Plutonic Rainbows

Google I/O used to have a recognisable shape. Android first, a few developer APIs, a hardware tease, then the bit where Search got a little stranger. This year the structure collapsed into Gemini. Not because Google forgot the rest of the company exists, but because the rest of the company now seems to exist as places where Gemini can be put to work.

The cleanest announcement is Gemini 3.5, Google's new model family, which starts with 3.5 Flash. Google says Flash is available now in the Gemini app, AI Mode in Search, Google Antigravity, AI Studio, Android Studio, Vertex-adjacent enterprise products, and the Gemini Enterprise Agent Platform. It is also the default model for the Gemini app and AI Mode globally. That is not a laboratory release. It is Google pushing a new base layer under its consumer and developer surfaces on day one.

The numbers are pitched with the usual violence. Google says 3.5 Flash is its strongest agentic and coding model yet, beats Gemini 3.1 Pro on several agent and coding benchmarks, and runs four times faster than other frontier models when measured by output tokens per second. I don't much like writing benchmark paragraphs because they age badly, sometimes by lunchtime, but the direction matters. In February I wrote about Gemini Deep Think as Google spending inference-time compute on hard reasoning. This release pulls the other way: make the everyday model fast enough and good enough that agentic work stops feeling like a special mode.

Then there is Gemini Omni, which is the showier thing and maybe the more revealing one. Omni takes text, images, audio, and video as input and starts by producing video as output. Google says the first model, Gemini Omni Flash, is rolling out to the Gemini app, Google Flow, and YouTube Shorts, with image and audio output planned later. The claim is not only better video. The claim is continuity: edit through conversation, keep characters stable, make the physics hold together, let one instruction build on the last.

That is where Google's advantage looks least abstract. A video model inside a chatbot is one product. A video model inside YouTube, Flow, Android, Search, and whatever Google decides Chrome should become next is something else. AP reported that the Gemini app has passed 900 million monthly active users, more than double the previous year. Even allowing for the fuzziness of app metrics, that is distribution almost no AI-native company can touch.

The less glamorous Android and Chrome announcements say the same thing in a more domestic register. Google says Gemini Intelligence on Android will automate multi-step tasks, fill forms with opt-in personal context, turn spoken mess into cleaner messages, and build custom widgets from natural language. Chrome on Android is getting Gemini summaries, app actions, and auto browse for chores like booking parking or updating an order, with confirmations before sensitive actions. I can feel the pitch hardening as I type it: Google doesn't want one agent. It wants every surface to contain an agent small enough to disappear into the verb you were already using.

I still don't know whether people want that much help. Some of this will be useful in the dullest possible way, which is usually the way software wins. Some of it will be exhausting, another layer of anticipatory cleverness between a person and a task they could have finished with three taps. But the strategic point is clear enough. OpenAI sells the interface. Anthropic sells the model and the caution. Google sells the operating environment in which the model is already waiting.

Most British towns had at least one. A wooden case with a sloped glass front, screwed to the brick outside the library or the post office or, in the small towns, the launderette. Inside, a square of cork or hessian half a centimetre too small for the frame, so the edges showed where the fabric had faded a different shade. Drawing pins. Index cards. Felt-tipped writing on the back of cereal boxes. Polaroid snapshots of missing cats curled at the corners from the damp.

The thing held the bounded life of a few streets, and it held it for as long as the pins kept their grip.

A noticeboard was not a classified column. The classifieds were typeset and counted by the line and paid for at a front desk; they aggregated upward into a paper that circulated across a county. The board outside the launderette aggregated nowhere. It was visible to perhaps two hundred people a week, all of whom lived within walking distance, and it carried things the paper would not have bothered to print. A piano teacher. A spare room, women only, references essential. A plea, in shaky capitals, for the return of a tortoise named Stanley. Knitting circle, Tuesdays, church hall, bring your own needles.

The content varied by the host. The library board ran to evening classes and political meetings, with the occasional yellowing leaflet about a footpath inquiry no one remembered initiating. The post office favoured laminated official notices about pension adjustments, surrounded by handwritten cards offering ironing services. The launderette tended toward the desperate: lost dogs, room shares, secondhand pushchairs. A board could acquire a tone the way a pub does, and locals knew which one to read for what.

Nothing on these boards was archived. Not by the library, not by the council, not by anyone. The cards came down when they got wet, or when the next pin needed the space, or when the cleaner decided the board was getting cluttered. Some cards stayed up for years because nobody wanted to be the one to remove the tortoise. Some came down within days. The board maintained itself through a slow informal consensus that left no record of what it had looked like the week before.

This is the part that interests me. A village in 1986 had a working memory that existed only in the present tense. You could not consult the noticeboard of last March. You could only consult the noticeboard of now, and the noticeboard of now contained whatever had survived since you last looked. The full texture of who in your street needed what, who was teaching what, who had lost what, was readable only by being there, and was unreadable forever the moment you weren't.

Replacements exist. The local Facebook group does roughly the same work, with infinitely more reach and an indelible archive of every cruel comment ever made about somebody's missing cat. Nextdoor performs a caricature of it. Both are searchable, both persist, both make the original noticeboard look quaint by comparison. Neither produces the same object, because the object's defining property was that it would not last. The noticeboard told you what your neighbours needed this week, and accepted that next week would need its own.

The wooden cases mostly survive. I see them still on village halls and outside the older library branches, their cork half empty, holding a single bus timetable and a poster for a flower show that happened in 2019. The pins are still there. The hands that used to put things up have moved indoors, onto screens, into feeds that never quite forget anything. The board waits anyway. A piece of public furniture for a kind of public attention that no longer collects in that shape.

Helmut Lang showed his Spring/Summer 1992 ready-to-wear in Paris in October 1992, and on the evidence of the runway images preserved by Vogue, the collection looked nothing like the brand most people now remember him for. Bolos. Chaps. Earth tones. Body paint. Quill-like breastplates that read as half armour, half costume. Beaded snakes wound around briefs. A Wild West reference no one would later attach to the word "Lang." It is the sort of show you would expect from a designer still finding his idiom, which by 1992 Lang nearly was. The flat-fronted tailoring and the sheer, layered cotton came shortly after, and they came with someone else in the room.

Melanie Ward, by 1992, was already a problem for the industry's hierarchies. She had spent the previous two years producing a body of work in The Face and i-D with the photographers Corinne Day and David Sims, a visual language stripped of styling tricks and built from secondhand clothes, sneakers paired with everything, and models who looked like they had just got off the bus. In June 1992 The Face put Ward and Day on its cover under the banner "Young Style Rebels, London's New Model Army," and the underground phase of what would soon be called grunge ended on that issue. By the autumn Vogue and Harper's Bazaar had begun adopting a softer version. Calvin Klein had hired her for the Kate Moss jeans and underwear campaigns shot by Sims. And Lang, watching from Vienna, had begun the conversation that would last thirteen years.

The partnership ran from 1992 to 2005, the year Lang retired. Across that period Ward is credited, by people who were in the rooms, with shaping the casting, the styling logic, and a great deal of the brand's public image. StyleZeitgeist's obituary specifies that her hand first showed clearly on the Spring/Summer 1994 catwalk, where she shook up Lang's casting with a London cool the Paris ready-to-wear circuit was not yet used to. By the mid-90s the language we now think of as Helmut Lang, sharp tailoring, sensual layered T-shirts, transparent organza and nylon over bare skin, was fully present. By the late 90s the language had spread outward into everything Ward also touched: the Klein campaigns, the Jil Sander adjacencies, Harper's Bazaar under Liz Tilberis.

What I find worth recording is the gap. October 1992 shows you a designer doing one thing. By 1994 he is doing another, and the second thing is the one that becomes the decade's default register for serious clothes. The brand that everyone copies is the brand after the meeting, not before. And the meeting, which lives in old interviews and a small number of magazine archives, is essentially undocumented in the way a hire would be documented now. There is no LinkedIn post. There is no announcement. There is a stylist who started doing the casting, and a designer who let her, and a partnership that read across the work for years before anyone outside the industry knew her name.

The other thing worth recording is how much of this period exists only in print. Ward herself told System Magazine that the first ten years of her work were not extensively digitised, that they survive in a small number of bound back issues and the imagination of people who happened to be paying attention to magazines at the time. Lang's pre-New York shows are partly online now, because Vogue Runway has been backfilling them, but the backfill is recent and patchy. A reader under thirty mostly cannot see what made the 1990s look the way they did. The cleanest visual evidence of the most influential stylist of the decade is locked inside paper that nobody scans.

There is a temptation in writing about this period to make the partnership sound inevitable. It was not. Lang could have stayed in the Wild West register. Ward could have taken a different magazine job, or signed exclusively with Klein. The thing that distinguishes 1992 is that neither happened. Two specific people, both already distinctive, found a shared idiom in time to define a decade. The October show was the last one before the shared idiom arrived. It is worth looking at in that light.

OpenAI has put Codex inside the ChatGPT mobile app, in preview, on iOS and Android, on every plan including Free and Go. The announcement landed this week and the rollout was already underway by the weekend. Windows users still have to wait, but for everyone else the coding agent now lives a tap away on the phone you already carry.

The framing in OpenAI's own messaging is interesting. It is not "write code on your phone." Nobody wants that, and the screen is the wrong shape for it anyway. What you actually get is a remote viewport into a Codex session running somewhere else: a desktop, a devbox, a server. The phone shows screenshots, terminal output, the agent's progress on a task it was already chewing through. You can answer a clarifying question, approve the next step, nudge it in a direction, or tell it to back off. The heavy lifting stays on the machine where your credentials and your local setup already live.

There is a secure relay layer in the middle, which is the boring detail that makes the rest of it work. Your laptop is not suddenly exposed to the public internet because you wanted to check on a bug from the supermarket queue. The relay is the kind of plumbing that gets invented once and then disappears into the background of everything else.

I keep thinking about what this changes about the rhythm of the work. The older model was: open the laptop, sit down, type for an hour, close the laptop. The intermediate model, where agents took on long-running tasks, already pushed against that rhythm. You'd kick off a refactor before lunch and come back to a pile of proposed diffs. Now that pile can find you on the bus. You can glance at a screenshot of a failing test and tell the agent to try the obvious fix while you keep walking. The latency between "agent needs a decision" and "decision arrives" collapses from hours to whatever it takes you to read a notification.

The free tier matters here in a way that I don't think OpenAI is quite shouting about. There are around four million weekly Codex users by the company's count, and the gating to date has been both subscription and surface area. Putting the surface on the phone, with no paywall for at least basic use, is the kind of move that grows habits more than it grows revenue. Habit is the thing every AI lab is actually competing for at this point. Paid plans follow habit; habit does not follow paid plans.

There is a counter-reading. Putting an autonomous coding agent within easy reach of every distracted moment is also a way of extending work into the gaps that used to be rest. The agent does not care that you are at the grocery store. It only cares that you have a phone and a connection. The same convenience that collapses the latency on a real decision also invites you to make half-formed ones while your attention is elsewhere. We will see what that does to the shape of code review and to the people doing it.

For now, the immediate effect is small and concrete: an update to the ChatGPT mobile app and the macOS Codex app, and a tab on your phone that wasn't there last week. Everything that follows from that is downstream.

Royal Mail's National Returns Centre has been in Belfast since 1992. A warehouse on the edge of the city is where everything that cannot be delivered in the United Kingdom eventually ends up, and the rolling estimate kept by the people who work there sits at around twenty million items. Wedding photographs, love letters, autographs, a Picasso sketch valued at three thousand pounds, expensive gifts the recipient never knew were on the way. The site, when The Guardian visited in 2003, was running on roughly ten million pounds a year and the volume was rising every year despite the arrival of email and text-messaging. Whatever the internet was doing to the postal service, it was not reducing the supply of mail that could not be made to land.

The building's job is to read what the rest of the system gave up on. An envelope arrives at Belfast because the address is incomplete, or the handwriting illegible, or because the recipient has moved and the sender wrote no return on the back flap. A team opens it, which is one of the very few places in British law where the implicit right to secrecy of correspondence can be set aside, and they look for clues. A name on a card inside. A return address on a cheque. A previous postmark. If they find nothing usable within roughly a month, the letter is destroyed; if the contents have any auction value, they are sold. The wedding photograph of the man in RAF uniform and his shy new bride that the Mirror ran in 2013, both faces still legible after seventy years, was found in a stack of unclaimed envelopes that nobody would ever see again outside this one room.

The British Post Office set up its first Dead Letter Office in London in 1784, and Edinburgh and Dublin opened similar offices shortly after, as the Postal Museum's archive guide records. By 1813 the system had been formalised as the Returned Letter Office, which sounded less morbid and was also more accurate, as the goal was always to put the letter back into the sender's hand rather than to bury it. The Victorian institution kept photographs of what it received. One image from around 1900 shows the room with dead pheasants, rabbits and hares hanging from the ceiling, sent as game gifts that never reached their country-house addressees and ripened in transit until the sorters had to do something. The renaming to Returned Letter was a public-relations decision. The work was the same.

What I find unsettling is that the institution is older than the post code, older than the typed address, older than the photograph itself, and it has survived every revolution in addressing without changing its basic function. Email did not shrink it. Two decades of broadband did not shrink it. The parcel boom of the 2010s, with its printed barcodes that should have been undeliverable-proof, did not shrink it. As long as human beings are willing to write a name and a town on the outside of a package and entrust the result to a national system, a fraction of those names and towns will be wrong, or the recipient will have moved, or the sender will have failed to include themselves on the envelope. The fraction is small. Multiplied across a country, it is twenty million.

The Belfast warehouse is the closest thing the United Kingdom has to a national memory of failed correspondence. It contains, at any given moment, a thousand wedding albums that the couple never saw, a thousand last letters from people who have since died, a thousand cheques nobody will ever cash. The room has no public catalogue and no public access. It does not appear on the Royal Mail tracking page. The most modern country in Europe runs an entirely analogue archive of its own undelivered intentions, and the only reason most of us never think about it is that we ourselves were lucky and our envelopes happened to land.

There was a moment, roughly between 1983 and 1987, when a serious AI researcher worked at a Symbolics. Not on one. At one. The machine was the workstation, the operating system, the editor, the debugger, the version control, the inspector, and the language all at once, and the language was Lisp from the microcode up. The LM-2 in 1981 was the first commercial repackaging of MIT's CADR design, and by the time the 3600 line shipped you could buy a desk-sized Lisp environment with over five hundred thousand lines of system code written in the same language you were writing your own code in. Down to the kernel. That is still an extraordinary thing to think about.

Symbolics was a spinoff from the MIT AI Lab, one of two companies that came out of a hacker schism around the turn of the decade, the other being Lisp Machines Inc. Symbolics got most of the hackers and most of the money. For a few years the bet looked obviously right. The Defense Advanced Research Projects Agency had decided that expert systems were a route to military advantage, and the Strategic Defense Initiative meant DARPA was buying. Symbolics machines went into labs and skunkworks and Star Wars feasibility studies. Revenues climbed. The company signed long real estate leases in California on the assumption the run would continue.

It did not. The mechanism is worth being specific about, because the standard summary, "AI winter killed it", obscures more than it explains. Two things happened in parallel. The first was that expert systems, the thing Symbolics had been sold as the platform for, turned out to scale badly. Rules accumulated, interactions between rules produced surprises, knowledge engineers were expensive, and the systems could not learn. The second was that Sun Microsystems and the rest of the RISC workstation industry caught up on raw compute and then surpassed Symbolics on price-performance, while compiler people figured out how to make Lisp run respectably on general-purpose silicon. A purpose-built Lisp machine had been the only way to run Lisp fast; once it wasn't, the entire commercial premise dissolved.

DARPA noticed before the company did. Funding cuts started biting in 1987. Russell Noftsker, the co-founder running operations, ended up in a power struggle with Brian Sear, the CEO the board had hired in 1986, over whether to pivot to selling Lisp software that ran on Sun boxes or double down on Symbolics' superior hardware. The board fired both of them. Sales fell off a cliff, the California leases turned into millstones, and Symbolics filed for bankruptcy in early 1993. A small remnant company kept maintaining Open Genera, a virtual machine version of the OS, for the customers who could not afford to migrate off it. By 1995 the Lisp machine era was over.

The instructive bit is that no single decision killed Symbolics. The technology was genuinely impressive. The customers were real. The leadership had founder energy and MIT pedigree. What killed it was the assumption that a specialised hardware platform could remain the cheapest way to run a particular kind of software, in a decade when general-purpose silicon was doubling in capability roughly every eighteen months. The platform was a moat that was also a clock. Every Moore's Law tick narrowed it.

You can see the same shape now in the argument over whether GPU accelerators stay a category, or whether the workload that currently justifies them gets absorbed by whatever comes next. Specialised hardware is enormously valuable while the gap is wide, and a liability when it narrows. The companies that bet on the gap staying wide are the ones who sign the California leases. There is no indication this lesson generalises, and every indication it does.

The laminated A4 sheet appears on the lamp-post outside the corner shop sometime in the night, and within a week the rain has buckled the plastic pouch and turned the photocopy blue at the edges. The notice is addressed to nobody in particular and to every adult in the postcode at the same time. A six-line description of the proposal in the flat register of development-control prose, the address of the site, the place where the full application can be inspected, and a date by which any written representation must be received. Twenty-one days from the date of first display, which is the statutory minimum councils must allow before they can determine an application.

The strange part is that the notice is the building's own ghost made visible before the building has gone anywhere. The shopfront is still trading. The Victorian terrace still has its tenants. The car park still has cars in it. And yet the sheet on the pole, in its unread and weatherproofed way, is already describing the place in the past tense, in the dry syntax councils use for things that have not happened yet and might not. Erection of a three-storey block of nine self-contained flats with associated parking, cycle storage, refuse store and amenity space. The proposal has not been approved. It may never be approved. But the sheet has introduced the future as a piece of municipal furniture, and the present has become the version of the street that the proposal will replace.

You can stand and read one of these for as long as you like and nothing will be clearer at the end than at the beginning. The description is technically complete. It is not legible in the sense of giving you any image of what will happen. The language is engineered to be neutral enough to survive an appeal. Active verbs are converted into nouns. Heights are given in metres above ordnance datum. Materials are described as proposed elevations finished in stock brick and standing seam zinc. The thing the prose refuses to do, by training, is to picture itself. You have to take the drawings down from the council's portal and find a planner who likes you to learn what the notice actually means, and even then it is provisional until the committee meets.

Telegraph poles get a different layer of paper. The Electronic Communications Code gives a broadband operator the right to put up a wooden pole of up to fifteen metres with only twenty-eight days' prior notice to the local planning authority, and no statutory requirement to consult the residents whose front gardens the pole will face. Some councils ask operators to laminate a courtesy A4 to the pole that is being installed, naming the next pole, the company, and a date. There is nowhere on the sheet for the householder to register dissent that would stop the installation. The notice is not asking. It is announcing. The grammar of the planning notice has been carried, intact, into a regime where the consultation period is decorative.

What this leaves on the British street is a continuous low layer of paper futures pasted onto its furniture. Most of the notices belong to applications that will be granted, often with conditions, and turned into the buildings they describe. Some belong to applications that will be refused, withdrawn, or appealed to inspectors and won on the third attempt two years later. Some belong to operators who will put the pole up regardless of what the council says. A few belong to proposals nobody at the council can later remember receiving. The sheet is taken down when the application is determined, or stays up until the rain unmakes the plastic and the staple lets go, or both at once. The street has by then long since stopped reading the lamp-post and resumed its own business, and the next sheet appears two doors down, addressed to the same nobody.

I think the unease they produce is not about the proposals themselves, most of which are unobjectionable infill that any honest assessment of the housing shortage would welcome. It is about the address. The sheet is written to a public that does not assemble to read it. The twenty-one-day clock starts on a day no one notices, runs through a window of attention no one has been given, and closes before most of the people the proposal will affect have understood that a clock was running. The notice is a working piece of democratic infrastructure that has been engineered to look exactly like litter. That is the trick the pole has learned, and it learned it long before any of us were paying attention.

The bandage dress is almost always credited to Hervé Léger, and almost everyone who repeats that has it wrong. Léger's version, the painfully tight strapless thing that became a 2000s red carpet uniform, did not arrive on a runway until the early 1990s. The silhouette he is famous for had already been on working models in Paris for half a decade by then, knitted in soft stretchy viscose by a small Tunisian man who was a foot shorter than most of his clients and who pinned dresses directly onto their bodies in the small hours of the morning. Azzedine Alaïa's spring/summer 1986 ready-to-wear was the moment the form arrived complete: a black satin-knit halter bodysuit, open at the back, married to a high-waisted bandage pencil skirt with diagonal seams and a side lace-up. ReSee's archive listing of the original S/S 1986 bandage ensemble is one of the rare places you can still see the construction up close.

What I find interesting is what the dress was answering. The mid-eighties was the high water mark of power dressing, a silhouette of padded shoulders and structured jackets borrowed from menswear and inflated, designed to claim space in rooms that had recently begun to admit women. Thierry Mugler had been building those rooms architecturally, in patent leather and chrome. Alaïa pulled in the other direction. His answer was a body that was not armoured but held; not enlarged but articulated. The Egyptian-mummy references he gave to interviewers were not idle, the wrappings were meant to evoke preservation and protection rather than constraint. The French philosopher Michel Tournier later wrote that the dresses allowed women to be held as tightly as possible while remaining free, a sentence that sounds like marketing copy and is in fact closer to the truth of how the material behaved. The Alaïa knit gave back when you breathed. Léger's bandage, when it eventually arrived, did not.

By October 1985 the trajectory was already legible to anyone paying attention. Earlier that year he had been named creator of the year at the Oscars de la Mode, France's short-lived Jack Lang-era state fashion prize, and had walked onto the stage with Grace Jones, who was wearing one of his fuchsia bandage dresses and a dazzling smile. The press had already christened the look the sexy mummy. Bill Cunningham, reviewing the spring 1986 Paris shows for Details, would write that the season had been defined by the major houses shamelessly copying the hourglass silhouette of Azzedine Alaïa. He was, by any working measure, the most imitated designer in Paris that year, and the imitation continued long after he had moved on, eventually attaching to a different name entirely.

There is a small footnote here about how authorship gets allocated in fashion. Alaïa never operated like a brand. He showed off-calendar by the late 1980s, refused the press machine entirely, hand-stitched garments directly onto Naomi Campbell and Linda Evangelista in his Marais studio, and let buyers and editors travel to him on his schedule rather than theirs. Léger by contrast was a brand from the start, marketed aggressively, quickly licensed, and by the late nineties the bandage dress belonged to him in the public mind, the way the wheel belongs to whoever sells the most cars. The historian Olivier Saillard, who curates the Fondation Azzedine Alaïa archive, has called the appropriation a strange contradiction; I think it is simply what happens when one party treats the work as the point and the other treats the work as the asset.

The original ensemble still sits in ReSee's vintage inventory at six and a half thousand euros, hand-selected for the archive by Pieter Mulier, the designer who now runs the Alaïa house. The satin knit, forty years on, is reportedly still soft.

Alexander Liberman, the editorial director of Condé Nast through most of the second half of the twentieth century, asked Peter Lindbergh a question. Lindbergh had told him, in plain terms, that he could not stand the kind of woman American Vogue kept putting in front of him. Over-styled, heavily made up, surrounded by limousines and small dogs, supported by a rich husband. Liberman, who had heard this complaint before but never from someone he wanted to hire, asked Lindbergh to show him what he meant instead.

Lindbergh picked the stylist Carlyne Cerf de Dudzeele, who liked gold Rolexes and serious jewellery and was therefore the wrong person for the job on paper. He told her not to worry about the clothes. Just bring some white shirts. He picked six models he liked at the moment: Estelle Lefébure, Karen Alexander, Rachel Williams, Linda Evangelista, Tatjana Patitz, and Christy Turlington. None of them was famous yet. He flew the whole group to Los Angeles and put them on the beach at Santa Monica.

What he produced was almost nothing. Six women in oversized men's shirts, mostly unbuttoned, almost no makeup, hair left to do what the wind did. Bare feet. Underwear, where you could see it. Black and white. They were laughing in some frames and looking past the camera in others. There were no clothes to sell because the only clothes were borrowed shirts.

He brought the prints back to New York and laid them out on a table in front of Liberman and the then-editor Grace Mirabella. In his telling years later, the two of them looked at each other and said little. The pictures were not what anyone had asked for, because the brief had been so vague it could not have been. They ran in the August 1988 issue of American Vogue. Within eighteen months, Lindbergh was working with the same group in different combinations, sometimes adding Naomi Campbell, sometimes Cindy Crawford, until in January 1990 he shot the British Vogue cover that everybody now reaches for when they want to date the start of the supermodel era. The Versace finale in Milan eighteen months after that was the same idea staged with music.

The Santa Monica picture is the one that gets less credit because it ran before anyone knew what they were looking at. There is no front-page narrative attached to it, no music video, no runway moment. It is a magazine spread that happened to contain six women whose names did not yet matter, in a setting that contained nothing else, photographed by someone who had spent his career arguing that the clothes were a pretext and the woman was the point. The thesis was the picture; the picture was the thesis. Everything that followed in the next six years, the big-haired Versace catwalks, the George Michael Freedom! '90 video, the perfume contracts, the trademarked first names, started here and worked outward from a shoreline.

What is strange in hindsight is how thin the apparatus around it was. No production designer, no celebrity, no set, almost no clothes. A photographer who had argued his way into a magazine he didn't like and a stylist who agreed, against her own instinct, to leave the jewellery in the case. The picture they made together did not announce itself. The supermodel decade walked out of it slowly, like the tide going back.