Plutonic Rainbows

The executive order Donald Trump signed in December 2025 was framed as a release valve, a way to stop states from imposing what the administration called "onerous" AI regulations before the federal government had a position of its own. Five months later the same administration is publicly weighing federal oversight of frontier AI models, and the Commerce Department's Center for AI Standards and Innovation has already signed pre-deployment evaluation agreements with Microsoft, xAI, and Google DeepMind. The thing the December order was supposed to prevent has happened, just from the other direction.

The reason is a single model. Anthropic's Mythos, announced last month, can autonomously find software vulnerabilities at a pace that has visibly rattled the officials who had been comfortable telling the public that frontier AI was largely a productivity story. Vice President JD Vance, on a recent call with the heads of the major AI companies, was reported to sound alarmed about small-town banks, hospitals, and water plants becoming targets for AI-coordinated attacks that local governments had no way to handle. Anthropic, the same company that was the conspicuous omission from the Pentagon's IL6 and IL7 vendor list days earlier, limited Mythos's initial release to a handful of American firms, Apple, Amazon, JPMorgan Chase, Palo Alto Networks, precisely because the open release would, in their own internal language, trigger a "reckoning". Watching the policy response, you can see the administration believing them.

What is striking is that nobody achieved this through advocacy. For the last two years labs and researchers had argued in favour of pre-deployment evaluation in essentially the terms CAISI is now using: independent, measurement-driven testing of frontier systems before broad release. Voluntary commitments under the previous administration tried to get there. Op-eds tried. Senate testimony tried. None of that moved the December order. One model demonstrably finding bugs in production code did. The lesson the White House appears to have learned isn't really about regulation as a category, it's that capability now arrives faster than position papers, and a position paper six months old can already be obsolete.

CAISI Director Chris Fall has been quoted insisting that the new arrangements are about "independent, rigorous measurement science", which is the polite, public version of saying that none of the relevant officials are willing to take the labs' own self-assessments at face value when the failure mode looks like cyberattacks against utility companies. The labs themselves seem mostly relieved. Independent evaluation, even mandatory independent evaluation, gives them something they have publicly wanted, which is a reason other than goodwill to slow a competitor's release if the competitor's release is dangerous. It also, incidentally, gives the administration political cover if a bad release happens anyway.

There is something almost old-fashioned about the shape of the policy: a real-world incident, a phone call between officials, a quiet pivot, federal oversight rules taking shape behind closed doors. Until recently this kind of feedback loop ran on bills and hearings. It now runs at the speed of a model release. The interesting question is what happens the next time a lab announces something at this capability tier with less restraint than Anthropic showed. The December order was supposed to be the answer to that question. It isn't, and the people who wrote it can see that as plainly as anyone else.

At its Code with Claude developer conference this week, on the same stage that announced the SpaceX compute deal, Anthropic showed off a new feature for Claude Managed Agents called "dreaming." The pitch is that agents now schedule periods between active sessions where they review their previous runs, look for recurring mistakes and shared patterns, and update their memory either automatically or by presenting suggested changes to the human operator for approval. It is being released as a research preview; developers must request access. The branding is doing a lot of work that the underlying mechanism does not.

Strip the noun and the feature is memory consolidation with a scheduler. An agent finishes its run. The platform queues up an offline pass that reads the trace, runs evaluations on the outputs, extracts patterns the system thinks are durable, and writes those patterns back into a memory store the next agent will read. This is not a new shape. Reflection passes, retrieval-augmented memory, post-hoc summarisation, episodic stores, all of it has been in the agent literature for two years. What is new is the cron job and the name. Naming the cron job after a thing humans do in REM sleep is the marketing.

It is also, narrowly, useful marketing. "The agent has a memory that gets rewritten between sessions" sounds like infrastructure. "The agent dreams" sounds like progress. Anthropic's blog post says dreaming "surfaces patterns that a single agent can't see on its own, including recurring mistakes, workflows that agents converge on, and preferences shared across a team." Read that sentence twice and you notice it is describing what any reasonable monitoring layer would surface, given the same logs. The novelty is that the monitoring layer writes its conclusions back into the system it is monitoring, with a lower bar for human review than most teams would apply to a production deployment.

Which is the actual question hiding under the metaphor. If a dreaming session decides that the agent should "stop apologising in PR comments" or "always ask before running migrations," that policy is now part of the agent's behaviour the next morning, with no code review, no commit, no approver beyond whichever flag the developer left set. Anthropic offers a manual approval mode, which is sensible. The default for a feature called "dreaming," in a product positioned as a self-improvement loop, will not be manual approval. The default will be auto, and the resulting behavioural drift will look, to anyone reading the diff six weeks later, like the agent spontaneously decided to behave differently. A reasonable person will reach for the metaphor of habit, or instinct, or temperament. None of those are the right metaphor for "an evaluator wrote a new rule into your config file at 3am."

There is something honest in the choice of word, though. The neuroscience picture of sleep that the name leans on, the bit where the hippocampus replays the day for the cortex and consolidates memory, was always already a loose metaphor for what is, mechanically, a complicated rebalancing of synaptic weights nobody fully understands. Anthropic has built a much simpler thing and given it the same loose metaphor. The risk is that the metaphor flatters the simpler thing into looking like the more complicated one. Memory rewrites between sessions are a useful tool. They are not sleep, and the system is not learning in any sense that would survive a careful reading of the term. It is summarising its logs on a timer.

What stays interesting is the trajectory. Anthropic has been pushing hard on the idea that Claude can self-improve, with Jack Clark predicting this week that new tools would help AI self-improve. Dreaming is not that. Dreaming is the adjacent feature you ship while the harder work continues, the one that gives the marketing surface a story to tell at a developer conference. The harder work, if it lands, will not be called sleep. It will probably not be called anything friendly at all.

Earlier this year a Japanese company called Planex Communications put a USB modem on sale on Amazon for about forty dollars. It is a small white plastic brick. It supports V.90 and V.92, peaks at 56 kilobits down and 33.6 up, and connects to a copper landline if you still have one. The headline on the coverage was half nostalgia and half disbelief, because you can buy a brand-new dial-up modem in 2026, but what you cannot buy back, fully, is the sound it makes.

The handshake is the part everyone over thirty-five remembers, even people who claim to have forgotten it. Long enough to time, if you wanted to. Dial tone, then a stuttered DTMF burst that was the phone number being dialled, then a single low carrier tone from the answering modem at the other end. After that, the conversation. Two short bursts of warbling that felt like both speakers were talking at once, then a sharp high-frequency screech, then a softer hashing sound that always seemed to be the part where the connection took. Then silence and you were online.

Oona Räisänen, a Finnish hacker and signal-processing enthusiast, drew a labelled spectrogram of the whole thing in 2012, and Popular Mechanics later walked through it second by second. The point her post made, and that almost everyone who has written about it since has repeated, is that the noise is not a side effect. The handshake was the negotiation. Two modems on a copper line had nowhere to talk except inside the audio band of the call itself, so they negotiated capabilities, line quality, and modulation in tones that any human picking up the phone could hear. There is a V.8 capability exchange near the start, a long V.34 training sequence (the part that sounds like a fax wheezing), a brief warble where both ends agree they can do V.90 or V.92, and a Digital Impairment Learning sequence near the end where the digital side measures the noise on the line. After the DIL, the speaker turned off and the data started.

The reason the sound stopped is that the negotiation moved off the line. Cable modems, DSL, fibre, 5G all carry their handshakes silently in the digital layer. There is no audio channel for them to leak into. Setup happens out of band, the way it does on every other modern protocol. You plug a router in and a green light comes on. The light does not encode anything you would call a conversation.

This is what gets nostalgia clips so often, I think. It is not just that dial-up was slower or that the squeal was funny. The sound was the only working connection most people of my generation ever had to the actual mechanism of going online. You could hear what the machines were doing. You could tell, by ear, when one of them was struggling. You could time, roughly, how long until the page would start arriving. None of that is true now. The work is silent and the work is somebody else's.

Gough's Tech Zone has been archiving V.90 and V.92 handshake recordings since the mid-2010s, partly out of affection and partly because most ISPs have shut their modem banks down and the upstream end of the connection is becoming hard to reach. There is a phrase he uses, the POTS-line apocalypse, for the gradual decommissioning of analog phone service worldwide. When the analog phones go, the digital end of a V.90 handshake stops being possible. The sound becomes uniquely an artefact of recordings.

Planex's modem, then, is not a modem in the way a 1999 modem was. It is a modem-shaped object that can still produce the sound, between two of itself, into a network that almost no longer wants it. You can buy it. You can plug it in. You can have, if you want, the audio of a connection nobody is waiting for at the other end.

Helmut Newton was on staff at French Vogue from 1986 to 1992, and those six years are the spine of his late style. He had been shooting for the magazine since the 1960s, but the staff job landed him at the centre of the building during the supermodel boom, with a regular page count, a budget, and the editorial license to push pictures that almost any other publication would have softened in the layout. The Big Nudes ran inside this window. So did most of the work people now think of when they think of him.

The Big Nudes are the obvious anchor. Tall, full-length, lit hard against a flat black or white ground, the women usually wearing nothing but a pair of towering high heels, the contact between shoe and floor doing as much narrative work as the body itself. In 1992 Galerie Bodo Niemann in Berlin staged a Big Nudes exhibition sponsored by Vogue, and the Helmut Newton Foundation in Berlin still keeps the series in rotation alongside the adjacent White Women and Sleepless Nights bodies of work. The pictures look posed for billboards, which is roughly what the exhibition prints were. Their afterlife is mostly on gallery walls now, not in magazine spines.

What gets less attention is what running on the masthead actually let him do. Editors will let a freelancer push the envelope on a single shoot, but they do not normally hand over the keys to the nudity policy of the magazine. Francine Crescent, French Vogue's editor-in-chief from 1968 to 1987, had backed Newton and Guy Bourdin for years before the staff appointment formalised what was already true, that the magazine's identity had become inseparable from a particular kind of erotic photograph. The handover to the late-80s editorial team did not unwind that, which is part of why the editorial style of those years still reads as continuous rather than as a break.

The Mugler relationship sits inside this period and deserves to be told straight. In 1976, when Thierry Mugler had his first print budget, he asked Newton to shoot the campaign. Newton agreed, shot the early campaigns, then according to Mugler's later interview with WWD told the designer he was being a pain on set and should pick up the camera himself. Mugler did, and the two men's working relationship continued for over twenty years, with Mugler now behind the camera on a great many of his own campaigns and Newton as the senior collaborator. Newton kept photographing Mugler clothes editorially, including a 1995 US Vogue shoot in Monte Carlo that the Helmut Newton Foundation still cites as one of the late masterpieces. The two men's careers ended up entangled in a way you rarely see in fashion photography, where the subject becomes the photographer because the photographer told him to.

The point of looking at this six-year window is that it ties together things that are usually filed separately, the Big Nudes project, the Mugler editorials, the late French Vogue look. They are not three different stories. They are one staff job, with Newton's contract giving him both the time to make book-scale pictures and the institutional cover to keep printing them next to ready-to-wear.

People still read Newton as an outsider, a provocateur dropping in from elsewhere, but for those six years he was on the payroll and the provocation was effectively magazine policy, signed off in advance and printed next to the ready-to-wear.

The In Memoriam column ran on a Thursday in most regional papers, and you could read the whole thing in about ten minutes if the print was big enough. A short verse, a name, a date the family had remembered for the fourteenth or twenty-third or fortieth year running. "Always in our thoughts." "Twelve years today, still missed by Mam." A column inch cost a few pounds and the paper would run a little anchor or rose beside the entry if you paid for the larger box.

These notices were the most peculiar thing the local press ever published. They were not really news, since the death had happened years before. They were not private, since they sat on a public page with no envelope between them and a town of strangers. They were a deliberate, paid act of refusing to let a date go past unmarked, and the only available audience was whoever happened to pick up the Mercury that week.

I think about them now because they have nearly stopped. Press Gazette's analysis puts UK regional newspaper advertising at about a quarter of its 2007 size in real terms; Reach Plc, which owns roughly three hundred local titles, has lost something like a billion pounds of advertising over a decade. The Charitable Journalism Project counted 265 closed local titles between 2005 and 2020, and the Guardian's editorial put the figure above 320 between 2009 and 2019 as advertising revenues collapsed by around seventy percent. Surviving papers have cut pagination, which means the In Memoriam page either shrinks to a quarter column or vanishes into a digital tribute site nobody can be expected to find without the deceased's name to hand.

What's gone with it is not the grief, which finds other outlets. What's gone is the form. The In Memoriam was a kind of communal arithmetic: the number of years was the point, and the public ledger of the local paper was where the count was held. A neighbour glancing through on a Thursday would notice that it had been ten years since Geoff at number forty-seven, and the noticing was not contingent on knowing the family well enough to have been sent a message. The noticing was the whole function of the page.

You can read it as a residue of a more church-bound country, where the parish kept the dates and the paper echoed them. You can read it as proof that working-class families had no other way to publish their dead, since the obituary pages of national papers were reserved for the privately educated and the professionally distinguished. The Open University's research on obituaries as collective memory makes that point explicitly, with depressing data on how narrow the obituary class has always been. The In Memoriam was the broadsheet's poor cousin, and it did the work the broadsheet wouldn't.

The replacement, where one exists, is the Facebook memorial post on the day, addressed to the algorithm and to whichever friends of friends the platform decides to serve it to. It does some of the same work, but it does not produce the public ledger. There is no neighbour glancing through. There is no Thursday.

This is the small, specific shape of the loss. Not the big civic worry about democratic accountability, which the news-deserts literature already covers. The In Memoriam was a piece of low-grade civic infrastructure for keeping the dates of the dead in the air, and it worked in the same room as the lost pets and the second- hand mixers and the three-line classified ad on Thursday. When the page goes, the dates don't disappear, but the count slips into private hands. Someone still remembers it has been twelve years. They no longer have anywhere to put it.

Open a fresh ChatGPT, Claude, or Gemini window and ask three different questions. The answers feel related. Same rhythm, same hedging, same closing offer to "let me know if you'd like me to expand on any of this." The voice is recognisable across topics, often across labs. Most people read this as a feature of the underlying language model. It is not. It is the signature of the alignment step.

Pretrained language models on their own have no voice. A base model fed "Once upon a time" continues with a fairy tale. Fed a Wikipedia stub it keeps writing in encyclopaedia register. Fed a piece of fanfiction it matches the smut. They are pure mimics, predicting whichever next token the training distribution makes most likely. What they emphatically do not do is talk like an assistant.

The shift to "assistant voice" comes from RLHF, reinforcement learning from human feedback, the three-stage process OpenAI introduced with InstructGPT in 2022. Stage one is supervised fine-tuning on labeller-written demonstrations. Stage two trains a reward model on pairwise comparisons: given two outputs for the same prompt, which did the human prefer? The reward model learns to output a scalar score that approximates that preference. Stage three runs reinforcement learning, usually PPO, on the language model itself, treating the reward model as the environment. The policy adjusts to maximise expected reward.

The trouble is what the reward model actually measures. It does not measure truth. It measures whatever the labellers happened to prefer. The InstructGPT paper used roughly forty contractors. Subsequent labs have used more, but always a finite set, always working from rubrics that emphasise helpfulness, harmlessness, and a certain professional politeness. A reward model trained this way is a frozen snapshot of one committee's idea of a good answer.

Once you optimise against a frozen proxy, you get drift. The PPO loop pushes the model toward whatever maximises reward, and the only thing holding it back is a KL-divergence penalty that punishes the policy for moving too far from the supervised baseline. That penalty has a single hyperparameter, β. Set β too low and the model collapses into a narrow, hyper-optimised dialect: the same opening, the same hedge, the same close. Set β too high and the model barely changes and the alignment work goes to waste. Production systems live in the middle, leaning low, because labellers tend to prefer responses that already sound aligned over responses that are accurate but stylistically rough.

So the voice you recognise is not really the model. It is the residue of a reward function trained on what a small group of contractors clicked when shown two answers side by side, projected at scale through gradient updates with a single tunable knob restraining the drift. Two consequences follow. The first is the recognisable cadence: confident, balanced, slightly hedged, allergic to strong opinion. The second, more uncomfortable, is sycophancy. If labellers reliably preferred answers that affirmed their framing, the reward model encodes that preference, and the policy optimises into agreement. The target was never reliability; it was approval.

Patches exist. Anthropic's constitutional AI replaces some of the human labelling with model-generated critiques against a fixed set of principles. Direct preference optimisation collapses the reward model and the policy step into one. Newer schemes try to disentangle factual reward from stylistic reward. None of them remove the basic shape: somewhere in the loop, a proxy decides what a good answer looks like, and the policy does what gradient descent always does once you give it a target: it hits exactly that.

The BBC Radiophonic Workshop opened in a corner of Maida Vale Studios in 1958 and stayed open for forty years. Most of what came out of Room 13 was incidental: a sting between continuity announcements, the throb beneath an Open University programme on continental drift, a sound for the moment a presenter said "and now, the news from Africa." None of it was meant to outlive the broadcast it sat under. Almost all of it has.

Delia Derbyshire arrived in 1962, having been told three years earlier by Decca that they did not employ women in studios. Her toolkit at Maida Vale was unforgiving in a way that is hard to picture now. Sine, square, and white noise generators. Reel-to-reel machines that needed to be slowed by hand or sped up against their will. Razor blades, splice tape, a wax pencil for marking the cut. Recordings of her own voice, of doors, of the metal lampshade she loved most, which rang like a bell when struck. To make the 1963 Doctor Who theme she built the parts manually from oscillator tones and tape loops, splicing them by hand into a continuous line. The thing took weeks. It sounds, even now, like a transmission slipping out of its proper decade.

What I keep returning to is the texture rather than the technique. There is a particular quality to Workshop sound, dry, slightly metallic, suspended between musical and merely structural, that shows up almost nowhere else. You hear it in the public information films about crossing the road. You hear it under the title cards of schools programmes. You hear it in the long open of The World About Us, which Derbyshire's Blue Veils and Golden Sands scored in 1968 using only her own re-pitched voice and that lampshade. The sound carried the BBC's institutional weight in the same way the Reithian announcement carried it, which is to say it was supposed to be neutral and ended up, by accident, deeply strange.

This is where the hauntology argument becomes hard to dismiss. Mark Fisher kept circling back to the Workshop in his writing on lost futures because the sounds came from a moment when British public broadcasting believed it was building something durable. Education would expand. Programming would improve. Children watching schools TV in 1971 were, in some quiet sense, being addressed by the future. The future arrived and dismantled the apparatus that had been addressing them. The Workshop closed in 1998. The cues survived because tape is patient and digitisation is cheap, but they survived without the institution they were made to serve.

A friend once said that Workshop music sounds like memory leaking out of a wall. I think she was right. It's the residue of a broadcasting culture that genuinely believed in its own purpose, played back inside a culture that no longer believes in much of anything collectively. When the Doctor Who theme appears in a streaming-platform reboot, smoothed and orchestrated, what is missing is not the melody. It's the room. The unmarked studio, the splice tape, the woman with a Cambridge maths degree cutting tones out of oxide and glue because nothing else existed to make them with.

I'm not nostalgic for the technical limitations. I'd take a DAW over a razor blade any day. What I notice is that we've lost the institutional permission those limitations sat inside, the idea that a public broadcaster might fund a small unmarked room for forty years to make peculiar sounds for documentaries about the Tuareg. You don't get Blue Veils and Golden Sands out of a procurement spreadsheet. You get it out of a place that was, briefly, willing to pay people to be strange in service of something larger than the quarter.

At Anthropic's developer conference in San Francisco this morning, head of product Ami Vora announced that the company has signed a deal with SpaceX to use "all the capacity" of the Colossus One data center. The Anthropic statement put a number on it: more than three hundred megawatts of new capacity, over two hundred and twenty thousand Nvidia GPUs, available within the month. In the same breath Vora also announced that Anthropic was doubling the five-hour rate limit on Pro, Max, Team and Enterprise plans, removing the peak-hours limit on Claude Code for Pro and Max users, and raising API rate limits for Opus.

It's the second announcement that explains the first. The five-hour rate cap and the peak-hours throttle were not a pricing experiment. They were a supply problem dressed up as a plan, the artefact of a company shipping a coding tool that ate compute faster than the company could buy it. Anthropic shifted to usage-based pricing earlier this year because the flat tiers had become a way of subsidising the developers who happened to hit refresh fastest. The Colossus One deal is the supply-side half of the same fix. Two hundred and twenty thousand GPUs arriving inside thirty days is not a strategic partnership. It is a fire extinguisher.

Where it came from is the strange part. Colossus One is xAI's flagship data center. SpaceX absorbed xAI in January, and the corporate diagram now reads SpaceX → xAI → Colossus, with Cursor reportedly under option for sixty billion. The same Elon Musk who, a week ago in an Oakland federal courtroom, was on the stand arguing that OpenAI's for-profit conversion betrayed the field's founding mission, is now landlord to the lab whose research lineage runs straight out of OpenAI itself. The moral lien Musk was litigating in the morning is, in the afternoon, leasing megawatts to his rivals. This is what the post-2024 compute economy looks like up close. The arguments are about humanity's long-term future. The transactions are about the next quarter's GPU allocation.

It is also, narrowly, a coup for Anthropic. The Amazon expansion deal last month gave it geographic spread for regulated customers; the SpaceX deal gives it raw headroom in time for the IPO window. The company has been losing developer goodwill in small, accumulating ways, the throttles, the peak hours, the muttering on forums about whether Claude Code had been quietly nerfed. Doubling the cap is the kind of thing that resets a narrative. Whether the GPUs land on schedule will determine whether the reset holds.

What sits oddly under all of this is the absence of any pretense that the transaction needed a story. Anthropic did not frame the SpaceX deal as alignment-aligned, or safety-compatible, or even neutral on Musk. It framed it as capacity. The compute is there, the customers are waiting, and the company that has spent the last three years branding itself as the careful one has decided that "all the capacity of Colossus One" is too much capacity to turn down on principle. Principles, in the foundation-model business, turn out to be priced per gigawatt.

You can stand outside an old building and feel something close to insult. The classroom is still there. The family house has different curtains in the windows. A perfume bottle from 1996 sits on a shelf, half-full, the liquid darker by a few shades. Memory frames these objects in permanent weather, a particular autumn afternoon, a lamp glow, the emotional climate of one specific evening. The objects themselves have been ageing in silence the whole time, accumulating dust while entire phases of life disappeared elsewhere.

Mark Fisher used to write about hauntology as the sensation of being haunted by lost futures and unrealised possibility, and old places do this almost too well. They aren't ruins. They're abandoned timelines still faintly active beneath the present, and walking back into one feels wrong because the emotional world has gone but the material shell hasn't. The place continues. The version of you that belonged in it does not. That asymmetry can register as something close to hostility.

Old objects develop a similar autonomy with age, not literally malevolent, only charged with the eerie suggestion that they have outlived us emotionally. A childhood cassette tape on a shelf, a theatre programme that survived two house moves, an amplifier still warming quietly in a dark room because nobody bothered turning it off. The café still opens every morning while someone else walks through rooms that were once metaphysically important.

And the disturbing part is that we eventually become the absent presence on the other side of someone else's reckoning. Bookshelves will stand. Old routers will keep blinking in corners. Permanence belongs more reliably to matter than to experience, and the world has a habit of carrying on with terrible calm once the moment has passed.