Plutonic Rainbows

China's National Development and Reform Commission posted a short notice on Monday saying it had decided to block the foreign acquisition of the Manus project and required the parties to unwind the deal. No press conference, no extended reasoning, just the kind of administrative sentence that takes two billion dollars off the board. Meta, according to the Wall Street Journal, is now preparing to disentangle a company it bought four months ago and has already integrated into Ads Manager and the Meta AI chatbot.

Manus's path to this moment is worth tracing because it explains why Beijing felt it had a case. The company was founded in 2022 in Beijing by Butterfly Effect, with an agentic AI product that Chinese state media briefly called the next DeepSeek. Last summer the founders, Xiao Hong and Yichao Ji, relocated the startup to Singapore, mostly to dodge the US chip export controls that were boxing in any frontier work done on Chinese soil. In December, Meta paid somewhere between two and two and a half billion dollars for it. By January, Chinese regulators were reviewing the deal. By late March, the two cofounders had been placed under travel bans. The Monday announcement was the back end of a process that had been running quietly since the ink dried.

What is striking is the legal mechanism. Bloomberg and CNBC both note this is the first time China has used the foreign investment security review measures it introduced in late 2020, the rules that established a dedicated office under the NDRC for screening deals with national security implications. The targets are unusual: a US tech company with effectively no mainland presence, and a startup that had already legally moved abroad. Beijing is asserting jurisdiction over a transaction conducted outside its borders by a company that, on paper, is no longer Chinese. Duncan Clark of BDA China summarised the message tartly to CNBC: founders will know that if you start in China, you stay in China.

There is an awkward subplot. Manus's product, the autonomous agent layer that Meta wanted, runs on Anthropic's Claude. If the company had remained in mainland Chinese hands, that dependency would have been severed by Anthropic's own restrictions on selling to Chinese entities. So even a successful Beijing "reversal" leaves Manus in an odd position, returned to a jurisdiction where its core engine isn't legally available to it. A former Biden official quoted in Ars Technica put the point plainly: if Manus had stayed Chinese, its product would have disappeared. Now it might disappear anyway, just for the opposite reason.

The timing is the loudest part. The blocked deal lands less than three weeks before Trump and Xi are due to meet in Beijing. Whatever the summit was going to be about, this is now part of it. The exclusivity unwind between Microsoft and OpenAI on Sunday and the Manus block on Monday read, taken together, as a week in which the commercial architecture of the AI industry moved from quiet renegotiation to public statecraft. One was a contract being loosened. The other was a contract being annulled.

I keep coming back to the relocation. Manus moved to Singapore to make itself acquirable by an American buyer, and the move itself appears to be what triggered the Chinese review. The gesture meant to neutralise the geopolitics turned out to be the geopolitics. There is a parallel with the DeepSeek-Huawei turn, where Chinese AI is being deliberately rebuilt on Chinese silicon: the direction of travel is the same, towards two systems that don't trade. Manus tried to step out of that pattern. The pattern stepped back.

In April 2021, Taro Kono, then Japan's minister for administrative and regulatory reform, announced that ministries in Kasumigaseki would stop using fax machines and switch to email. The cabinet's reform body followed up in June with a deadline of thirty June, after which no more thermal-paper handshakes would pass between the agencies that run the country. The directive went out on 7 June. The objections came back so quickly that the Hokkaido Shimbun stopped trying to count individual ministries and reported a round figure of more than four hundred formal complaints. A cabinet body, told by a minister to digitise, replied in writing that digitising would be "impossible."

The cabinet stood down. Exceptions widened until they swallowed the rule: disaster response, anything involving members of the public and businesses with established fax habits, anything touching the police or the courts. By July 2021 the fax-free ambition had become a footnote, and Kono moved on to the war he could win, against the floppy disk. By July 2024 the Digital Agency could announce that all 1,034 regulations governing floppy-disk submissions had been scrapped. Kono told Reuters they had won the war on floppy disks, and did not mention the other war.

The fax is still there, four years on. A November 2025 survey by Japan's education ministry put fax usage in public elementary and junior high schools at 71.7 per cent. The same survey found 91 per cent of schools requiring at least one hanko stamp on parental paperwork. A government policy from 2023 had named fiscal 2025 as the year both would end in education; fiscal 2025 came and went, and the numbers had moved by a few points at most.

The reason the fax survives is not really about fax. It is about what the fax delivers. A faxed document arrives as paper, already in the form a hanko can stamp, already physical enough to bind a hospital, a school, or a contractor to the words on it. The Group 3 standard most of these machines still run on was set in the early 1980s and has not been meaningfully updated since. That is its appeal: four decades of forensic familiarity, a log entry on both ends, and the comforting friction of a thing sent slowly enough that nobody can claim they did not notice it arrive.

Japan's bureaucratic trust infrastructure calcified around this combination of paper, stamp, and telephone-line acknowledgement. To pull the fax out is to pull out the load-bearing piece of a much older arrangement, and the ministries who filed those four hundred objections were not defending a machine. They were defending the procedure that machine certifies. In 2020 a respiratory specialist at a public hospital tweeted that COVID case numbers were still being handwritten and faxed to the health ministry, and the ministry relented within weeks. The fax ban, when it finally came in that narrow domain, came not from a digital minister but from a doctor with two thousand followers losing his patience in public.

Kono's loss in 2021 was not really a defeat by the fax. It was the moment a digital reform programme noticed that the analogue infrastructure it had inherited was not a habit but a legal order.

By the time Gianfranco Ferré sat down to plan his spring-summer 1995 haute couture collection, he had been at Dior more than five years. The architectural shock of the Ascot–Cecil Beaton debut had already softened into something the French press grudgingly treated as legitimate. Eleven collections had passed under his hand, each given a name in the manner Christian Dior himself had once observed. He titled this twelfth one Extrême, a single word with no hedging in it.

The centrepiece was a sheath dress called Floridante, cut in bright yellow lace and embroidered to the brink of structural collapse. The Dior Héritage collection in Paris still keeps it catalogued under that name, photographed by Laziz Hamani for Alexander Fury's Assouline volume. Around the Floridante, Ferré built a runway of saturated colour, exploded floral prints, and silhouettes the in-house notes described as "lively and striking as an etching." The New York Times called the show "decadent" and "dreamlike," in the slightly suspicious tone American critics reserved for Italian designers who had not lost their nerve.

The reference points were openly pop. Warhol's flowers, Pollock's splatter, the saturated commercial colour of postwar American print. Ferré had been a guest at the architectural end of fashion for years, written about as the architect of fashion long before he reached avenue Montaigne, and now he was making clothes that wanted to behave like silkscreens. The off-the-shoulder overcoat in the finale, with a bow nipped high at the waist over a floral column gown, was Dior's New Look reread through a flower that had been photographed too closely.

What's interesting is the timing. January 1995 was the moment the mood in Paris and Milan was tilting hard toward black, toward deconstruction, toward the anti-ornament minimalism that Miuccia Prada and Helmut Lang and the Antwerp graduates were defining as the next decade's vocabulary. Ferré, six years into the most prestigious post in French couture, went the other way. He went toward saturation. He went toward print. He went toward the kind of yellow you only see in a Warhol cow.

He explained it in a sentence that reads, on the page, like a mission statement and a defence at the same time: "I am trying to respect the kingdom and the power of couture." Couture, in his reading, was not a register that should follow the ready-to-wear mood. It was supposed to lead it, or refuse to follow it, or do something the rest of the system could not afford to do. By 1995 that was already an unfashionable position. Couture houses were losing clients, the haute couture client base was contracting into single digits per atelier, and the argument for ornament was increasingly framed as nostalgia rather than craft.

Extrême is one of the collections that disappears in the standard narrative of Ferré at Dior, sandwiched between the early architectural triumphs and the Indian-themed finale that ended his tenure. It deserves better. The Floridante in particular is the moment you can see his method working at full volume, the body sculpted by the cut, the surface released to do something almost reckless on top of the structure. Architecture underneath, pop art on the skin.

The Galliano announcement was still ahead. The press, the buyers, and the French establishment had no way of knowing the ground was already shifting beneath the avenue Montaigne. Ferré, on the evidence of Extrême, did know, and chose to answer with more colour rather than less.

Hackney Council costed Clissold Leisure Centre at two million pounds and a 2000 finish. By the time it opened on the first of February 2002, almost three years late, the figure was thirty-two million. By 2007, after the centre had been forced shut for the better part of four years, the total cost of the building plus its remedial works sat at forty-five million. The original budget was off by a factor of more than twenty. The building had cracks in the squash courts, a roof that leaked across the entire centre, glass walls that retained fetid water around the pools, and a sports-hall floor that warped within twelve months of being walked on.

The official defect schedule, leaked to the Guardian in 2004, ran to fifty-nine numbered items. Defect thirty-two read: roof leaking across whole centre. Defect thirty-three read: roof sweating with condensation. Defect fifty-nine read: water damage to sports-hall floor causing warping and lifting at less than twelve months, with injuries sustained by users. Marcus Fairs, then editor of Icon, called the place "a monument to architectural arrogance and local government ineptitude." A swimming coach told the Evening Standard that rain poured through the roof into the pool from a jug, and that the open-plan changing village made the building a child-protection risk. The Muslim and Hasidic Jewish residents of Stoke Newington, whom the centre was nominally meant to serve, found the changing arrangements unusable on the day it opened.

The architect, Stephen Hodder, was an established and award-winning name. Gleeson, the contractor, was a familiar fixture in the procurement world. Hackney Council, however, had no architect's department of its own. This is the part Jonathan Glancey kept returning to in his Guardian piece a few months after the closure: the borough's local authority architects, the people who would have quietly stress-tested the brief and the drawings before a single tile was ordered, were gone. They had gone the way of council housing in the seventies, down the plughole, said Glancey, and by the time the lottery turned up with prodigious funds for bright new buildings there was no one in the town hall capable of holding a designer to account. The municipal swimming pools that Clissold replaced, late-Victorian baths at Haggerston and Whitechapel, had been thrown up by such departments and ran for a hundred years with few problems.

Hackney sued Hodder. Hodder denied responsibility and blamed the council's inability to host the project. In 2005 a confidential settlement was reached in which Hodder Associates and Gleeson paid Hackney an undisclosed sum without acknowledging fault. The centre reopened, partly, in December 2007.

What sticks about Clissold is that the building was not, by 2002, an outlier. It was the typology. The lottery-funded millennial leisure centre arrived as the proud successor to the seventies leisure-centre boom, and that boom had been a municipal project run by people who knew the limits of their own boroughs. The successor was a contracted-out icon flown in from outside, photographed for a touring exhibition called 12 for 2000: Building the Millennium, endorsed by Chris Smith on the steps of the British Council. Prophetic words, Glancey wrote. The replacement had been costed by people who had never seen the originals work, and the originals had worked because someone in the building knew the borough.

Forty-five million pounds bought Hackney a model of British architecture for an exhibition tour, and a leisure centre that could not keep the rain out of the swimming pool.

Anthropic announced nine new connectors today that wire Claude directly into the software professional creatives actually use. The list reads like the contents of a working freelancer's dock: Adobe Creative Cloud (Photoshop, Premiere, Express), Affinity, Blender, Ableton, Autodesk, Splice and several others. The announcement on Anthropic's site frames it as an extension of Claude Design from earlier in the month, but the ambition is broader. This is the company trying to sit inside the apps where the work happens, not on a separate tab where the work gets summarised.

The technical mechanism is MCP, the Model Context Protocol Anthropic introduced last year and which has since become the de facto standard for letting an LLM read from and write to outside tools. Each connector is a small server that translates Claude's requests into the host app's native API. The Blender bridge, for example, exposes Blender's Python API as natural language: ask Claude to instance a hundred copies of an object along a curve with random rotation, and it does the equivalent bpy calls. The Ableton connector is more modest, it indexes the official documentation and answers questions, rather than opening a session and arming a track. The Adobe one sits somewhere in between, able to pull assets from Creative Cloud into Claude's context and trigger actions back inside Photoshop and Premiere.

It is worth being clear about what this is and what it is not. None of these connectors replace the practitioner. The Verge notes that Anthropic itself is careful in the announcement copy: "Claude can't replace taste or imagination." The pitch is repetitive manual labour. Renaming layers, batch-tagging clips, building out a hundred variations of a packaging mock, sourcing a sample pack that fits a brief, generating the boring scaffolding around the interesting decisions. The interesting decisions remain a human problem. The argument is that if the boring scaffolding gets cheaper, the interesting decisions get more time.

Whether that argument survives contact with reality depends on which side of a creative team you sit. Senior people who already delegate the scaffolding to juniors will probably love this. The juniors whose job was the scaffolding will not. The historical pattern when tooling absorbs entry-level tasks is not that the work disappears, it is that the bottom rung of the ladder gets sawn off and the people who were supposed to climb it go elsewhere. The studios that are healthiest in five years will be the ones that figured out how to keep training people through the gap.

The strategic read is that Anthropic is now doing to the creative suite what it already did to coding and is trying to do inside the federal government. Pick a high-value professional vertical, ship a connector that makes Claude useful from inside the workflow, accumulate the kind of sticky usage that survives the next model swap. OpenAI, as the Atlantic noted this morning in a separate piece, is reliably about three months behind on this playbook. The follow-on Codex-for-Photoshop announcement should land before August.

The thing I will be watching for is the long tail. Nine connectors on launch day is a press release. Ninety connectors in two years, maintained by a healthy third-party community using the MCP spec, is a platform. Anthropic has been quietly betting that MCP becomes the USB-C of agentic tooling. Today's launch is the loudest evidence so far that the bet is being placed at the application layer, not just the infrastructure layer.

The pool sat under a galvanised-steel pyramid. That was the detail that lodged in your head if you ever swam there. Faulkner Brown finished Bletchley Leisure Centre in 1974, a year when the British state still believed it could provide a tropical interior to a Buckinghamshire new town, and the answer it arrived at was a space-frame tetrahedron over a heated pool, glazed to the sightlines, surrounded by sub-tropical planting that the maintenance crew pruned during the night shift. The building came down in 2010, replaced by something with windows that did not need to keep the heat of the tropics inside an East Midlands winter.

There was a peculiar dread to those rooms even when they were new. The chlorine fog hit your sinuses in the entrance corridor and stayed with the rest of the day. The acoustics of tile and concrete turned a child's shout into a prolonged metallic ring that bounced off the roof and came back wrong, half a beat late, slightly higher in pitch. The fluorescent lighting was always too much for the eye and not quite enough for the camera, so every photograph from the period has the same washed-out yellow that makes the swimmers look as if they have been preserved rather than caught. You walked through a footbath of disinfectant before the water itself, and the smell of it was the smell of being small and slightly afraid.

Otto Saumarez Smith's Lost World of the British Leisure-centre Boom, published in 2019 in the Twentieth Century British History journal, makes the case that these buildings were the last serious municipal architecture this country produced. His phrase is "holiday atmosphere", lifted from the trade press of the period: pop imagery, fun, sun, the unashamed lowbrow populism of a council that thought a working-class family in Bletchley deserved to feel for an afternoon as if they were in Mallorca. Twenty-nine degrees centigrade, year-round, paid for out of the rates.

That model began to collapse the moment the funding mechanism behind it did. Coventry Sports Centre, the zinc-clad "Elephant" of 1976, closed in 2020. Swindon's Oasis, with its perspex dome, went the same year. Bradford's Richard Dunn Sports Centre, the Brutalist tent-roofed complex inspired by Kenzo Tange's Yoyogi gymnasium, shut in 2019 after a council decision driven by the cost of upkeep, and during the pandemic it served as a temporary morgue. The Twentieth Century Society's campaigns manager Oli Marshall told Dezeen in 2022 that every pool in the country was now at risk: the pandemic, the global chlorine shortage, the energy price shock, all of it falling on a building type designed in an era of cheap municipal gas.

What survives mostly survives by listing. The Oasis got Grade II status in 2021. Richard Dunn got it in 2022. The buildings themselves are mothballed, half-derelict, surrounded by hoardings, the tropical planting long dead. You can find drone footage on YouTube of the empty pools, the tile still gleaming, the ghost of a footbath at every threshold.

Five hundred leisure centres have shut since 2010, and the pyramid roof is a kind of period marker for the moment a place like Bletchley could still be promised the tropics by its own council. The dread of those rooms came from the gap between the promise and the building's own admission that the promise was provisional. The chlorine fog, the metallic echo, the yellow light, were the texture of an institution that knew, somewhere, it would not last.

David Silver came out of stealth this morning with $1.1 billion in seed money and a thesis that runs against the prevailing weather. Ineffable Intelligence, his new lab, plans to build models that learn from their own experience rather than from the internet's archive of human writing. The round, reported by CNBC, is the largest seed in European history. Sequoia and Lightspeed co-led it. Nvidia, Google, DST Global, Index, and the UK Sovereign AI Fund piled in. The post-money valuation is $5.1 billion before there is a product, a paper, or, as far as anyone outside the cap table can tell, a model.

Silver is not a celebrity outside the field, but inside it he is the person who designed AlphaGo and AlphaZero. The work that made DeepMind's name in 2016 was reinforcement learning, agents that played themselves until they were better than anyone alive. AlphaZero learned chess from scratch in nine hours and then beat Stockfish, which had spent two decades absorbing every game humans had ever recorded. The lesson Silver took from that, and which he is now raising a billion dollars on, is that human data is a ceiling. Anything an AI can only learn from us cannot, by definition, exceed us.

The pitch deck almost writes itself. The frontier labs are spending the GDP of small countries on training runs that cannot continue at this clip, because the supply of high-quality human text is genuinely close to exhausted. The fix the industry has settled on is to pay forty labelers to write rubrics and hope the model generalises. Silver's bet is that the whole labelling layer is a distraction, that the actual gradient runs the other way, that a system which generates its own curriculum from interaction with an environment will skip the human bottleneck entirely. "Our mission is to make first contact with superintelligence," is how he put it in the press release. I flinched a little at the phrasing, but the technical claim underneath is real and not new. He has been making it for years.

What is new is that Sequoia is now writing a nine-figure cheque to fund the experiment in public. The talent flight from the big labs has been gathering for eighteen months. Mira Murati's Thinking Machines, Ilya Sutskever's Safe Superintelligence, and now Silver. Each one walks out of a hyperscaler with a plausible technical story, a Rolodex full of researchers, and a venture market willing to fund a five-billion-dollar valuation on day zero. The implicit critique is the same in every case: whatever the labs are doing inside, it isn't ambitious enough.

I am not sure the experience-based path scales the way the self-play games did. Chess and Go are closed worlds with crisp reward signals. The physical world is messy, slow, and expensive to simulate at fidelity. Reward hacking is the default outcome, not the edge case. The companies that have tried to do open-ended RL at the scale of, say, robotics, or science, or software engineering, have spent years learning how hard the reward-design problem actually is. None of that is solved by money, although money buys time to keep trying.

Still, a billion dollars is a real signal, and the people writing the cheques are not naive. The bet, as I read it, is not that Silver will reach superintelligence. The bet is that he will produce one or two genuinely novel results within three years that the existing labs cannot replicate without rebuilding their training stacks from the ground up. That is enough to clear the seed. Whether it is enough to clear the next round, when the science gets specific and the costs go vertical, is a question for 2028.

Six months after the last restructuring, OpenAI and Microsoft have rewritten the deal again. The new agreement, announced this morning on OpenAI's blog and almost immediately on every business desk in New York, ends the exclusive licensing relationship that has been the load-bearing wall of the partnership since 2019. Microsoft keeps its rights through 2032, but those rights are now non-exclusive. OpenAI can ship its models on Google Cloud, on AWS, on whatever infrastructure a customer happens to prefer.

That last clause is the one that matters. The Financial Times reported in March that Microsoft was considering legal action to block a $50 billion enterprise deal between OpenAI and Amazon, on the grounds that the original exclusivity language covered exactly that scenario. The new terms quietly retire the threat. Azure remains the first port of call. OpenAI ships there first, "unless Microsoft cannot and chooses not to support the necessary capabilities," at which point everything else is fair game. In practice, that escape hatch is wide enough to drive a fleet of TPU pods through.

The financial side has been turned around. Microsoft has paid a revenue share to OpenAI for years; that stops now. OpenAI continues to pay 20% back to Microsoft, but the obligation is capped and runs only through 2030, not indefinitely. So Microsoft trades open-ended upside on a runaway customer for a fixed ceiling and a clean exit window. That is not the move of a company that thinks the next two years will be the most valuable two years.

I am trying to read the body language and finding it hard. The official framing is mutual, friendly, "next chapter." The market read is more guarded; Microsoft shares dropped about 1% on the announcement, and Wedbush put out a note calling the deal a net positive because it ends a year of "limbo." Limbo is a generous word for it. Denise Dresser, OpenAI's revenue chief, sent an internal memo earlier this month complaining that the partnership had "limited our ability to meet enterprises where they are." Enterprises, in that sentence, means customers who already buy Bedrock from AWS and Vertex from Google and would quite like to add ChatGPT-class inference without having to re-architect their procurement.

There is a longer arc here that the press release does not mention. OpenAI's restructuring last October turned the nonprofit-owning-a-for-profit into something closer to a conventional company. Microsoft's stake, depending on whose valuation you trust, sits somewhere between $135 billion (NYT) and $225 billion (Forbes), or roughly 27% of OpenAI on paper. Wedbush thinks today's renegotiation clears a runway for an IPO, and that part feels right. Public-market filings work better when your distribution rights are not contingent on a single counterparty's willingness to keep the relationship warm.

What hasn't changed: the money still goes in circles. Microsoft funds OpenAI, OpenAI buys Azure, Microsoft books the cloud revenue, and the headline number gets bigger every quarter. The new agreement loosens the loop without breaking it. OpenAI gets to court Bezos and Pichai while still paying its toll to Redmond. Microsoft gets locked-in cloud spend without the political cost of being the only adult in the room.

It is the kind of deal you sign when both sides have grown up and noticed the other one is now their main competitor's landlord.

The plan was not to repel the invasion. The plan was to slow it down. A Southern Command memo dated 22 June 1940 set the intention plainly: divide England into small fields surrounded by a hedge of anti-tank obstacles, close the gates behind the German armour, and let the mobile columns round up the cattle. The cattle, in this metaphor, were the Wehrmacht. The hedge was concrete.

General Edmund Ironside had three weeks. By the end of June 1940 the Home Defence Executive was passing plans for thousands of pillboxes and tank-traps along beaches and at nodal points, and work began everywhere at once. Around 28,000 hardened field defences went up between June 1940 and February 1941, when the order came down to stop building. Ironside was already gone by then, replaced in late July by Sir Alan Brooke, who disliked fixed lines and preferred mobile reserves. The pillboxes remained. They had cost too much to demolish.

About a quarter of them are still here.

Roughly 7,000 concrete pillboxes survive across Britain, most of them along the GHQ Line, the 400-kilometre stop-line running from Somerset to the Medway, designed as the last fallback before London and the industrial Midlands fell. In Surrey it follows the Wey from Farnham to Shalford, then the Tillingbourne to Wotton, the Pippbrook to Dorking, the Mole to Horley. The pillboxes sit at roughly 500- metre intervals, sighted to cover the river or the road, often half-buried in undergrowth now. In Essex, between Great Chesterford and Canvey Island, over a hundred FW3 boxes still exist; about forty of them are highly visible from the A130, hexagonal concrete drums sitting in field corners as if they had grown there.

What strikes you about the surviving boxes is how indifferent the landscape has been to their purpose. A Type 22 hexagonal pillbox in a Suffolk field is now a shelter for sheep. A Type 24 on the Stroudwater Canal is a fishing platform. The loopholes have been blocked up with breeze-block, or left open, or filled with empty cider cans. Iron hooks for camouflage netting still poke from the roofs. The Kent Archaeological Society's Victor, surveying a pillbox in Tonbridge, noted with quiet pride that it was bubble-level and vertical, showed no sign of having been dislodged, and was only superficially damaged at the firing apertures. They built them to last a week of bombardment. They are lasting forever.

This is what makes them hauntological in a way that, say, an abandoned Victorian railway viaduct is not. A viaduct is a thing that worked, then stopped working. A pillbox is a thing built for a future that never arrived. The men who poured the concrete in the wet summer of 1940 believed, with reasonable confidence, that German armour would shortly come along that road. Most of those men did not live to know how precisely they had been wrong, and how completely. The boxes are a frozen anticipation, a fortified flinch. The country built itself a defensive crouch and then quietly forgot it was crouching.

There is also the matter of category drift. Henry Wills published his survey Pillboxes: A Study of UK Defences in 1985, and the Pillbox Study Group has been recording sites since the 1990s. In 2003 Historic England (then English Heritage) issued Power of Place, the document that finally reframed the boxes from eyesores to be cleared into part of the historic landscape worth keeping. The bureaucratic sentence had taken sixty years. By then, half of the originally-surveyed boxes had already been lost, mostly to motorway construction, gravel extraction, and the steady incremental tidiness of farmers who wanted their fields back. Listing came late, as listing usually does in Britain.

What the surviving boxes carry, more than military history, is a kind of bureaucratic embarrassment. Nobody quite knew what to do with them in 1946, and nobody quite knew in 1976, and the default position of the British state when it doesn't know what to do is to leave the thing alone and let weather and ivy do the work. This is also how we treat asbestos garages, redundant village telephone exchanges, and the sort of public information films that arrived in primary schools without anyone deciding they should. The pillbox is the same gesture in concrete: an institutional shrug that lasts seventy years.

You can still walk a stop-line. Bring an Ordnance Survey map and the Pillbox Study Group's locations, and follow the Wey or the Stroudwater or the upper Thames. The boxes appear roughly where they should. Some have plaques. Most do not. Sheep wander in and out of them. Children climb on them. The country has metabolised them so completely that the question of what they are for has stopped being asked, which is, I think, the only honest answer it could have arrived at.