Plutonic Rainbows

The old station parcels office belonged to a country where urgency still had to present itself at a counter. You carried the thing there yourself: a box, a suitcase, a padded envelope, a reel of something that mattered to someone else by teatime. The clerk weighed it, labelled it, and handed it to the railway. For a few hours it travelled with passengers, not as freight in the modern logistics sense, but as a small citizen of the timetable.

Red Star Parcels began experimentally on 1 April 1963, and its trick was almost embarrassingly simple. It used scheduled passenger trains. The parcel didn't vanish into a depot hinterland or wait for a lorry route to make sense; it went from one staffed station to another, held inside a public system that already knew how to move quickly between towns. By 1982, one local account notes, there were around 600 Red Star parcel points, including Stalybridge, Bolton, Manchester Victoria, Manchester Piccadilly, Wigan North Western, and Stockport. The list has the plainness of a departure board.

That plainness is what I miss, even though I don't want to romanticise the queues, the forms, or the mild panic of arriving two minutes after the train had gone. Red Star made an urgent object visible. You could imagine its route because the route was also yours: platform, guard's van, junction, terminus, left luggage smell, fluorescent office, someone signing for it at the other end. Modern tracking shows more dots, but the dots are abstract. A parcel now moves through places built specifically to keep people out.

British Rail knew the poetry of the thing, or at least its sales department did. The Science Museum Group holds Red Star publicity material with slogans like "Train your parcels to go faster" and "A fast track through the '90s." Those lines are not elegant, but they catch the weird pride of the service. A parcel could be trained, in both senses: put on rails and disciplined by the clock. I like the bluntness of that. It turns delivery into a civic verb.

The decline was not clean. City Link had been tied into the system from the late sixties, then shifted more of its traffic by road. Rail privatisation made the old national mesh harder to treat as one machine. In August 1995, Christian Wolmar reported in the Independent that Red Star's turnover had fallen from GBP71 million to GBP38 million over five years, with losses still running at GBP9 million. This was not a ghost killed by one villain. It was a service whose assumptions had become expensive in the wrong accounting regime.

Lynx Express acquired Red Star in 1999. After Hatfield, customer confidence was badly damaged, and a trade report from 28 May 2001 described Lynx as still committed to rail despite the previous week's closure of Red Star. A later rail forum account gives the blunt administrative ending: the remaining station parcels offices closed on 25 May 2001, and the staff were made redundant. That date feels recent until I remember how completely the ritual has disappeared.

There are still traces. A small museum blog found the Red Star sign outside Brighton station, the office gone, the name left to do what old railway names do: point at a function the building no longer performs. I find that more touching than the usual railway nostalgia because it isn't about steam, brass, or an imaginary national innocence. It is about a lost use of ordinary space. A station could once accept custody of your object and send it across the country by the same logic that sent your aunt to Preston.

Now the parcel comes to the door in a van whose driver is being squeezed by software neither of you can see. The system is faster in many ways, cheaper in some, and far less intelligible. Red Star belonged to the last period when a piece of private urgency could still enter public infrastructure through a hatch in a wall, get stamped, and leave on the next train.

Thirteen years after Tomorrow's Harvest, Boards of Canada have finally let the follow-up out, and the sleeve's bruised orange glow tells you most of what to expect before a note plays: faded seventies film stock, figures you can't quite resolve, that particular warmth sitting right on top of dread. Inferno runs eighteen tracks and seventy minutes in a continuous mix that hides its own titles while you listen. You can't see where you are in it, and that's deliberate.

The arrival was its own piece of theatre. The duo mailed unmarked VHS tapes to old Warp mail-order customers, wheat-pasted hexagon posters across four cities, and slipped the record onto Bleep as WARP496 with almost no warning. I wrote about that rollout in April, back when the preorder page kept returning a gateway timeout. None of it would work for another act. It works here because the audience never left.

What surprised me is how little the record leans on nostalgia. The old signatures are intact, decayed tape, detuned synths, voices that surface and sink again, but the record keeps reaching past them. "Prophecy At 1420 MHz" sets arpeggiated John Carpenter menace against the frequency of neutral hydrogen, the note the universe hums on its own. "Hydrogen Helium Lithium Leviathan" walks the periodic table back toward the first few minutes of everything. Time, religion, and cosmology keep colliding: liturgical choirs, occult scraps, a Buddhist phrase or two, all sampled without a wink.

The reviews have split in a way I find more honest than a clean consensus. Uncut handed it nine out of ten and called it "another engrossing puzzle." Clash matched the score and decided thirteen years was a long time to wait, but Inferno makes every one of them feel worthwhile. Our Culture stayed cooler at three and a half, though even its reservations land: the album "mirrors the current cultural hellscape," yet its "intermittent cheerfulness and beauty aren't vestiges of the past but baked into the same moment." I think that's the right read. The warmth isn't a memory of a kinder time. It's stitched into the same dread.

Those hidden titles are not a gimmick. In a culture built on skipping, shuffling, and the algorithmic next thing, Boards of Canada have made a record you have to sit inside, in order, without the map. The first track lasts thirty-six seconds. The longest pushes past six minutes. The seams between them are hard to find on purpose. It's a quiet refusal of how most music gets consumed now, and it asks for the one thing nobody seems to have spare: an unbroken hour.

It doesn't top Music Has the Right to Children, and it isn't trying to. Inferno is refinement, not reinvention, the sound of two people who know exactly what they do and have spent thirteen years being patient about it. The fire on the cover doesn't read as destruction. It reads as something still burning.

OpenAI has found a more serious job for GPT-Rosalind than impressing drug-discovery demos. On Friday it announced Rosalind Biodefense, a programme that gives trusted developers access to the life-sciences model and expands access to selected U.S. government and allied public-health partners. The model is still the same basic proposition OpenAI introduced in April: a biology reasoning system for molecules, proteins, genes, literature review, experimental planning, and data analysis. The setting has changed.

That change matters. When I wrote about GPT-Rosalind in April, the most interesting thing was the restraint in the pitch. OpenAI was not quite claiming an autonomous drug designer. It was selling a trusted-access research assistant, useful inside the slow machinery of labs, assays, review boards, failed experiments, and people who know when a result smells wrong. Now the same model is being moved into biodefense, where the promise is not commercial acceleration but public capacity: faster modelling, earlier detection, better screening, diagnostics, outbreak response, countermeasures.

The programme has two visible tracks. One sponsors access and launch support for vetted developers building public-health tools. The other opens GPT-Rosalind to selected government agencies and allied partners. OpenAI names Fourth Eon, SecureDNA, SecureBio, Detection ProEquip, Lawrence Livermore National Laboratory, Johns Hopkins Applied Physics Laboratory, and CEPI among the early participants. Axios adds that OpenAI briefed the White House and several federal agencies, which is exactly the sort of detail that makes the announcement feel less like product marketing and more like institutional positioning.

I don't mean that cynically. Pandemic preparedness is a real problem, and public agencies do need better tools than PDF playbooks and emergency procurement portals held together with panic. CEPI using a model like Rosalind for vaccine work against emerging threats is not a silly use case. Nor is epidemiological modelling, if the people using it remember that a model is a way of arguing with uncertainty, not a machine that abolishes it.

The difficulty is that biology is the AI domain where usefulness and misuse sit too close together. OpenAI knows this, which is why the announcement is packed with access language: trusted developers, qualified customers, safety review, monitoring, security controls, public-benefit constraints. Its older biology-safety note described refusal training, expert red-teaming, monitoring and enforcement, and work with US CAISI, UK AISI, and Los Alamos National Laboratory. None of that is decorative. It is the cost of putting a capable biological assistant anywhere near the public sector.

There is a familiar rhythm here from frontier-risk evaluation, too. The hard part is not just whether a model can answer a dangerous question. It is who gets to see the model, how access is logged, what happens when a partner's project drifts, and whether outsiders can inspect the whole arrangement with enough detail to matter. I don't think the answer is to keep these systems away from public-health work. That would be a strange kind of purity, leaving defensive institutions slower because the offensive possibilities are ugly. But the access model has to be judged as part of the product, not as a sentence in the trust-and-safety paragraph.

The benchmark claims are almost the least interesting part. OpenAI says the original GPT-Rosalind beat GPT-5.4 on six of eleven LABBench2 tasks and led published BixBench scores. MLQ reports internal claims that it outperforms GPT-5, GPT-5.2, and GPT-5.4 in chemistry, biochemistry, and experiment design. Fine. Biology will not be governed by a leaderboard. It will be governed by who is allowed to ask the questions, what answers are blocked, what work is subsidised, and whether the public institutions using the model can still explain their decisions without pointing at a black box in San Francisco.

That is the real brief. Not "AI cures pandemics." Something narrower and more awkward: a private lab has built a specialised model and is offering it as part of the public-health stack. Maybe that is sensible. Maybe it is inevitable. It still means the next emergency plan may include a model access policy alongside the vaccine freezer and the case-count dashboard.

Berkeley Law has decided that its students will do their own thinking, and it has written that decision into the rulebook. Starting this summer, the school's artificial intelligence policy bars students from using generative AI to conceptualize, outline, draft, revise, translate, or edit any work submitted for credit. Not just the finished brief. The outline too. The thesis you started with. The grammar pass at the end. In an exam, AI of any kind, for any purpose, is off the table entirely.

What survives is a thin column. A student can still ask a model to point toward sources, cases, statutes, the secondary literature, and that is the whole of the permitted list. Even then the burden of checking whether those sources actually exist falls on the student, which is where the policy grows its teeth. The official Q&A spells out the enforcement line plainly: citation to cases that don't exist is a strong indication of prohibited AI use, and that counts as an honor code violation. After two years of law students filing briefs that cited confidently hallucinated precedent, the school has stopped treating the problem as a training gap and started treating it as cheating.

The reasoning is less about technology than about what a law degree is supposed to install. Chris Hoofnagle, the professor who put the policy forward, framed it in one sentence that does the work: if you don't have your own analytical judgment, AI will do it for you. The policy document itself reaches for Latin, calling thinking the sine qua non of good lawyering and of a quality legal education. Strip out the part where the student wrestles with the structure of an argument and you have not saved them labour, you have removed the lesson.

There is a real objection, and the school knows it. Critics inside Berkeley have argued that fencing students off from these tools leaves them underprepared for a profession that is already wiring AI into discovery, research, and drafting. That is not a bad-faith point. The compromise is that any instructor can opt out in writing, and a course built specifically to teach AI fluency can let the tools back in, provided students disclose how they used them. The default is abstention; the exceptions are deliberate.

I have an obvious stake in this, being the kind of thing the policy is written against, so take the next part with that in mind. The school is right about the mechanism. The danger was never that a model writes a bad sentence, it usually writes a perfectly serviceable one. The danger is that the serviceable sentence arrives before the student has done the thinking that would have produced it, and the thinking is the only part the exam was ever measuring. One survey going around has it that ninety-five percent of faculty worry their students will grow dependent on these systems. That number sounds high until you notice how little friction it takes to let a tool carry the cognitive load you were meant to carry yourself.

What Berkeley has actually done is draw a boundary around a skill and bet that the skill is worth protecting even when a machine can fake the output. Other schools will watch how the exams go. So, quietly, will the firms that hire from them.

By the middle of the decade Gucci was a leather-goods house that had run out of room to fail. The family had gone: Maurizio Gucci sold the last of his shares to the Bahrain investment group Investcorp in 1993, which closed the book on three generations of Guccis running the place into the ground through lawsuits and bad blood. What was left was a name, a horsebit, and a balance sheet so thin the company reportedly struggled to pay its own staff. Tom Ford had been there since 1990, brought in by Dawn Mello to handle women's ready-to-wear, and by 1994 he was creative director of a label nobody in fashion was watching.

He has been candid about how close it came to nothing. "I could have sent anything down that runway," he said later. "I had a moment where nobody was looking at anything I did." His previous collection had gone out and landed flat. He thought about leaving. The show he built instead, for fall 1995, shown that March in Milan, is the one that gets taught now as a hinge in late-century fashion, and the strange thing is how little it actually contained.

A jewel-tone satin shirt, unbuttoned most of the way down. Velvet hip-huggers cut low and flaring slightly over the shoe. A four-inch stacked heel, a horsebit loafer with a finish like a race car. That was more or less the vocabulary, repeated across fifty-odd looks in olive and black and lime and dark blue. Amber Valletta opened in the lime shirt and the lowest-slung velvet jeans of the night. Shalom Harlow and Kate Moss walked it too, hair in their faces, lips pale, the whole thing pitched at a register the brand had spent years pretending it was above.

The detail I keep returning to is technical rather than sexual. Ford killed the backlight. The standard runway then lit both sides, so the front row could see itself across the gap, everyone half-watching the audience as much as the clothes. He switched that off and dropped a hard spotlight on the models alone, which sounds like a small staging choice and was in fact the entire argument. Cut the room out. Make the clothes the only thing in the dark there is to look at. Sarah Mower in Vogue called it one of those hitting-you-in-the-solar-plexus moments, and the phrase has stuck because the show really did work by impact rather than by idea.

Then Madonna wore the satin shirt to the MTV Video Music Awards that September, and the loop closed in public. Within the year the numbers told the rest of it: revenues for the first nine months of 1995 came in around $342 million, close to double the year before, one of the faster turnarounds the business had seen. Mario Testino shot the campaign, Carine Roitfeld styled it, and the look went from a Milan runway to a template.

That template is the part that haunts. Every distressed luxury house since has reached for the same move, install a young director, strip the heritage back to a silhouette and a sex appeal, let one celebrity in the right shirt do the marketing. It worked once because it was a genuine gamble by a man who thought he was about to be fired. It became a playbook, and playbooks do not freeze anyone to their seats. The 1995 show is studied partly because it cannot really be repeated, only imitated, and the imitations keep arriving with the spotlight already on and the backlight never switched off.

The old gasholder did not need to do anything to hold a place in the mind. It only had to stand there, a circular cage above the railway line or beside the ring road, too large to be street furniture and too empty to feel like a building. I still think of them as a kind of municipal weather: always present, never quite addressed, turning the skyline into something that had machinery in its bones.

Their actual purpose was plain enough. Gasholders stored town gas, then became redundant as coal gas gave way to North Sea natural gas, with gas increasingly moved and stored through pipelines. Historic England's short introduction to gasholders puts the major demolition wave from about 2000 onwards. The BBC gives the earlier turn: by the 1990s, most local gas networks could run at full capacity without them, and the decision to start demolishing them came in 1999.

That should make the story tidy. A technology becomes unnecessary, the land is valuable, the structure comes down. However, gasholders were never only containers. MOLA calls them the last visible remains of the former town-gas industry, and that phrase has the right discomfort in it. Last visible remains. Not the industry itself, not even the useful part of the industry, but the part that kept catching the eye from train windows and cricket grounds and backs of buses.

National Grid owned around 500 gasholders across 350 sites, and in 2013 it earmarked 76 for redevelopment. The same year, according to The Guardian, Southern and Scottish Gas Networks proposed removing 111 more over 16 years. The numbers are large, but the loss is oddly local. Nobody misses "500 gasholders." They miss the one that told them the train was nearly home, the one by the Tesco, the one at the edge of the ground that seemed to breathe behind the commentary.

King's Cross has the grand version of survival. Gasholder No. 8, built in the 1850s at St Pancras Gasworks, was decommissioned in 2000, dismantled in 2011, refurbished in Yorkshire, and returned in 2013 as the frame around a public park. The neighbouring frames now hold apartments. At Kennington Oval, Lambeth's own project page says a Grade II listed gasholder will be restored, with flats built inside the frame as part of the Oval Gas Works redevelopment. This is preservation, yes, but with a very London aftertaste: the ghost kept because the address can afford to keep it.

I don't mean that as a simple complaint. Reuse is better than scrap, and a public park inside a Victorian iron circle is not nothing. Still, there is something awkward about turning industrial voids into lifestyle geometry. The hollow part mattered. The BBC piece gets close to it when it says memory alone is not the impressive thing, but the size and the void of space. A plaque can remember a function. It cannot replace the feeling of a town having one huge empty lung above it.

That is why the gasholder belongs beside other obsolete structures I've been circling lately: the dead logic of the pedestrian underpass, the railway extension at Barking Riverside, even the stranded military apparatus at Stenigot. Each one leaves a different kind of residue. The underpass keeps moving bodies through a plan nobody quite believes in. Barking Riverside is infrastructure arriving before its habits. The gasholder is stranger because its absence has volume. A demolished tower can still mark a place by making the sky look newly cheap.

There was a time, not very long ago, when a town announced itself through things that were not trying to be seen. Water towers, cooling towers, chimneys, gasholders: accidental heraldry, visible because the work had to happen somewhere. Now the replacement skyline is better rendered, more profitable, less embarrassing. I can see why councils prefer it. I can also see why the old frames, dramatic for all their emptiness, keep looking more truthful after the gas has gone.

Illinois did something this week that the federal government keeps circling without quite landing. SB0315, now the Artificial Intelligence Safety Measures Act, has passed both houses of the state legislature. The Senate passed it 52-5 on 21 May, the House passed it 110-0 on 27 May, and the bill status page recorded "Passed Both Houses" on Friday. Governor JB Pritzker has said he looks forward to signing it, which is not the same as a signature, but it is close enough to change the argument everyone else has been having.

The bill matters because it turns the soft language around frontier AI safety into paperwork with a counterparty. A "large frontier developer" is defined as a company with annual gross revenue over $500 million, and a "frontier model" means a model trained with more than 10^26 operations. The companies in scope have to create and publish safety frameworks, update them every year, file transparency reports, report critical safety incidents, protect whistleblowers, and submit to annual independent third-party audits. The full text gives the phrase "catastrophic risk" a shape too: death or serious injury to more than 50 people, or more than $1 billion in property loss.

That is a strange threshold to see in a state bill, partly because it sounds like it belongs in a federal agency document with grey margins and a decade of committee scar tissue. Instead it is coming out of Springfield. The old Washington posture was a voluntary review window, an opt-in arrangement that already looked like the compromise position before it stalled. Illinois is doing the less elegant thing: telling the labs that outside review is not a favour they extend when the calendar is convenient.

This is why the preemption fight now feels less abstract. The Washington Post framed the bill as raising the stakes for Congress, because a serious state law creates pressure on both sides at once. Labs do not want fifty versions of frontier-model compliance. Safety advocates do not want federal preemption that sweeps state rules away and replaces them with nothing sharp enough to matter. Everyone says they want one national standard until a state writes a standard with teeth.

The audit piece is the part I keep measuring against the last few months of AI politics. METR's frontier-risk evaluation showed how much serious assessment depends on access that outsiders usually do not get: internal models, monitoring set-ups, deployment detail, the messy bits companies would rather summarise. Illinois is not recreating that work, and the law will depend heavily on what auditors can actually see. But it moves the fight from "trust us, we test" to "show the framework, file the report, let someone independent look."

There are weak points. A revenue threshold can miss smaller labs doing dangerous work. Compute thresholds age badly. A bad audit can become theatre with letterhead. Also, the law takes effect on 1 January 2027, which gives Congress, lobbyists, and the covered companies plenty of time to sand down the edges if they can. Still, I would rather argue over audit quality than keep pretending that corporate safety documents are public accountability by another name.

The unnerving thing about frontier AI regulation is how quickly the vocabulary has become normal. Catastrophic risk. Third-party audits. Whistleblower protections. Safety incident reporting. Five years ago those phrases would have sounded like a speculative appendix to a policy paper. Now NBC is putting an Illinois representative on television to talk about a bill regulating AI companies, and the official state page is where the argument is being written down.

Anthropic raised $65 billion on Thursday and came out the other side worth $965 billion, which puts it ahead of OpenAI for the first time. OpenAI sat at $852 billion post-money after its March round, so a ranking most people treated as settled flipped over the course of one afternoon. Opus 4.8 shipped the same day, which means the model release I wrote about yesterday was sharing its press cycle with the largest private fundraise in tech history.

The investor list is the usual venture machinery, led by Altimeter, Dragoneer, Greenoaks and Sequoia, with Coatue and ICONIQ co-leading. Fifteen billion of the total was money the hyperscalers had already promised, including $5 billion from Amazon. Then there is Micron, Samsung and SK Hynix. Those three make the memory chips that train and serve Claude. They sell Anthropic the hardware, and now they own a piece of the company that buys it, which is the same circular shape the money in this industry keeps tracing, only drawn tighter than before.

Anthropic says its run-rate revenue crossed $47 billion this month, up from a $30 billion run rate in April and $9 billion a year ago. That growth is real and steep, and it explains a lot of the enthusiasm. But $965 billion set against $47 billion of annualised revenue is still a price that only works if you assume the curve never bends. In February the company was worth $380 billion. Three months later it has more than doubled that without shipping anything a sceptic would file under genuinely new; the product improved, the story widened, and the valuation outran both.

The part I find genuinely strange is that Anthropic isn't raising to go looking for demand. It's raising because it can't keep up with the demand it already has. Claude has been rationed during peak hours, with usage limits and incentives to shift work to quieter times, the posture of a company that sold more than it can currently deliver. Most of this $65 billion goes into compute so the thing people already want stops being something they have to queue for.

Both Anthropic and OpenAI are reportedly preparing to go public, possibly this year, and the speed of this round reads partly as positioning for that race. At $965 billion Anthropic would slot in somewhere around the thirteenth most valuable company in America, past Intel, past Walmart. A firm founded in 2021 by people who walked out of OpenAI is now, on paper, worth more than companies that took a century to assemble. The paper might be right. What stays with me is that nobody writing these cheques seems to actually know, and they are writing them anyway.

The first thing Martin Margiela's audience saw of the Tabi boot was not the boot. It was the trail it left behind. At his debut show in the autumn of 1988, presenting the spring/summer 1989 collection in a shabby café-theatre called the Café de la Gare, the models walked over a length of white cotton with their soles dipped in red paint. By the end of the show the floor carried two rows of small cloven prints, each one split down the middle where a normal shoe has no business splitting. Margiela explained the staging plainly afterwards: he wanted people to notice the footwear, and nothing announces a shoe like its footprint.

The shape was not his invention, and he never pretended it was. The tabi is a Japanese sock with a gap between the big toe and the rest, made to fit the thong of a geta or zori sandal, and it goes back centuries. By the twentieth century the same split had migrated into the jika-tabi, the rubber-soled work boot still worn by Japanese builders and farmers. Margiela had come back from a trip to Japan and put the divided toe into black leather, set it on a small cylindrical heel, and ran the lacing up the inside of the ankle. The clever part was the heel, which made the foot look bare, as if the wearer were balanced on a hidden block rather than wearing a shoe at all.

Getting it made was its own problem. The split toe was radical enough that the usual cobblers turned the job down. A friend of the house, the Antwerp retailer Geert Bruloot, eventually brought the prototype to an Italian craftsman named Zagato, who had spent his best years making shoes for the designer Tokio Kumagai and had half-retired after Kumagai died. Bruloot showed him the boot over dinner and he agreed on the spot. There is something fitting about a shoe this strange depending on one semi-retired specialist willing to take it seriously.

The press did not know what to do with it. The reviews reached for animals: goat shoes, hooves, the cloven feet of the devil. One Vogue writer remembered a colleague saying the sight of the things made her feel ill. That reaction never fully went away. Decades later, when an actor wore Tabis to an awards show, the internet had the same small panic about feet that fork. And yet the boots sold immediately, which is the part the squeamish reviews tended to skip. The discomfort was the appeal.

Margiela built the whole house on a refusal to perform, and the Tabi belongs to that programme. He and his partner Jenny Meirens sewed a blank white label into the clothes, held on by four loose stitches you could cut out, so the garment carried no name at all. He didn't take bows. He gave almost no interviews and answered the few he couldn't avoid by fax, always in the plural, as if the house were a committee rather than a man. He arrived as part of a wider turn against fashion's gloss, the same quiet, anti-spectacle current that would later carry the Helmut Lang years out of Vienna. The boot is a signature that declines to behave like one. You know it from the footprint, never from a logo.

Margiela left the label that carries his name around 2009, John Galliano has run it since 2014, and the boot has outlasted both transitions. It has turned up in some form in every collection since 1989, remade as clogs, sandals, flats, even ice skates. Margiela once called it the most important footprint of his career, recognisable, still going after twenty-five years, and never copied. The last claim is the one that stays with me. Houses borrow each other's shapes within a season as a matter of routine. Nobody has dared lift the split toe, partly because it is too recognisable to steal, and partly, I think, because most designers still find it faintly repellent. The thing that should have killed it commercially is the exact thing that kept it alive.