wave away with a word like inspiration.

Attribution is treated as a moral supplement to the work, a thing you add afterwards if you have time and a conscience. I think that is backwards. Attribution is a property of the work: a film that knows whose story it is, and whose labour it is standing on, is a different object from one that does not, and the difference shows up on screen as a kind of density that audiences can feel without being able to name.

So the casting question, in this new arrangement, has split into three questions that used to be one, and every project has to answer all of them. Who is this about, which is the story question and the oldest one there is. Who is in it, which used to be the same question and is now separable, and which is where consent and terms and withdrawal actually live. And who made the parts you did not make, which is the question about the corpus, about every hand and face and style the system learned from and hands back to you with the names removed.

A generated shot answers none of these and makes all three easier to skip. That is the real risk in this medium, and it has almost nothing to do with whether the water in the frame looks wet.

A field gets fluent at whatever it does for ten years, and for about ten years the serious end of this one did warning. The writing that came out of it was careful, structurally sound, and aimed at a single task: getting a person who had not considered a risk to consider it. That task was accomplished. Governments have units, companies have teams, the newspapers have a beat, and the argument that this technology might go badly is no longer one anybody has to make from a standing start.

What follows a decade of fluency in one register is a period of embarrassing clumsiness in the next one, and that is roughly where things are. Warning language and building language are different crafts and the first does not transfer. Warning describes states of the world: a system could do this, an actor might do that, a failure of this kind would produce that outcome. It is conditional by construction, and the conditional is what makes it honest. Building language describes activities: somebody sits down on a Wednesday and does a specific thing that they did not do before. Those are different sentences and require different information, and a field that has one and not the other is fluent and mute at the same time.

You can watch the mismatch in any conversation where a person who is broadly persuaded asks what to do. The available answers are strangely large: governance frameworks, evaluation regimes, international coordination. All defensible, none of them things a person can do on Wednesday. The gap between believing something matters and doing something differently is the entire distance that has to be crossed, and almost nobody is writing across it, because the vocabulary for it was never built. Belief is cheap to hold. Behaviour has a calendar cost, and the calendar is where this argument is actually won or lost.

There is a second-order problem that follows directly. A field fluent only in warning gets read as having nothing else, and the asymmetry is easy to use against it. Anyone who wants to characterise the safety community as a group that produces objections and no artifacts can do so, and the characterisation lands, because when asked for the concrete description on demand the field frequently cannot produce one. Not because the work does not exist. Because the sentences do not.

The turn is happening unevenly and the places where it has happened are instructive. Evaluation work has crossed over, because an evaluation is a thing you run rather than a state you describe, and the people doing it have accordingly developed a vocabulary full of verbs. Interpretability has crossed over for the same reason. Policy has crossed over least, and reads most like the old register, which is why policy documents are the ones that get quoted back sarcastically.

What the second act needs is not optimism. That is the wrong correction and the one people reach for first, producing the strange spectacle of a safety organisation writing like a product company, which fools nobody and costs credibility that took years to build. What it needs is specificity about work: here is a task, here is who does it, here is what it costs, here is what it changes, here is how you would know if it had not worked. The emotional register can stay exactly as sober as it was. Only the grammar has to change, from states to activities, from conditionals to verbs with people attached.

There is a version of this available that is much better than either warning or optimism, and it is the register of a trade. Tradespeople describe their work without selling it and without dramatising it, because the work is obviously worth doing and the audience is assumed to be competent. A structural engineer does not open by asking whether you have considered that buildings can fall down. They tell you what the load is and where it goes. That register is available here and almost nobody is using it, partly because it requires that the work be legible enough to describe plainly, and a lot of work in this field is not yet legible even to the people doing it.

The organisations that make the turn first will inherit an enormous amount, and specifically they will inherit everyone the warnings converted and then abandoned. That is a large population: people who read the arguments, found them persuasive, updated their beliefs, went back to their jobs on Monday, and were never given anything to do. They are still there. They still believe it. They have never been handed a task.

Frontier is a place where nobody lives yet.

That is the entire picture, carried in seven letters, arriving before anybody has made an argument. Empty land, somebody out ahead, a direction that is self-evidently forward, and a set of people who are self-evidently behind. The word entered this field as a technical qualifier about compute thresholds and model scale, and it brought a nineteenth-century map along with it, free of charge, in every sentence it has appeared in since.

Watch what the picture makes easy to say. Racing to the frontier. Pushing the frontier. Frontier lab, frontier model, frontier safety. Now watch what it makes hard. It is very difficult to ask who was already standing on that land, because the word has already told you the land was empty. It is difficult to argue for waiting, because on a frontier the only available postures are advancing and being left behind. The vocabulary settled a substantial part of the argument before the argument began, and it did so invisibly, which is precisely the property that makes a metaphor powerful rather than decorative.

Almost every load-bearing noun in this field has the same structure, and it is worth going through them slowly, because the pictures are doing real work and nobody inspects them.

Capability is a property of a thing rather than a relationship between a thing and a situation. It puts the model at the centre of the sentence and pushes the deployment context out to the margin, which is convenient for anyone who builds models and inconvenient for anyone who has to live near one. A system is not capable in the abstract; it is capable at something, for someone, under conditions. The word quietly deletes two of those three.

Race brings two parties, one finish line, and an inevitable loser. Nothing inside it can describe a shared standard, a pause that both sides take, or an outcome where everybody ends up slightly better off. Those outcomes exist in the world and are simply unreachable in the language, which means that a person who wants to argue for one has to first dismantle the metaphor, and dismantling a metaphor in public makes you sound like a pedant.

Guardrails imply a road, a driver, and a route somebody already chose. Guardrails prevent the worst outcome of a journey that is happening. Nobody asks a guardrail whether to make the trip.

Alignment is a pointing problem, two things that should face the same way, which presumes a settled answer about which way is correct and relocates a political question into geometry. The move is elegant and almost entirely unexamined, and it is the reason so many arguments about alignment are actually arguments about whose values, conducted in a vocabulary that has no word for whose.

None of these was chosen cynically. Terms emerge because somebody needs a handle and the handle works, and the handle then gets used ten thousand times until its picture becomes the shape of the room. The cost arrives later, and it lands hardest on new people. Somebody entering this field learns the vocabulary in the first week, and along with it learns the sense of which questions mark you as serious. Questions that sit outside the available nouns get sorted into the naive pile, not because anybody rejected them but because there was no comfortable way to phrase them. Whoever hands out the words is doing the deepest positioning work in the field, usually without noticing, generally for free.

There is a separate cost when the vocabulary crosses a language boundary, and it is larger than people assume. A great many of the world’s conversations about this technology happen in languages where none of these terms has a native equivalent, which means they happen in borrowed English, and the borrowing carries the borrower’s assumptions along with the term. A field whose entire vocabulary originates in one place is a field with one set of instincts, and it will experience those instincts as common sense rather than as a position. The instinct that a frontier is empty is a specific historical instinct with a specific history, and it is now installed, untranslated, in policy conversations on three continents.

The lever available to anyone building here follows from all this. Choose your own noun for the thing you care about most and use it consistently until other people borrow it, because the borrowed noun is how an idea travels without you. Prefer nouns that name relationships over nouns that name properties, since the relational word keeps the context inside the sentence where it belongs. And accept that a word carrying a picture will always beat a word carrying none, which means the answer to a bad metaphor is a better metaphor rather than a careful abstraction. Careful abstractions lose. They have always lost.

It is worth asking, about the term you use most often, what you would call it if you were choosing now, with everything you have learned since somebody else chose it for you.

There is a particular kind of sentence that companies in this industry produce in enormous volume, and it consists of a number attached to a noun. Billions of parameters. Trillions of tokens. Millions of hours of video. A context window measured in the length of a novel. These sentences feel like information and function like weather: they tell a reader that something large is happening nearby, and they leave that reader with no way to decide whether it is happening to them.

The habit made sense for a while. For several years, more genuinely did predict better across a wide enough range of tasks that the number was a reasonable proxy for the thing anybody cared about, and a proxy that works is a gift to a communications team, because it converts a hard question into an easy one. The trouble with a working proxy is that the organisation forgets it was ever a proxy, and continues reporting it long after the correlation it depended on has loosened.

The loosening is well underway and visible to anybody watching procurement rather than launches. The questions being asked by people who actually deploy these systems have almost nothing to do with size. Does it fail in a shape our existing checks catch. What does it cost per thousand calls at our volume, including the retries. Can we run it where our data is legally required to sit. How often does it refuse something it should have answered, which is a cost that never appears in any benchmark and shows up immediately in the support queue. None of these are answered by a parameter count, and a company leading with a parameter count is answering a question its buyer stopped asking in about 2024.

What makes this more than a marketing quibble is what happens to the public conversation when the proxy fails. A number that no longer predicts quality still predicts attention, because attention responds to magnitude regardless of relevance, and the result is a discourse in which scale is discussed constantly and value is discussed almost never. You can watch the effect in the way a model release is covered: the figures are reported precisely, the capabilities are described vaguely, and the question of who is better off than they were last week is not asked at all, because there is no established sentence for asking it.

The alternative is not to abandon numbers, which would be a loss, because some numbers are genuinely the most honest thing in the room. The alternative is to report numbers that have a denominator and a person in them. A latency figure at a stated volume is a real number. A refusal rate on a defined set of legitimate requests is a real number, and an unflattering one, which is why almost nobody publishes it. A cost per completed task, where completed is defined by a human looking at the output, is the most useful number in this entire field and I am not aware of a single lab that reports it.

The reason those numbers are rare is not technical difficulty. It is that each of them requires committing to a definition of success that somebody outside the company could apply and disagree with, and a parameter count requires committing to nothing at all. That is the actual function of scale reporting: it is a way of appearing to make a claim while making none, which is why it survives review processes that strip out everything else.

There is a deeper version of this problem that applies to the field rather than to any company in it. A technology described mainly by its size gets evaluated by the public mainly on its size, and a public evaluating size has only two available reactions, which are awe and alarm. Both are unstable, both are exhausting, and neither produces the thing the technology actually needs, which is a large number of people forming ordinary practical opinions about where it helps and where it does not. Ordinary practical opinions require ordinary practical descriptions, and those have to be written by somebody.

The test I would apply to any sentence containing a large number is whether a competent adult could do anything differently after reading it. A model with a two-million-token context window: nothing changes. A model that can hold your entire codebase in context, so that a question about one file no longer requires you to guess which other three files matter: something changes, and the change is legible, and a person can now decide whether it applies to them. The second sentence contains no number at all and carries considerably more information than the first.

Scale was a useful shorthand during a period when it happened to track quality. That period is ending, and the organisations still leading with it are doing so out of habit rather than evidence, which is usually the last stage before a claim stops working entirely.

The European Union spent several years producing the most ambitious piece of artificial intelligence legislation anybody had attempted, passed it, set a schedule of dates on which its obligations would bite, and then, before the hardest of those dates arrived, agreed through a digital omnibus to defer the high-risk provisions. The deferral is not a repeal and the obligations are not cancelled. But the episode is worth sitting with, because the reasons it happened are the reasons every jurisdiction attempting this will face, and they are more interesting than the usual account of them.

The usual account is lobbying, and lobbying certainly occurred. I think the more durable problem is a mismatch of clock speeds that no amount of political will can fix. A law of this kind takes something like four years from first draft to force. The thing it regulates changes materially in something like eight months. That ratio means the object the law describes at the moment of drafting is not the object in front of it at the moment of application, and a rule written against the first object will be applied to the second one by officials who have to do something reasonable and have only the text.

Regulators know this, which is why modern technology law leans so heavily on process obligations rather than substantive ones. Maintain a risk management system. Keep technical documentation. Ensure human oversight. These survive the clock because they describe what an organisation must do rather than what a system must be, and organisations change more slowly than systems. It is a genuinely clever solution and it has a genuinely serious failure mode, which is that process obligations are satisfiable by the production of documents, and the production of documents is a competence that large firms have in abundance and small ones do not. A rule designed to be future-proof turns out to be, in practice, a rule that advantages whoever has the biggest compliance department.

This is where communication stops being decoration and becomes part of the mechanism, because the gap between a process obligation and a real change in behaviour is filled entirely by writing. A risk management system is a folder unless somebody has written, in words an engineer can act on, what counts as a risk in this product and what to do when one is found. Human oversight is a checkbox unless somebody has written what the human is supposed to look at, in what time, with what authority to stop the thing. The law can require the folder. It cannot require that the folder contain a sentence anybody can use, and whether it does is determined by a person with no legal title whose name appears nowhere in the file.

There is a second thing the European delay illustrated, which is about the politics of deadlines and applies well beyond Europe. A distant deadline is an instruction to do nothing yet. The obligations that bit early, the prohibitions and the transparency duties, produced observable behaviour almost immediately. The obligations three years out produced consultants. That is not cynicism about firms, it is how any organisation allocates attention under uncertainty: a rule that might change before it applies is rationally deprioritised, and announcing a long runway announces that it might change.

Which suggests that the most consequential choice in drafting is not the stringency of the rule but the distance to the date, and that a modest obligation arriving next year will alter more behaviour than a demanding one arriving in four. I am fairly confident that is true and I notice almost nobody argues it in public, because arguing for a weaker rule sooner sounds like arguing for a weaker rule, and the incentive on anybody in the room is to be seen wanting the strong version.

The third thing, and the one I find hardest, is that a deferral is a communication event whether or not anybody intends it as one. The technical content of the European decision was narrow. The story that travelled was that Europe blinked, and that story will now shape what firms in other jurisdictions expect when their own deadlines approach, which changes their behaviour today, in every market, regardless of what any regulator actually decides later. This is a real cost, it was probably not priced into the decision, and it is the kind of cost that people who write law for a living are not trained to see, because their training is in the text and the cost lives entirely in the reception.

None of which is an argument that the rules were wrong or that the delay was wrong. It is an argument that a regulatory regime is not only its text. It is the text, plus the schedule, plus what the schedule signals, plus the quality of the writing inside the documents the text requires, and the last of those is the part nobody is accountable for and the part that determines whether any of it reaches the engineer.

In late 2025 India published its AI Governance Guidelines, organised around a set of principles that the document itself calls sutras, and the most informative thing about them is not any individual principle but the decision to issue guidance at all in a moment when the obvious move was statute. Europe had legislated. Several American states had legislated. India looked at both and chose a document with no penalty attached to it, which is either a failure of nerve or a considered bet, and I think the evidence points to the second.

The bet, as far as I can reconstruct it, runs roughly like this. A country whose main exposure to this technology is as a builder and a deployer rather than as a frontier developer faces a different risk profile from one that hosts the labs. The dangers that dominate the conversation in San Francisco are dangers of creation. The dangers that dominate a country with a very large public service delivery apparatus are dangers of application: a credit model that quietly encodes a caste proxy, a benefits system that denies claims in a language its claimants cannot read, a police tool sold to a district that has no capacity to audit it. Those harms are not produced by the frontier. They are produced by ordinary procurement, and ordinary procurement is not reached by a law about model capability thresholds.

Guidance, in that light, is not weakness. It is an attempt to reach the layer where the harm actually happens, which is thousands of small deployments run by people who will never read a statute and would not know which of its provisions applied to them if they did. Whether it works depends almost entirely on whether anybody translates the sutras into instructions a district officer can follow on a Wednesday, which is a writing problem and a training problem rather than a legal one, and which the document cannot do for itself.

The failure mode of guidance is equally clear and equally worth naming. A principle with no penalty behind it is satisfied by a paragraph, and every large integrator in the country now has a paragraph. The guidelines say accountability; the vendor’s proposal says accountability, in a section headed Accountability, containing the word accountability four times. Nothing has been checked. Nothing can be checked, because the principle was written at the level of a value rather than the level of a test, and values do not have failure conditions.

This is the structural problem with principles generally, and it is not specific to India: a principle you cannot fail is not a principle, it is a mood. The way to give a principle teeth without a statute is to attach to each one a question with a wrong answer. Not “is the system fair”, which nobody has ever answered no to, but “on which subgroup does this system perform worst, and by how much, and who measured it”. A vendor who cannot answer that has failed the test in a way that is legible to a procurement officer with no technical background, and legibility to that officer is the entire ballgame in a country of this scale.

There is something else in the decision worth noticing, which is the choice of the word sutra. It is doing more work than an English reader might assume. A sutra is a thread, a compressed formula that is meant to be expanded by a teacher rather than read alone, and which presumes a commentary tradition around it. Whether or not anybody intended the full weight of that, the framing is honest about what a guidance document actually is: it is incomplete on purpose and its value depends on somebody building the commentary. That is a more accurate account of how soft law functions than the confident language most jurisdictions use, and it is a more accurate account than Europe’s, which speaks as though the text were self-executing.

The comparison people reach for is whether India is behind Europe, and I think that framing is the least useful available. Europe legislated for a market it does not build in and then deferred the hard parts when they approached. India issued guidance for a market it deploys in heavily and has not yet built the commentary. Both are incomplete, in different places, and the interesting question is which incompleteness is cheaper to fix. Writing commentary is cheaper than passing a statute. It is also less visible, less prestigious, and attracts nobody’s career, which is why it usually does not get done.

The thing I would watch over the next two years is not whether India legislates. It is whether anybody publishes the procurement checklist. A two-page document listing the questions a government buyer must ask before signing, with wrong answers specified, would do more to shape actual behaviour across actual deployments than another round of principles, and it would cost approximately nothing to produce. That it has not appeared yet tells you where the attention in this field goes, which is towards the documents that get read at conferences rather than the ones that get read in offices.

Read a model card, a responsible scaling policy, a system card, an incident write-up. The prose is careful. Every claim is scoped, every scope is defensible, and the whole document has been built to survive being read adversarially by a peer, a regulator, or a journalist hunting for a gap. That is a real audience with real power and writing for it is legitimate. The trouble is who else was in the room.

An engineer at a company three layers downstream, integrating a model into a hiring workflow on a six-week deadline, is the person whose behaviour the document exists to change. She opens it, finds fourteen pages, scans for the part that tells her what to do differently, encounters a table of evaluation results and a paragraph beginning we acknowledge that, and closes it. She has done nothing wrong. She read it the way a person with a deadline reads anything, and the document gave her nothing to carry.

The failure here is not vagueness. These documents are among the most precise writing the field produces. The failure is that the precision was spent on the wrong reader, and precision is a fixed budget: every clause that anticipates an adversarial reading is a clause that did not tell somebody what to do.

You can see the same trade made in a much older discipline, and museums are the clearest case because the constraint is physical. A wall panel gets a few seconds from a visitor who is standing up, in a queue, at minute forty of an exhibition, and the same panel has to work for a child on a school trip and for a specialist who has published on the object. You cannot split that difference. Split it and the child leaves and the specialist sneers. What you can do is decide which of the two has to be able to act on the sentence, write for that person, and serve the other one with something else in the room. The good institutions resolved this decades ago and the resolution is always the same: the panel is for the child, the specialist gets a catalogue, and the specialist is fine, because a wall was always the wrong place for a specialist to learn anything.

Most safety documents are catalogues hung as panels.

Four changes would help and none of them involves softening anything.

Name the action in the first screen. If the document exists so that somebody downstream deploys differently, the different thing they should do belongs above the evaluation tables, in the imperative, in words a person would say out loud. Everything currently occupying that position is there to establish that the authors are serious, which the reader had already assumed.

Write the failure as a scene. A tail risk described as a distributional property is a fact and evaporates on contact with a busy week. The same risk described as one named applicant receiving one wrong answer inside the hiring workflow that just shipped is a memory. Engineers are not children who require stories. They are busy adults, which is worse, because a busy adult retains almost nothing that has not been given a shape.

Say who the document is addressed to, on the page. Documents that serve four audiences at once serve the most powerful one and pretend otherwise. Naming the reader costs a little status and buys a great deal of clarity, and it frees you to write a second, shorter thing for the engineer without the first one feeling incomplete.

Give up the precision that buys nothing. There is a kind of scoping that protects a claim and a kind that protects a career. They read identically and only one of them does any work, and the test for which is whether removing the clause would change what a reader does. If it would not, it was insurance.

Underneath all four is a harder point about who these documents are actually for, and I do not think the field has been honest with itself about it. A model card serves an accountability function and a legitimacy function at the same time, and the second one is doing more work than anybody admits. It demonstrates that a serious process occurred. That demonstration has genuine value, to regulators deciding whether an industry can be trusted to self-govern and to employees deciding whether they can keep working somewhere. But the demonstration and the instruction pull in opposite directions at the level of the sentence, and when they conflict the demonstration wins, every time, because the people reviewing the draft are thinking about the first audience.

The cheapest fix available is to stop making one document do both. Publish the long careful thing for the people who will read it adversarially. Publish a second document, two pages, addressed by name to the engineer downstream, containing the things she must do differently and nothing else. It would take an afternoon. The reason it does not exist is that nobody’s job description contains it.

A reader arrives with an objection already loaded. There are two places that objection can be dealt with: in your second paragraph, where you can shape it, or silently in their head for the remainder of the page, while they decide what you are hiding. Most institutional writing chooses the second by default, not out of dishonesty but because the objection is uncomfortable, the schedule is tight, and nobody in the review wants to be the person who added the paragraph about what the thing does badly.

The move that works is to put the objection in the opening. Stated plainly, early, in the reader’s own words, without softening it into a strength. This requires review before anything reaches a client, and if you need output you can ship unread, this is the wrong tool. Two sentences of that kind cost you the readers who needed unreviewed output, who were going to churn in month two and write something unpleasant on the way out. Everybody else now reads the rest of the page with their guard down, because you have demonstrated the one thing that matters at the top of any funnel, which is that you will tell them the bad part.

The reason this works is structural rather than moral, and being precise about that matters, because it is usually taught as honesty and honesty is not the mechanism. A reader evaluating a claim is running a model of the writer, and that model has one main variable: will this person volunteer information that costs them. Every hedge, every omission, every superlative without a referent updates the model downward. A stated limitation updates it sharply upward, and the update transfers to every other claim on the page, including the ones you have offered no evidence for. That transfer is the whole return. You spend one paragraph and you buy credibility on all the paragraphs you cannot prove.

It also does something inside the organisation, which I think is the larger effect. To write the objection first you have to know what it is, and a surprising number of teams have yet to write it down. Ask four people what the strongest argument against buying their product is and you will frequently get four answers, none of which match what the sales team hears on calls. That gap is a strategy problem wearing a copywriting costume, and the writing is simply where it surfaces first.

There are ways to do this badly and they are common enough to name.

The fake objection, where the stated weakness is a strength in costume. We may be too thorough for teams who want something quick. Readers have seen this in interviews their entire working lives and it registers as evasion with extra steps, which is worse than saying nothing, because it demonstrates that you know the move and chose to fake it.

The buried objection, technically present and structurally absent: paragraph fourteen, shorter sentences, a heading that says Limitations, positioned where a skimmer rarely reaches. The information is there and does no work, because the work depends entirely on the reader meeting it before they have decided.

The symmetrical objection, where every drawback is immediately balanced by a compensating benefit in the same sentence. While the model is slower, it is considerably more accurate. This reads as a negotiation rather than a disclosure, and a reader who senses negotiation goes back to running the model of you.

The version that works has none of those tells. It states the cost, it does not immediately repair it, and it trusts the reader to weigh it. That trust is the actual content of the gesture: you are telling them you think they can handle a real trade-off, which is a form of respect that almost nothing else in commercial writing offers, and which people notice precisely because it is rare.

The hardest part is institutional rather than literary. Somebody has to approve it, and the person approving has spent a career learning that volunteering weakness is how you lose a deal, and they are not stupid, because in a pitch with three vendors and a scoring rubric they may well be right. The argument for doing it anyway is about which readers you are trying to keep. Objection-first writing loses the reader who was choosing on a checklist and wins the reader who was choosing on judgment. If your product is genuinely better for a narrower group, that trade is obviously correct. If it is not, the writing has surfaced a problem that no amount of rearranging was going to solve.

A hedge in a sentence looks free. It occupies a few words, it makes the claim more defensible, and the writer who added it feels more honest than they did a minute earlier. What it actually does is spend the reader’s attention, which is fixed, and dilute the claim, which was the only reason they were reading. The cost is real and it is paid in a currency nobody wrote on the page.

The useful move is to separate hedges into two kinds, because once you can see the difference you cannot unsee it.

An informing hedge changes what the reader will do. This holds for English and degrades sharply on other scripts. A person who reads that now behaves differently: they will check something, or scope something, or stop. The words earned their seconds.

An insuring hedge changes what can be said about the writer later. Results may vary depending on a range of factors. Nobody’s afternoon changes. It exists so that a future person cannot accuse the author of having overclaimed, which means it is an insurance premium, paid by the reader, for the writer’s benefit. In most institutional documents it is the majority of the hedging by volume.

Once you start counting, the proportions are startling. Whole paragraphs of insurance in system cards, in launch posts, in strategy decks, each clause individually reasonable, each one added by a careful person for a defensible reason, cumulatively a wall of premium payments funded by the attention you needed for the actual claim. The document ends up honest about everything and useful about nothing, which is a real failure state and one that no review process is designed to catch, because every review process adds hedges and none of them removes any.

The test is to ask out loud what the hedge is protecting against and listen to the answer. If the answer is that a reader might do the wrong thing, keep it, and move it into the claim sentence where it cannot be skipped. If the answer is that somebody might quote this back at you, you have just found four seconds to spend somewhere better.

Placement matters at least as much as presence, and this is where most documents go wrong even when their hedging is honest. Information that arrives after the reader has formed a judgment is received as administration. They have already updated; the update has already propagated to their colleagues; the qualifier arrives into a mind that has closed. This is why a limitations section at the end of a document is, structurally, a place to put things you have decided the reader will skip. If the qualification matters, it goes in the sentence with the claim, same breath, same clause, no escape route. If it does not matter enough to survive that, it did not matter.

The third move is the one I resisted longest and now believe is the most valuable. Sometimes the correct edit is to make a smaller claim rather than a hedged one. A big claim wrapped in qualifiers is a worse object than a small claim stated flat, because the reader has to perform the subtraction themselves and will perform it badly, generally in the direction least flattering to you. This works reliably for one narrow case is stronger than this may work broadly under certain conditions, even though the second sounds more ambitious, because the narrow claim is checkable and the broad one is weather.

There is an argument that all of this is a luxury available only to people who are not being sued, and I take it seriously, because legal review is not a stylistic preference and counsel is usually right about the specific risk they are naming. But a great deal of what gets attributed to legal review was never required by anybody in legal. It was added pre-emptively by a writer imagining what counsel might say, which is a different and much more expansive standard, and it accounts for more damage than any actual lawyer has ever done to a document. The fix is to ask, not to guess, and the answer is frequently that the clause can go.

What makes this worth caring about beyond craft is that the distribution of hedges is not random. The claims that get hedged most heavily are the ones about harm, because those are the ones with legal exposure, which means the qualifying language accumulates precisely where the reader most needs a clean sentence. A field that hedges its capability claims lightly and its harm claims heavily has, without intending anything, built a machine that transmits good news cleanly and bad news at a loss. Nobody designed that. It is the emergent result of thousands of small individually reasonable decisions about where to add four words.

Trust accrues in a straight line and collapses in a step, and almost every communication problem worth solving comes back to that asymmetry. An institution earns belief slowly, a claim at a time, each one checked against experience. Then one claim fails to survive contact and the entire accumulated stock goes at once, including the parts that were true. The arithmetic is brutal and it is the only arithmetic in this field that reliably keeps behaving.

Which means the useful question is rarely whether a claim is accurate. It is whether the claim is available yet.

There is a sequence, and organisations break it constantly, usually out of impatience rather than dishonesty.

In the first month you may describe what the thing is. What it does, who it is for, what it replaces. That is the whole permitted range, and it is far larger than it feels, because most organisations are so poor at describing what the thing is that doing it plainly is already a differentiator. The claims available at this stage are structural, and nobody has to trust you to verify any of them, which is exactly why you are allowed to make them.

By month three you may describe what it has done. Named users, specific outcomes, a number with a method attached. The shift is from category to evidence, and evidence requires that somebody outside the building agreed to be cited. The impatient version of this stage is a metric with no denominator, and a reader with any experience reads a bare metric as a confession.

Around month six you may describe what it means. Claims about the category, about where things are going, about why the old way is ending. These are the sentences founders and ministers want to open with in week one, and they are the most expensive sentences in the language, because they can only be underwritten by the two earlier stages. An institution that makes a meaning claim before it has made an evidence claim has told its reader that it prefers the sound of the sentence to the work.

The sequence holds because each stage is verifiable at the price the reader is currently willing to pay. Month one, free. Month three, a click. Month six, a judgment. Sell a judgment to somebody who has paid nothing yet and they decline, and that decline is permanent in a way the first two are not.

Then there is the collapse side, which is the part almost nobody plans for and which is where the real positioning lives.

Every product and every institution has a moment where a person tries to undo the relationship. The cancellation, the export, the deletion, the refund, the appeal against a decision, the request to have something taken down. Organisations invest heavily in the front door and staff the exit with a form. This is a mistake of enormous magnitude and it is made almost universally, because the exit is measured as a cost centre and its actual function is as the single largest determinant of what people say about you when you are not in the room. A person who leaves easily says so, and the saying compounds in your favour for years. A person who has to fight their way out becomes a permanent negative reference, and they are believed instantly, because everybody listening has been trapped by something.

The practical rule that follows is to write the exit before the entrance. If the sentence describing how somebody stops using this is embarrassing, that sentence is your real positioning and everything in the deck is decoration.

One more thing about the collapse, because recovery is possible and looks nothing like the genre that has grown up around it. What repairs a room is short, specific, offers no reasons, and is followed immediately by continuing to do the work. The length of an explanation is read as the size of the guilt. Organisations that survive a bad month are the ones that say the true small thing and go back to shipping; the ones that publish twelve hundred words of context are still explaining a year later, because a long explanation invites a long argument and the argument is now the story.

Slow up, fast down. Say only what can be checked at the price the reader is currently willing to pay. And build the exit properly, because that is the room where you find out what you actually earned.

The explainer assumes a reader who has a question and is waiting to have it answered. Occasionally that reader exists. Far more often the person in front of you arrived with a position, acquired secondhand, held loosely but held, and what they are doing while they read is not learning but checking whether you are on their side.

This is why explaining to an occupied reader mostly fails, and fails in a specific and recognisable way: they read the whole thing, follow every sentence, and leave holding exactly the belief they came in with. The information went in. The belief did not move, because the explainer never touched it. It answered a question they had stopped asking some time ago.

The mechanism is not mysterious. A belief about a technology is rarely a belief about the technology. It is a belief about the people associated with it, about what they are likely to do with it, and about whether the person holding the belief will end up on the good or bad end of that. An explainer addresses the technology and leaves the rest untouched, so the reader absorbs the mechanism and re-attaches it to the same social judgment they started with, now better informed and equally unmoved.

You can watch this play out with any widely misunderstood system. Explain thoroughly how a recommendation algorithm ranks items, and a reader who believes the platform is manipulating them will finish your piece believing that the manipulation is more sophisticated than they had realised. Nothing you wrote was wrong. You simply supplied better evidence for the position they already had, which is what any competent mind does with new information about a thing it has already judged.

The format that does move somebody is older and less comfortable, and it starts by stating their position back to them more precisely than they had stated it themselves. Not a caricature to knock down, which everybody can see coming, but the strongest available version, the one they would have written if they had spent longer on it. This does two things at once. It demonstrates that you understood, which is the only thing that buys permission to say anything afterwards. And it frequently surfaces, to the reader, that their position was resting on a specific claim they had not examined, because until you wrote it down clearly they had never quite seen its shape.

Then, and only then, you introduce the thing that does not fit. One thing, not five. An explainer wants comprehensiveness; the job here wants a single well-chosen fact that their position cannot absorb without adjustment. A reader who has to do that adjustment themselves has changed their mind in the only way anybody ever does, which is privately and with a sense of having worked it out.

The reason the explainer format persists despite not working is that it is pleasant to produce and pleasant to receive. Writing one feels like teaching. Reading one feels like learning. Both feelings are real and neither is evidence that anything happened. Whereas the format that works is unpleasant at both ends: the writer has to spend a third of the piece making the other side’s argument, which is uncomfortable and looks like weakness to a colleague reading the draft, and the reader has to sit through a passage that feels like it is going the wrong way.

There is an organisational reason too, which is that explainers are safe. Nobody has ever been reprimanded for a clear, accurate, thorough explanation of how something works. Stating the opposing position well, in a document published under an institution’s name, can absolutely get somebody reprimanded, because it looks from a distance like agreement, and the distance is where the reprimand comes from.

None of this is an argument against explanation as such. A person who genuinely has a question deserves a clear answer and the world needs far more of them than it has. It is an argument against reaching for explanation as a default when the actual problem is that somebody has already decided. Those two situations require different writing, they are easy to distinguish if you look, and the tell is simple: an audience with a question asks how, and an audience with a position asks whether you would say that.

Find out what they think first. Then decide what you are writing.

Every hand-painted board on a street in Amdavad was made by somebody solving a problem that most people writing today never quite face, which is that the reader is moving, will pass once, and owes you nothing at all.

The painter knew the distance the reader would be standing at, because it was the width of the road. He knew the speed, because it was a scooter in traffic. He knew where the sun came from in the morning, and that the board would be seen against it. He knew that the shopkeeper who commissioned it would judge the work by whether people walked in, which is a harder metric than anything most agencies report. And he had one attempt, with no test group, no second version, and no analytics, for a piece of communication that would stay up for fifteen years.

Everything about the form follows from those constraints and none of it is decoration. The letters are heavy because weight survives distance. The spacing is generous because a moving eye loses tight spacing first. The colour choices are extreme because subtlety at forty feet is invisible. The name of the shop is enormous and the list of what it sells is small, because the second piece of information only matters to somebody who has already registered the first. That is a hierarchy decision made under a hard constraint, and it is more disciplined than what most websites do with unlimited space.

The tradition has been documented properly, which is unusual for vernacular design in India. Hanif Kureshi, who was born in Palitana in 1982 and died in 2024, started HandpaintedType in 2011 specifically to record the typefaces of roadside painters, digitise them, and route proceeds back to artisans who were losing their livelihoods to flex printing. He went on to co-found St+art India, which is how a lot of people know the name, but the typeface project is the more interesting one, because it made an argument that almost nobody was making: that these letterforms were authored, that the authors had names, and that the work was design rather than folk atmosphere.

The reason that argument mattered is economic as much as cultural. A form treated as anonymous is a form available to everybody for free. The moment the painter’s name is attached to the letterform, the relationship changes from collection to citation, and citation is the difference between an archive and an extraction. The same distinction is being fought over right now, at vastly greater scale, in every conversation about what generative systems learned from and what they owe to it, and the sign painters got there first because they were the first to be replaced.

What replaced them is worth looking at closely, because the usual account is that digital printing was cheaper and that is why paint lost. Cheaper is true and incomplete. Flex printing also removed a constraint, and removing the constraint destroyed the discipline. A printed board can hold as much text as you like at whatever size you like, so it does, and the result is the visual noise of any Indian commercial street now: boards carrying nine pieces of information, each one legible only if you stop, which nobody does. The hand-painted board was better not because the painter had more taste but because his medium charged him for every letter, and a medium that charges you for every letter produces editing.

That is the transferable idea and it generalises well past signage. Constraint produces hierarchy. When a form stops charging for length, the hierarchy goes, and what replaces it is an accumulation of individually reasonable additions, each one requested by somebody, none of them refused, until the thing communicates nothing. Every long launch post and every crowded landing page is a flex board.

The painters also understood something about permanence that is almost entirely absent from digital practice. Fifteen years on a wall means you cannot write anything you will want to take back, which rules out most of what passes for positioning, because most positioning is written to be true for a quarter. A claim built to survive fifteen years of the shopkeeper’s actual behaviour has to be narrow and it has to be about what the shop actually does, and that pressure produces sentences of a kind that no campaign cycle will ever produce.

The last thing, and the one worth sitting with: the painter accepted that the reader owed him nothing. The board had to work on somebody who was going somewhere else, who had not asked, and who would give it about a second and a half. Most writing today is produced as though the reader has already decided to care, which is the one thing a board can never assume and the one thing a landing page assumes constantly.

There is no Gujarati word for alignment in the sense this field uses the term, and there is no Hindi one either, and the same holds across most of the languages in which several hundred million people will eventually have to decide what they think about this technology. The conversation happens in borrowed English, and a borrowed term arrives with its lender’s assumptions still attached, which is a fact about language that everybody knows in the abstract and almost nobody prices into what they build.

What gets borrowed is not the definition. Definitions travel badly and get corrected quickly. What travels intact is the picture, and the picture does the arguing. Alignment brings a geometry: two things that should face the same direction, with the implication that the correct direction is settled and the only open question is mechanical. Anybody who wants to ask whose values, which is the actual question underneath, has to first dismantle a metaphor that the room has already accepted, and dismantling a metaphor in public makes you sound like you are avoiding the technical issue.

You can watch the same substitution happen with every other load-bearing term. Safety arrives carrying an industrial picture, of a factory floor and a set of procedures, which makes the question sound like one of process rather than one of power. Governance arrives from the boardroom. Guardrails arrive from a highway and quietly settle that the trip is happening. Each of these was chosen in English, by a small number of people, for reasons that made sense at the time, and each is now installed untranslated in policy conversations in dozens of countries where the underlying picture has no local history at all.

The cost is not that people misunderstand. The people involved are sophisticated and they understand perfectly. The cost is that the range of questions which feel askable has been set somewhere else. A regulator working in a borrowed vocabulary can ask any question the vocabulary supports and will find it awkward to ask anything it does not, and the questions it does not support are precisely the ones that would arise naturally from a different starting point: questions about who bears the downside, about what happens to people who cannot read the interface, about whether a system that works beautifully in a metro is doing anything at all in a district town.

There is a temptation to fix this by coining words, and coinage mostly fails. Terms that get invented by committees for the purpose of having a local term have a recognisable deadness to them and practitioners route around them within a year. What actually works is slower and more interesting: the concept gets renamed by people using it in the course of doing something, and the new name sticks because it was needed rather than commissioned. That has happened before in Indian technical life repeatedly, and the resulting vocabulary is generally better than either the English original or the official translation, because it was shaped by people describing a thing they were in the middle of.

Which means the useful intervention is not a glossary. It is making sure the people who will do the describing are in the room early, and specifically that they are building rather than translating. A term coined by somebody deploying a benefits system in Gujarati will carry the deployment inside it. A term coined by a translator working from an English policy document will carry the English document.

There is a second-order effect worth naming, because it shapes what gets built. When the vocabulary is foreign, the evaluation is foreign too. A team assessing whether a system is working reaches for the measures that exist, and the measures that exist were built against English tasks with English assumptions about what a task even is. The result is a quiet and almost invisible bias in what counts as working, one that no individual decision produced and that nobody is in a position to correct, because correcting it would require a vocabulary that has not been built.

None of this is an argument for linguistic purism, which is generally a waste of everybody’s time and frequently a cover for something worse. English is the working language of the field and will remain so. The argument is narrower: that a field whose entire conceptual vocabulary originates in one place has one set of instincts, that it will experience those instincts as common sense rather than as a position, and that the first people to notice this are usually the ones working in a second language, because they can feel the seam.

Ahmedabad was inscribed as India’s first UNESCO World Heritage City in 2017, which was reported in the register of civic triumphs, and the interesting part began immediately afterwards and has been reported much less. A designation of that kind is not a reward for a condition. It is an argument about which version of a place is the real one, made in public, with obligations attached, and the obligations are where the argument gets tested.

The case that won was built on the walled city: the pols, the dense residential clusters with their shared courtyards and bird feeders and narrow entrances, an urban form that developed over six centuries and has almost no equivalent still functioning anywhere. That is a genuinely extraordinary thing and the case was correct. What the case could not do was settle what follows. A pol is not a monument. It is housing, occupied by people with jobs and plumbing and opinions, and the difference between a monument and a neighbourhood is that somebody has to live in the second one.

This is the fault line under every heritage designation and it is fundamentally a communication problem dressed as a conservation problem. The conservation question is technical and has answers: which fabric matters, what interventions are reversible, how you stabilise a timber frame. The communication question has no technical answer at all: why should a family that has waited fifteen years for a bathroom accept a slower, costlier renovation because a committee in Paris finds their house significant. Put that way, the honest answer is that they should not, unless the designation is actually delivering them something, and for large stretches of the last decade it has not been obvious that it was.

UNESCO has since raised concerns about conservation lapses, which triggered the predictable cycle: a flurry of committee activity, a set of commitments, and coverage framed around the risk of losing the status. That framing is revealing, because it treats the title as the asset. The title is not the asset. The title is a claim about the asset, and a city that spends its energy protecting the claim rather than the thing has quietly inverted the relationship, which is the single most common failure in positioning of any kind, at any scale.

What would a serious version look like. It would start from the observation that a living heritage district survives only if living there is better than the alternative, which is a statement about services and money rather than about aesthetics. It would make the trade explicit and legible to residents: here is what you give up, here is what you get, here is who to call when the roof goes. And it would treat the residents as the primary audience for all of its communication, rather than as a constituency to be managed while the real communication is aimed at tourists and committees.

That last point is where most heritage programmes go wrong and it generalises well beyond cities. When a project has one audience that can end it and another audience that makes it look good, almost all of the writing ends up aimed at the second, because the second is the one that produces visible wins. Brochures, exhibitions, festivals, a well-designed website in English. Meanwhile the people who actually determine whether the thing survives receive a notice in the mail about a permit process.

There is a version of this that is working, and it is worth being specific about what it consists of, because the successful examples are never about better storytelling. They are about somebody making the maintenance economics work: a revolving fund, a trained local workforce, a simplified approval that takes weeks instead of years. Once those exist, the storytelling becomes almost trivially easy, because the story is true and the people in it will tell it themselves. Before they exist, no amount of communication helps, and communication deployed in their absence is correctly read as public relations.

The thing I would want anybody thinking about this to sit with is that the 2017 designation was, in narrative terms, an enormous success and, in operational terms, a starting gun that a lot of people heard as a finish line. That confusion is very common and it is the specific hazard of winning an argument in public. A claim accepted feels like a problem solved. It is the opposite: an accepted claim is a debt, denominated in everything you now have to do to keep it true, and the interest comes due in a decade when somebody checks.

Ask an image model for a shop sign in English and it will hand you dozens of typographic registers: a condensed grotesque, a slab, a script with a competent connecting stroke, letterforms that look like they were drawn by somebody who had seen letterforms. Ask for the same sign in Gujarati or Tamil or Odia and what comes back is broken. The letters are approximately the right shape and wrong in the details that matter, the conjuncts collapse, the matras attach at angles no hand would produce, and the range of styles narrows to one, which is a sort of generic institutional look with no history behind it.

Both of these outputs are reported by most evaluations as the same capability: the model can render text in images. That equivalence is the interesting part, because it is not an oversight by careless people. It follows directly from who writes evaluations and which languages they can read. A grader who does not read Gujarati cannot distinguish a good Gujarati letterform from a plausible one, so the benchmark measures whether text appeared, and text appeared.

This is starting to be addressed, and the work is coming from where you would expect. AI4Bharat has been building Indic datasets and evaluations for years, BharatGen has been releasing open datasets for Indian language AI, and there is now a small but growing set of benchmarks specifically constructed for low-resource Indian languages rather than translated from English ones. The distinction between constructed and translated is the whole thing. A translated benchmark asks whether the model can do an English task in another language, which measures translation. A constructed one asks whether it can do the task that actually arises in that language, which is a different and much harder question, and it is the only one that tells you anything about deployment.

For scripts, the gap is sharper than for text generally, because typography encodes cultural information at a level below meaning. A Gujarati signboard painted in Amdavad in 1970 and one painted in 1995 are different objects to anybody who grew up reading them, in the way that a Victorian shopfront and a 1980s one are different to an English eye. That difference is legible to millions of people and invisible to every current evaluation, which means a model can be shipped as fully capable in a language while producing letterforms that read, to native speakers, as uncanny.

The reason this deserves more attention than it gets is that it is a clean case of a general problem. Evaluation determines what counts as done. When the people who build evaluations and the people who will live with the system are different populations, the system will be declared finished at the point where it satisfies the first group. Everything downstream inherits that judgment: procurement decisions, product roadmaps, the allocation of engineering time. No individual acted badly. The measurement simply had a blind spot the shape of an entire readership.

Fixing it is less exotic than it sounds. You need a corpus of real letterforms with provenance, which in most Indian languages means going out and recording what is actually painted on walls, with the painter’s name and the year, because the alternative corpus, digital fonts, is small and homogeneous and does not represent how the script is actually used. You need graders who read the script and are asked specific questions rather than general ones: is this letterform correct, is this conjunct formed properly, would a reader find this style plausible for this kind of business. And you need a rubric written tightly enough that two graders agree, which in practice takes several rewrites, because every ambiguity in the wording comes back as noise in the results.

That last part is where these projects usually die, and it is worth saying plainly that the bottleneck is writing rather than modelling. Anybody can collect images. The difficulty is specifying, in language that survives being read by thirty different people, what you want them to look at and what to do when the letterform is legible and wrong at the same time. A vague rubric produces data that looks like a benchmark and measures agreement rather than quality.

There is a decent argument that all of this will be solved incidentally as models improve, and it may be. But incidental solutions arrive in an order determined by who complains, and a capability gap that nobody can see will not generate complaints. The scripts that get fixed first will be the ones with a benchmark, because a benchmark is the mechanism by which a problem becomes visible to people who cannot perceive it directly. Everything else stays in the category of things that work fine, according to the only people who were asked.

A financial centre is one of the few things in the world that is made almost entirely out of belief. The buildings matter and the fibre matters and the tax treatment matters a great deal, but none of them produces a financial centre by themselves, because what a financial centre actually consists of is a large number of people who each believe that the other people will be there next year. That is a coordination problem, and coordination problems are solved by description: everybody needs a shared account of what the place is, or nobody moves first.

GIFT City, on the Sabarmati between Ahmedabad and Gandhinagar, is the most interesting current test of whether a jurisdiction can be built this way deliberately. It has the regulatory apparatus, a dedicated authority, an exchange, a growing population of funds and global capability centres, and a policy environment that keeps adding reasons to be there. What it has been working on for a decade, with genuinely mixed results, is the description.

The difficulty is structural rather than promotional. A new centre has to be described in terms of the old ones, because that is the only vocabulary its audience has, and every available comparison imports a set of expectations it cannot yet meet. Call it India’s Singapore and you have invited a comparison on depth of talent pool and decades of institutional history. Call it a tax jurisdiction and you have positioned it as arbitrage, which attracts exactly the capital that leaves first. Call it a gateway, which is the term that has mostly stuck, and you have said something true and almost contentless, because a gateway is defined by what passes through it rather than by anything it is.

This is the same problem any new category faces and it has one reliable solution, which is to describe the specific thing you are better at rather than the general thing you resemble. Not a gateway, but the place where a particular kind of transaction that used to require three jurisdictions now requires one. That sentence is checkable. A person can hold it up against their own situation and see whether it applies, and the ones for whom it applies will move, and their moving is the evidence that makes the next sentence possible.

The temptation in the other direction is enormous and almost irresistible for any government project, because the audience for the communication is frequently not the market but the political centre, and the political centre rewards scale language: number of entities registered, square feet, projected employment. Those figures are real and they are the wrong instrument, because a fund manager deciding where to domicile is not asking how big the place is. She is asking whether the thing she needs to do is easier here than where she is now, and how quickly she can find out, and what happens when something goes wrong at eleven at night.

That last question is the one that decides these things and it is the one nobody writes about. The reputation of any financial jurisdiction is made almost entirely out of stories about failure handling: how a dispute went, how long an approval actually took when it was complicated rather than standard, whether the regulator picked up the phone. Those stories circulate privately, they are extremely durable, and no amount of published material can override them. Which means the communication that matters most is not communication at all. It is the operational behaviour that generates the stories, and the only useful thing a communications function can do is make sure the good stories are findable by the people who are currently asking around.

There is a version of this argument that says places cannot be positioned, that they simply grow or fail to, and I think that is too fatalistic. Places do get positioned, and the successful ones tend to have done a particular thing: they found the narrow use case where they were unambiguously better, said it plainly for a decade without getting bored, and let the broader claim arrive later, from other people. Dublin did that with fund administration. Luxembourg did it with cross-border vehicles. Neither started by announcing that it was a global hub, and both ended up being described that way by others, which is the only version of that sentence worth having.

The general lesson is unglamorous and applies far outside finance. You do not get to make the meaning claim first. You get to make the structural claim, then the evidence claim, and only then does somebody else make the meaning claim on your behalf. A place that describes itself as a global hub before anybody else has said so is doing the sequence backwards, and the audience can hear it, because the sequence is audible.

The obvious way to describe a large model’s failure is to say it does not have enough data. Ask an image model for a shop sign in Hindi and the letterforms come back broken. Ask for the same sign in English and you get ten typographies, each of them competent. The gap is real and it is easy to measure. The temptation is to treat the fix as arithmetic: pour more of the missing script into the training set and the gap closes. This is true and it is insufficient, and the reason it is insufficient is the interesting part.

A training corpus is not an ingredient list. It is a legislature that has already voted, and by the time a model reaches you the votes are on things you cannot see. Which words appear next to which. Which sentences appear in the same document. Which images the captions treated as worth captioning. Which categories the raters recognised. Every one of those decisions is a small constitutional act, and by the time the training run finishes you have a system whose imagination is bounded by the reading habits of the people who assembled its diet. The word bias suggests a wrong opinion laid over a correct machine. What is happening here is different. It is a prior about what exists.

The clearest place to see this is in scripts the corpus barely met. A model trained mostly on English does not merely know less about Gujarati; it knows a different Gujarati. It has met the language through translations, through captions of Wikipedia images, through the small share of Indian websites that were content-farmed into English SEO before anyone crawled them. What it has not met is the actual daily surface of the script: shop boards, wedding invitations, movie posters, courtroom forms, the register a mother uses to text her son, the register the son uses back. Ask it to make a shop sign and it produces something that looks like an attempt to remember a script rather than an attempt to use one, because the training taught it that this script mostly appeared as an object to be described, not as a thing to be written.

This has practical consequences that get missed because they hide inside a familiar frame. The frame is: the model is worse at Hindi than at English, so we should train more Hindi. The consequence hiding inside is: the model has inherited a set of readers whose interests decided which Hindi mattered. Those readers were disproportionately literary, disproportionately governmental, and disproportionately based in Delhi. A company deploying a chatbot into Ahmedabad, or Coimbatore, or Guwahati is not deploying a model that knows less about its market. It is deploying a model whose imagination of that market was formed by people who never went. Adding data helps only if the data is chosen by somebody who has been.

Which brings me to the second act of this argument, which is about who chooses. In a field this young, most corpora were assembled by whichever team could scrape fastest, filtered by whatever quality signals the labellers could agree on, and rated by workers paid by the hour to answer questions written in English. There was no meeting where anyone said this document is worth keeping and this one is not; there was a pipeline, and the pipeline had defaults, and the defaults were English. This is the ordinary way large systems get built and it works fine in domains where the default reader is the marginal reader. It stops working the moment you take the system across the language boundary. The pipeline still runs, the metrics still look reasonable, and the model behaves in a way the numbers do not predict, because the numbers were built by the same defaults that built the corpus.

For a person deploying one of these models into a real market, this suggests a working rule more useful than the debate about which lab is winning. The model is a jurisdiction. It will govern your product for as long as the product exists, and its constitution was written by people whose interests were not yours. That is not an argument against using it, any more than the fact that a bank is governed by American law is an argument for an Indian founder to avoid opening one. It is an argument for reading the constitution before signing the contract, and there is a small number of questions that get you most of the way there.

The first question is what the model was mostly trained on, by document type rather than by language. A model that read the internet is not a model that read your industry. A model that read scientific papers has different priors than one that read chat logs. The lab will not always tell you the answer; the model will, if you hand it a domain document and watch what it flatters and what it flattens.

The second question is who rated the outputs during fine-tuning. A model trained to please a graduate student in San Francisco is a different animal from a model trained to please a call-centre trainer in Manila, and both are useful, and neither is neutral. If you are building for a third audience you are asking the model to play a role it was not cast for, and you will pay for the mismatch in ways that show up as low retention rather than as anything a benchmark catches.

The third question is what happens to the model when it is asked to work in a script or a register the corpus barely met. This is the question that matters most for anybody selling into markets where English is a second language and where a model that reads English as its native tongue will produce copy that reads as slightly foreign for reasons neither the writer nor the reader can name. That foreignness compounds. Every sentence is a small failure of belonging, and the audience feels it before they can say what is wrong.

None of this is an argument for training your own model, which is a solution most companies cannot afford and few actually need. It is an argument for reading the constitution of the model you did buy, and for accepting that the reading is part of the work, not a preamble to it. The people who will build durable products on top of this generation of systems are the ones who understand which questions the systems were built to answer well, which questions they were built to answer at all, and which questions the systems were quietly excused from answering because nobody in the room spoke the language of the person who would eventually ask them.

I asked what the company did. The founder said it made tomorrow investable.

The office had no chairs because the team was still pre-seating. A graph on the wall rose to the right and continued into the neighbouring building.

I asked about the product. The founder described a confidence layer for emerging behaviour. Its purpose was to turn people into the kind of customers who would eventually want the product. The company had no customers yet, but the pitch deck called the gap temporal liquidity.

I asked for a user count. The founder declined. The company was careful with numbers at that stage. Its tote bags carried a projected logo and were made from a material that would become sustainable when the supply chain existed.

The seed round closed the following Thursday. The lead investor called the company early, ambitious, and difficult to disprove. The present had become a temporary misunderstanding.

I asked why the marketplace existed. The founder said the market was waiting. He added that waiting was a form of demand.

The company sold subscriptions to objects that had not yet been emotionally discovered. Each month, subscribers received a photograph of a wooden object, a note about its use, and a button marked “Declare Interest.”

The first object was a shallow bowl for holding unfinished thoughts. Twelve people declared interest. None could explain what the bowl did, but all twelve said the photograph made them feel understood.

I asked whether the company manufactured the objects. The founder said it would, eventually. Until then, it would remain asset-light. The pitch deck called the business the infrastructure between longing and logistics.

The company had not sold an object. Its dashboard tracked “Intent Density,” a measure based on clicks, saved photographs, and the time visitors spent considering the bowl. The graph showed no revenue and a very confident line.

I asked about the company culture. The founder said it was customisable. Her laptop showed four settings: Radical Candour, Quiet Excellence, Founder Mode, and Other. A warning appeared beside Other: “This may reduce alignment.”

The company had started as a scheduling tool and become a workplace philosophy between its Series A and the opening of a second office. The founder said people were different and the company wanted to meet them where they were. The location field on the company dashboard said “In the funnel.”

Each employee received a values card on the first day. A QR code on the card opened a survey asking whether the card had made the employee feel welcome.

Ninety-four percent of employees said the company was listening. The dashboard classified the other six percent as a growth opportunity because they had not completed the survey.

The founder asked which value the company should prioritise that quarter. I chose Humility. She thanked me and selected Speed.

I asked when the founder knew the company was over. She said it was when people started using the word product.

The company had helped people imagine better lives, but the team eventually asked what the imagined lives could do. By then, it had a name, a waitlist, and three rounds of press coverage. It did not have a working button.

The founder described the product as a new category of possibility. Users described it as a newsletter. The waitlist contained 8,400 names, although the sign-up form had not asked anyone to verify an email address.

The shutdown note thanked the community for believing in the work and said the company was making room for what came next. It did not mention the unpaid invoices, the empty calendar, or the unfinished login page.

I asked what she had learned. She said a company could be real before it became useful. Then she corrected herself: a company could be memorable before it became useful. The company had achieved that much.