Plutonic Rainbows

A toy world is a forgiving place to build an intelligent machine. Put a few blocks on a table, give the program a neat description of them, and search can look uncannily capable. Add the furniture, the door, a person walking through it, the million irrelevant details a real room contains, and the trick becomes a bill. The machine has far too many possible next moves to examine.

That objection sat at the centre of Sir James Lighthill's review of British artificial intelligence. Submitted to the Science Research Council in 1972 and published the following year as Artificial Intelligence: A General Survey, the report did not argue that computers could do nothing useful. It argued that ambitious work kept running into the same wall: as the world a program needed to represent grew, the possible searches through that world grew violently faster.

Lighthill called this combinatorial explosion. I prefer the plainer description: intelligence was expensive precisely where it appeared to be general. A system could solve a carefully bounded problem by carrying a great deal of human knowledge inside it. Let the boundaries loosen, and the method began spending its time discovering how much it did not know. That is less a philosophical defeat than an engineering one, but funding committees tend to recognise engineering defeats quickly.

There is an awkwardness here. John McCarthy, in his response to Lighthill, did not deny the explosion; he objected that AI researchers had known about it all along. He argued that heuristics existed to reduce just this kind of search. The dispute was not between a sober mathematician and fantasists who had never met complexity. It was about whether the field's workarounds amounted to progress, or merely made small demonstrations look more impressive than they were.

Britain's later account of this period is more careful than the neat myth. A parliamentary history of AI evidence records the report's connection with the British AI winter, while resisting the idea that one hostile document simply switched off a living field. The historian Jon Agar makes a similar correction in his reassessment of the report: Lighthill mattered, but his report also landed inside an argument about what publicly supported science was for.

This is why the story feels different from an ordinary wrong prediction. The earlier American decision to reduce machine-translation support, which I wrote about in Pierce's Verdict, had a similar trap built into it: evaluate a field by what its present methods can deliver, and you may be entirely fair while still starving whatever method comes next. Lighthill's case is harder, because the expensive search never went away. Bigger machines and learned representations did not abolish the problem. They changed which parts of it we could afford to tolerate.

I distrust histories in which the critic is made foolish by subsequent success. He had found a live wire. General systems still become costly when the world they must handle expands, even if their failure now shows up as a compute invoice or an answer produced without enough grip on the facts. The frustrating part is that a serious objection can be both true and mistimed.

The surviving document has the dry force of a public expenditure review: show me where this scales. It is not a glamorous demand. It is also the question that keeps returning whenever an impressive demonstration asks to be mistaken for a reliable world.

Dunmail Raise is not where I would choose to discover that a car had stopped cooperating. On the A591, the old AA telephone box appears in the records as a point of practical help: box number 487, placed on the Keswick-Grasmere road, ten miles south of Keswick, in the AA's historical list of roadside boxes. What survives now is less a telephone than a small black-and-yellow instruction in how fear once worked.

Before a mobile could turn a breakdown into an administrative nuisance, the road retained stretches where contact had a location. You had to reach the box. Weather, darkness, an injured passenger, a car that couldn't be left unattended: each fact counted while the telephone stood elsewhere, fixed and indifferent. I don't miss that vulnerability, but I do recognise the peculiar dignity of an emergency system that admitted distance rather than disguising it.

The AA's own timeline records the turning point with unusual neatness. In 1968 its wooden sentry boxes were phased out, apart from those protected by listing or retained in scenic places, as the network peaked at 787 boxes. In 2002 the telephones were decommissioned because mobile phones had made them redundant. Between those dates, the box moved from active equipment to something the landscape could keep after its reason for being had gone.

Number 487 had acquired another kind of protection before that final switch-off. Historic England lists the box as Grade II, first listed on 27 January 1987, its photograph still showing the black structure with AA yellow lettering and the number set out on the eaves. Listing is an odd form of aftercare. It can preserve the shelter, the paint scheme, perhaps the exact scale of a door, but it cannot preserve the moment when opening that door altered the odds of getting home.

I am tempted to call the box comforting, and that isn't quite honest. Its whole design assumes a failed journey. It belongs to the old grammar of the road: know the route, note the last petrol station, be aware that the next human voice may require a walk. The smartphone has improved most of this beyond argument. It has also thinned the visible evidence that we depend on systems at all. A call now seems to rise from the hand, not from a maintained network, charged battery, mast and contract.

There is a yellow severity to an AA box that the red public telephone kiosk doesn't have. A red box could be social, even faintly theatrical; the black and yellow box speaks only of trouble and the organisation summoned to deal with it. On an exposed road that narrow purpose must once have been a relief. Now it makes the surviving structure unusually stark: an emergency verb left behind after the sentence has changed.

Passing box 487 today would not make me want the old arrangement back. I would still check the charge on my phone and keep driving. Yet the little listed box marks something that constant connection has made difficult to feel clearly: help used to occupy a place in the landscape, and until you reached it the road was allowed to keep you waiting.

On Saturday, DeepSeek announced that the 75% discount on its flagship V4-Pro model is no longer a discount. It's the price. The promotion was due to expire on 31 May; instead the company locked the new rates in indefinitely. Output tokens now cost $0.87 per million. Cached input sits at $0.003625. The standard input rate is $0.435.

For context: GPT-5.5 charges $5 per million input and $30 per million output. Claude Opus 4.7 is $5 in, $25 out. Decoder ran the comparison and put V4-Pro at roughly 34 times cheaper than GPT-5.5 on output, and about 52 times cheaper once you cross GPT-5.5's long-context tier above 272K tokens. Those gaps aren't margin; they are a different business model wearing the same product shape.

The framing of "price war" has been used for every Chinese model release since the original V3 in late 2024, and it has become a tired phrase. What is actually new here is the supply story. V4-Pro is the first Chinese frontier model that runs natively on Huawei Ascend 950 silicon rather than Nvidia. DeepSeek told customers at launch a month ago that prices would ease once Ascend 950 supernodes started arriving in volume, and warned that until then the Pro tier could cost up to twelve times more than the lighter Flash model because of compute constraints. Saturday's lock-in is the company saying the supply problem is solved well enough to commit to the new floor.

That is the part worth paying attention to. The token price is a headline; the chip pivot is the structural fact. US export controls on the most advanced Nvidia parts pushed Chinese buyers toward Huawei. A second layer of restrictions on chipmaking equipment slowed Huawei's own ramp. Both pressures are still in place. What changed is that DeepSeek has decided the Ascend pipeline is reliable enough to price against, which is a different kind of bet than running benchmarks on borrowed hardware.

The interesting question is what this means for buyers who do not live inside Anthropic or OpenAI's stack. The flagship tax was already collapsing inside the Western labs; you could get 95% of the quality for a third of the cost by dropping from full GPT to a mini variant. DeepSeek's move is a more aggressive version of the same trick, only the discount goes to roughly two cents on the dollar and the savings are independent of which tier of Western model you benchmark against. For a CTO running document analysis or codebase review across a million tokens of context, the math is no longer close.

The complications are real. Training-data provenance for V4 is opaque, Anthropic has openly accused DeepSeek of distillation against earlier Claude generations, and routing enterprise traffic through a Chinese API still trips most large companies' procurement processes. Whether that gets resolved through audited deployments, private hosting, or just slow erosion of caution is the actual question. Reuters reported DeepSeek is chasing a $45 billion valuation off this strategy. The playbook is Amazon Retail circa 2002: give up margin, take the demand, build the moat. The new wrinkle is that the moat is silicon made in Shenzhen.

There's a smaller observation underneath all this, which is that the phrase "frontier model" is starting to do too much work. V4-Pro is not the most intelligent model in the world. It is the cheapest model that is intelligent enough for almost everything, and that has become a separate axis of competition entirely.

Vogue's note on Gianni Versace's Fall 1992 ready-to-wear collection carried a sentence that has stayed with me longer than any of the photographs from the show. "Getting strapped required eight hands." The clothes were beautiful in the usual Versace way, all glow and gold and a kind of Roman-emperor confidence, but the line betrayed how much engineering it took to land the silhouette. A woman couldn't walk into one of these looks alone. The dress had become a small infrastructure.

Versace called the show Miss S&M, in the sort of unembarrassed register he had been working in for years. The runway pulled in the supermodels of the moment, the same handful of women whose faces Lagerfeld was reading against Rose Macaulay on the Chanel runway earlier that year. What separates the two shows is that Lagerfeld's was a literary footnote in couture; Versace's was a hardware shop. PVC, oxblood leather, silk straps with steel buckles, and the safety pins that would soon define the house: not yet the giant chrome ones that held Liz Hurley together at the Four Weddings premiere two years later, but the smaller silver ones already migrating from punk shorthand to evening-dress structure.

The reception split along a fault line that fashion writing still finds awkward. Helmut Newton, whose own work had been engineering women into hardware for thirty years, told Vogue he loved it. Suzy Menkes, less so. "I don't want women to be sex objects or any of that," she said immediately after the show, then added, "But, after all, women have a right to choose." It is the kind of line a critic delivers when the work has refused to give her a clean exit. She had walked into the show with a position and walked out with half of it.

Versace himself gave the season its post-show punctuation. At an AIDS benefit in New York a few weeks later, the same looks turned up on the guest list, and he crowed to The New York Times, "Last night, there were two hundred socialites in bondage." The quote is archived in the trade press as a kind of victory lap, although what it actually marks is the moment a collection's vocabulary crossed from runway to red carpet without anyone losing their nerve. The clothes were never meant for a dungeon. They were meant for the front of a benefit photograph.

The other thing the show did, almost by accident, was institutionalise the Medusa head. Versace had been using the face on stationery and press packs, but Fall 1992 is the season the trade press dates as the emblem becoming the house's permanent logo, set into safety pins, clasps, buckles and buttons rather than printed on labels. By the following summer the Medusa was as inescapable as the safety pin itself, and the safety pin had stopped meaning punk and started meaning Versace.

What I keep coming back to is the eight hands. The whole moment sits inside that small piece of stagecraft. A collection presented as transgression that required, to actually wear, a quiet conspiracy of dressers backstage. The clothes look like rebellion, but they behaved like couture, and the gap between the two is where Versace made his case.

The Chambre Syndicale finally let him in. For the Fall-Winter 1992-93 season, Thierry Mugler was invited as a guest member of the haute couture calendar, and the decision was less generous than overdue. Couture in the early nineties had a problem. Buyers were ageing into the seats, press coverage was thinning, and the houses kept showing variations on the same drape. What the schedule needed was someone who treated tailoring as engineering and the runway as a venue for argument.

He answered by booking the swimming pool at the Ritz.

There is a Numéro retrospective of nine Mugler collections that walks through this debut with some precision. The pool gave the collection its name. Twenty seamstresses moved into the atelier and worked the season around the corset, which is the structural detail worth pausing on. Most couture houses at that point were assembling the silhouette from the shoulder down. Mugler started from the waist and built outward, the way a coachbuilder starts from the chassis. The 1989 bodywork-bustier that Naomi Campbell wore in the Buick collection had already proved the principle on ready-to-wear. The couture debut was the same logic stretched to the disciplines the Chambre Syndicale measures you against: hand-finishing, fitted-to-the-body precision, no shortcuts.

The Ritz pool is a strange room for clothes. The tiles bounce sound around the edges in a way no proper auditorium would tolerate, and the chlorinated humidity is hostile to silk. None of that mattered, because the room is also a stage set, art-deco depth and water-light and chrome handrails that pick up flashbulbs. The venue did half the work of arguing that couture could still surprise. The other half was the clothes, which the FIT Fashion History timeline catalogues alongside his ready-to-wear pieces from the same year, the bustier and corset traditions running in parallel between the two calendars.

What it bought him was a ten-year permanent membership in the couture calendar, which is the unglamorous answer to why this debut matters. Mugler now had a decade of January and July slots. He used it to do the Cirque d'Hiver anniversary show, the chrome gynoids and Cardi-B-shell-dress couture of Fall-Winter 1995-96, and Les Insectes in Spring 1997 with Galliano newly at Dior and McQueen at Givenchy in the next seats over. The Ritz pool is the show where that runway access was paid for, in 20 sets of hands working sleeves that took weeks instead of hours.

The footnote that always gets cut from the legend is that Mugler had been ready for this membership for a decade. The 1984 Zénith spectacle, the Too Funky bustier worn down a 1992 runway, the Atlantes mermaids and the Buick bodywork: the structural vocabulary was already complete. The Chambre Syndicale spent ten years deciding what a body could do, and then handed him the keys to a room he had already furnished.

Moxon Architects photographs Barking Riverside station at dusk, its metal screen catching a warm orange light while the square in front of it remains almost theatrically clear. It is a handsome London Overground terminus, clean, accessible and oddly expectant. I am used to railway buildings carrying the grime of the place they serve. This one seems to be waiting for the place to catch up.

The station opened to passengers on 18 July 2022. Barking Riverside Limited, the developer, describes an extension of 4.5km from the Gospel Oak to Barking line, arriving in a new public square with the Thames riverfront five minutes away. That wording is cheerful and quite precise. A station is usually sold as a way out: a short walk from home to a train, a reduction in the indignity of the commute. Here it is also a way in, an entrance provided before the district has settled into whatever ordinary habits will eventually define it.

Moxon's account of the design makes this feeling structural. Designed with Atkins and Burns & Nice, the station folds its working rooms into the footprint of the viaduct: ticket office at ground level, platforms overhead, an outer skin of stainless-steel panels moving from solid to perforated. Trains do not simply pull into a platform. They approach through a screened object that gradually admits movement, people and light. It is a threshold with an unusually confident idea of what lies on both sides.

Most haunted infrastructure has suffered a withdrawal. A pier has lost its steamers; an underpass has lost the precinct it was meant to feed. I recently wrote about a bus shelter standing after its last bus, the small cruelty of a public promise surviving after its timetable has gone. Barking Riverside is the inverse problem. The service is present. The square and interchange are present. What has not yet acquired a fixed shape is the life that makes a terminus feel inevitable rather than provisional.

I do not mean that the neighbourhood is vacant, or that a new railway is some melancholy mistake. That would be an easy aesthetic lie, the kind made by anyone who prefers a photograph of an empty platform to the inconvenient fact of people needing homes and transport. The developer's account of the opening is properly pleased with itself: trains running, buses connecting, river services close by, the journey to Barking compressed. This is infrastructure doing the decent thing and arriving early.

Early still has a strange atmosphere. Somewhere in east London there is a terminal destination on a departure board whose name sounds both settled and conditional. Barking Riverside. Not an old town absorbed by the railway, not a demolished works remembered by a branch line, but a name rehearsed aloud by automated announcements until daily life grows around it. Children will eventually regard the stainless-steel screens and the raised platforms as boring local facts. That is the desired result, and it requires an interval in which the future has opened its station but has not stopped looking like a rendering.

Pope Leo XIV released his first encyclical this morning, and the photograph the wire services chose was not of the pope alone. It was of Christopher Olah, Anthropic co-founder, standing at a Vatican lectern as the document was unveiled. The Associated Press headline put it bluntly: the pope calls out AI companies even as he hosts Anthropic. That tension is the whole story.

The encyclical is called Magnifica Humanitas, roughly forty-two thousand words across eighty-three pages, and its argument is that artificial intelligence is the labour-and-dignity question of this century in the way industrial mechanisation was the question of the century the Catholic Church first wrote about in Rerum Novarum. Leo, an American mathematician by training, has been telegraphing this framing since the days after his election in May 2025, when he named AI as the defining challenge of his papacy. So the document itself was not a surprise. The supporting cast was.

The Vatican could have launched a 42,300-word manifesto on AI with any number of partners. It picked a frontier lab presently locked in a legal fight with the Trump administration over access to its own technology, and it put that lab's chief interpretability researcher at the podium. Olah is not the policy face of Anthropic; he is the mechanistic-interpretability lead, the person whose work is closest to the actual question of whether anyone understands what these models are doing internally. Choosing him rather than Amodei was a curatorial decision, not an accident. The Vatican wanted the scientist, not the lobbyist.

What the encyclical itself says is harder than the press summaries suggest. Leo writes that it is "not permissible" to delegate irreversible lethal decisions to AI systems, which lands as a direct rebuke to autonomous-weapons procurement and, by extension, the deregulatory current the current US administration has set running. He calls on developers to "disarm AI" in a sense that goes beyond weapons: disarm it of the assumption that profit organises its deployment, disarm it of the cultural authority to displace whole classes of work without negotiation, disarm it of the presumption that speed is its own justification. The CNET write-up notes that the phrase is already the line being quoted back, and you can see why. It is the kind of formulation that survives translation.

What is interesting about Anthropic's presence is that the lab has, for two years, been the loudest internal voice arguing roughly this position from inside the industry. The pope disagrees with most of how Anthropic operates and has now said so in the most formal document a pope can produce. Anthropic showed up anyway, and let him. That is either an extraordinary act of public humility or an extremely sophisticated piece of positioning, and the honest answer is probably both at once. The encyclical will be quoted in regulatory hearings for years. Olah was in the photograph.

In 2012, three researchers from the University of Toronto submitted an image classifier to a contest and opened a much larger argument. Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton's model scored a 15.3 per cent top-five test error in the ILSVRC-2012 competition, while the second-best entry managed 26.2 per cent. The gap, set out in their original NeurIPS paper, wasn't a marginal win that needed a persuasive press release. It was a result other computer-vision researchers had to explain.

What holds my attention isn't just the winning number. The paper says the network trained for five to six days on two NVIDIA GTX 580 GPUs, each with 3GB of memory. Graphics cards were no longer merely drawing a synthetic world for a screen; here they were helping a network sort the visible world into categories. That shift now seems inevitable because it succeeded, but few changes in computing are inevitable before someone makes them run.

AlexNet didn't arrive through a single brilliant trick. It joined a deep convolutional network to a sufficiently large labelled image set and a GPU implementation quick enough to train it at useful scale. The authors also used non-saturating units and dropout, choices that mattered, but the awkward physical detail remains decisive: the experiment had to fit across two graphics cards. Software theory met memory limits, heat and a week of waiting.

IEEE Spectrum's history of AlexNet gets this balance right. ImageNet, CUDA and neural networks had each been developing before the contest, without producing this particular shock on their own. Krizhevsky had already written GPU convolutional-network code and extended it for ImageNet and multiple GPUs. A dataset, an unfashionable method and hardware built for another market met at exactly the useful moment.

There is a version of AI history that makes progress sound like an orderly succession of better ideas. I don't trust it. Plenty of ideas sit in papers for years because they are too expensive, too slow or too fiddly to win an ordinary comparison. Then a machine built for videogames makes a previously impractical calculation bearable, and the field suddenly discovers that its taste has changed. After a result like 15.3 against 26.2, skepticism starts to look less like rigour and more like a backlog.

It is tempting to treat this as the direct origin story of every generative model now taking up server halls and headlines. That would flatten too much: an image classifier is not a language model, and today's systems contain many later inventions. Still, AlexNet exposed a habit that remains with us. We talk about intelligence as though it lives solely in algorithms, while the winning idea often depends on which computation can be bought, powered and repeated enough times.

Two GTX 580 cards are modest hardware by current standards. Their place in this story is useful precisely because they look modest now. A research field can pivot on a benchmark result, but it can also pivot on an engineer finding that hardware from one culture of computing is suddenly good enough to remake another.

Of every ten corporate networks the UK AI Security Institute pointed Anthropic's Mythos at, six fell. With OpenAI's GPT-5.5-Cyber the number was three. Politico published the result this morning, buried in a Sunday explainer aimed at people only now catching up to why Washington keeps holding closed-door briefings about a model nobody outside a small circle has touched.

The framing matters. AISI is not a marketing arm. It is the British government's testing body, the closest equivalent to an independent referee that AI cybersecurity capabilities have right now, and it just put a clean ratio on the offensive gap between two frontier models. Two-to-one. British AI Minister Kanishka Narayan, in a statement to Politico, allowed himself the line "cyber capabilities in leading AI systems are advancing much faster than we expected," which is the polite ministerial register for "this is worse than the brief said it would be."

Mythos has been the subject of increasingly anxious coverage since Anthropic released Project Glasswing's initial findings last week. The numbers there were also impressive: more than ten thousand high or critical vulnerabilities surfaced across partner software, with all the patch-pipeline pain that follows. But ten-thousand-vulnerabilities-found is a defensive metric dressed in offensive clothing. It is "look how much we caught." Six-out-of-ten corporate networks taken over is different. That is a head-to-head capability test on the attacker side of the ledger, run by an arms-length government body, and the result is not subtle.

The other quotes Politico collected ratchet the picture upward. Cloudflare's chief security officer described Mythos as a "real step forward" in AI's ability to find vulnerabilities and write the code to exploit them. Broadcom called its own internal findings "jolting." An unnamed member of the House Homeland Security Committee left a closed-door Anthropic briefing reporting that Mythos had broken into his bank account with ease. Each of those is the sort of detail that, if anyone else were saying it about an unreleased lab model, would read as marketing. Saying it about a model the lab has pointedly declined to ship lands differently. Evans, quoted in the same piece, says the plain thing: "these model developments mainly are advantages for attackers rather than defenders." That is the AISI ratio in English.

The hardest part to sit with is the implicit assumption that the defenders' tooling will catch up. Glasswing's whole pitch rests on that. Give the model to a coalition of two-dozen large companies and government agencies first, run the bugs into the patching queue, and the attackers will arrive at a hardened landscape. That arithmetic only works if the defenders' side of the equation is willing to do its half of the work at the rate the model is producing it, and the patch-capacity story suggests the institutions are not. Even if they were, the AISI test implies a model exists today that is materially better at offense than its sanctioned defensive twin. Three out of ten was already the previous frontier. Six is not a step on the same staircase.

Anthropic's argument has always been that capability of this kind will arrive whether they ship it or not, and the most responsible move is to be the lab that demonstrates the upper bound under controlled conditions. AISI's number is the form that demonstration takes when it leaves the briefing room. It is a useful number. It is also the kind of number that gets cited later in unrelated testimony.