Plutonic Rainbows

From London this evening the Moon covers 91.3% of the Sun, with the partial phase running 18:17 to 20:06 and maximum at 19:13 BST. That sounds like most of an eclipse. It isn't. The last sliver of photosphere keeps the sky blue and the shadows sharp, and without a forecast telling you to look, you might miss it entirely. Cardiff and Belfast reach 93%, Dublin 94%, and not one of them sees the corona.

Madrid and Barcelona make the point harder. Both fall a fraction outside the path and settle for 99%, and the corona stays inside the path. Totality lands on Greenland, western Iceland and northern Spain, plus a corner of Portugal, and the longest anyone gets is 2 minutes 18 seconds, 45 km off the Icelandic coast. Spain gets it low and gold just before sunset, which is either the best version of the thing or a good way to let one hill ruin your evening.

Fifteen million people will stand inside the path; another 283 million get 90% or better. The second group isn't having a smaller version of the first group's evening, which is the whole reason anyone books a flight to Bilbao for two minutes of dark. Europe's last mainland totality was 1999, and Britain's next is 2090.

Chloé Bakalar left OpenAI in July, about eleven months after joining from Meta, where she had been chief ethicist since 2019. The Financial Times reported it this week, and by that account she was the company's only dedicated ethicist, with no replacement named.

OpenAI's response is the part worth keeping. Asked to confirm there is now no ethicist on staff, a spokesperson said AI ethics "doesn't live with one owner or team" and that ethical considerations are "deeply embedded into the model building process driven by a number of teams across research." As a principle that's fine, and probably right; a lab where ethics is one person's job has already lost. As a description of who actually does what, it can't be audited from outside. The principle also arrived only once the owner had gone, and the job itself was created in August 2025.

The exodus framing built up around this deserves less confidence than it's getting. The FT piece is single-sourced, neither Bakalar nor the company has given a reason, and her LinkedIn still lists the role, so even the shape of the exit is thinner than the coverage around it. Reading a departure as a warning shot, when nobody involved has said anything, produces a story that writes itself and then gets cited as though someone reported it.

What needs no motive is the shape of the org. Superalignment stopped existing as a standalone team in 2024, Mission Alignment was disbanded earlier this year, and its former head Josh Achiam left in July, the same week as Safety Systems head Johannes Heidecke. Each of those sat outside the teams shipping the models. Chief Research Officer Mark Chen told Wired the goal now is safety work integrated with frontier-model development, earlier and closer to launch decisions.

That might genuinely be the better design, and the distinction worth holding onto is between responsibility and authority. Spreading responsibility across every team is good practice. Authority to stop a launch is a different thing, and it has to sit somewhere a deadline can't reach it. The behaviour that needs catching is rarely the behaviour anyone asked for, which is why whoever raises it needs a job that survives the raising. A reviewer inside the build shares the shipping deadline; a dedicated one doesn't, and that's most of the point of having one.

Anthropic will start marking what Claude produces. Text gets an imperceptible watermark woven into the words themselves, and supported files (.svg, .png, .jpg) get signed C2PA provenance metadata. Models launched in the EU on or after 2 August 2026 carry it from launch, older models are being retrofitted during the grace period the law allows, and the marking is applied at the model level worldwide, across the API, Claude Code, Cowork and the cloud partners. None of it is optional and none of it is visible.

The driver is Article 50 of the EU AI Act, in force since 2 August, which requires machine-readable marking of synthetic content that is "effective, interoperable, robust and reliable as far as technically feasible". Anthropic signed the accompanying code of practice, and the code is voluntary while Article 50 is not. Marking obligations are arriving through other doors too. California bundled watermarking guidance into its procurement rules back in March.

The C2PA half is close to decorative. The manifest comes off with a re-save, a format conversion, a screenshot, or an upload to a platform that rewrites metadata on the way in, and The Verge notes it is sometimes stripped by accident. Marks written into the pixels themselves hold up far better. OpenAI's checker was solid enough that people used it two days ago to identify an anonymous model on Arena. Metadata riding alongside a file has never managed that.

The text watermark is the consequential half, and it is the one with nothing published behind it. Anthropic doesn't name the scheme, hasn't released technical documentation, and hasn't shipped the detection tool it says it is working on. The properties it does claim are that the mark rides inside the text, so copy-paste carries it, while heavy editing, paraphrase or translation can leave nothing detectable, as can a passage that is simply short.

Those two properties invert who gets caught. C2PA fails indiscriminately, coming off for anyone who re-saves the file. The text mark is selective, and selective in the wrong direction: someone pushing Claude's output through a second model to pass it off as their own clears it without trying, while someone who pasted Claude's tidy-up of a paragraph they wrote themselves keeps it. The careful user stays marked.

Anthropic is straight about the limits. A detection means content may have been processed by Claude, not that Claude wrote it, and no detection proves a human did. That is the sentence that matters, and it sits in a help-centre article. Detection tools are promised rather than delivered, so nothing has gone wrong yet. Once they arrive, the result will be read out in misconduct hearings and HR meetings by people who never opened the caveat, treated as a test result rather than a hint, the same way plagiarism-checker percentages got treated. Marking the output is the right call. Shipping the marks a long way ahead of any means of interpreting them is not.

An anonymous image model called mona-lisa-1 turned up on Arena today, and several people independently flagged it as OpenAI's within hours. @AiBattle_ posted the first sighting this morning, having run the outputs through OpenAI's own verification tool and got a hit. That matters more than it sounds: OpenAI watermarks its images with SynthID via a partnership with Google, and the checker looks specifically for a watermark originating from OpenAI. One commenter noted that the other accounts passing the story around had left that detail out.

So the provenance is reasonably solid. What the model actually is stays wide open, and the early reads are strikingly muted. @synthwavedd ran comparisons through the day and called it a slight improvement in realism and noise over GPT-Image 2, not a big jump. @WolfRiccardo posted raw outputs two hours earlier and framed it as rumour. On Reddit the verdict was that "it doesn't seem deeply different".

Blind arena testing won't settle it. Everyone is running different prompts with no seed control and no fixed comparison set. @synthwavedd reads the "-1" suffix as a promise of further checkpoints, which fits April, when three tape-themed codenames ran on Arena before shipping as GPT-Image 2 on the 21st.

No outlet has covered mona-lisa-1. For now it's a few thousand views on X and a Reddit thread that can't tell whether it's looking at anything new.

Two more frames have surfaced from the Alberta Ferretti autumn/winter 1992 show in Milan, alongside the five I wrote about in July. The question I keep being asked about this collection is whether it laid any groundwork, and the honest answer is narrower than either yes or no: nothing in these clothes anticipates the house Ferretti became. Whether it mattered at the time is a separate question, and one I'll come back to.

What the name means now is settled. Lisa Armstrong, writing in the Times, pinned the signature as a kind of Valentino look for a younger, slightly more bohemian woman, and the house itself sells that reading back in the language of chiffon, slip dresses and lace. Transparency and drift. It is the register that put Uma Thurman in crimson at the Oscars in 2000 and never really left the red carpet afterwards.

None of it is in these pictures. Helena Barquilla comes down the runway in a red wool jacket cut short and nipped, the shoulder built up, chinchilla at the collar and swallowing both cuffs so her hands emerge from grey fur in black leather gloves. Under it run brocade trousers in the same red, gold urns and cartouches printed the length of the leg, and the shine on them is doing all the work the jacket refuses to. That's the third chinoiserie piece I've now seen from this one show, after the mandarin-collared brocade jacket and the beaded suede column, so the theme was running properly through the collection rather than turning up once as a novelty. Heavy, opaque, structured, warm. The opposite of drift.

Whether Milan noticed, I can't tell you, and the reason is itself worth recording. Vogue Runway's archive does reach back to 1992, and its entire fall season that year comes to three houses; Ferretti is not among them. The Fashion Model Directory indexes the show and then answers 403 on every page that would name the cast, and Bloomsbury's catalogue entry sits behind a subscription wall. What survives in the open from these years is the business rather than the clothes: the Milan showroom at Palazzo Donizetti in 1994, New York two years after that, Bergdorf Goodman and Neiman Marcus by mid-decade. An expansion is easier to record than a collection.

So the clothes are free to be of their moment. The purple suit is period Milan almost to the point of parody: violet bouclé jacket and short skirt, a magenta leather shell zipped to the throat underneath, black suede gloves again, and a belt of gilt baroque medallions wide enough to be structural. Gold hardware over bright wool was general across the seasons either side, Krizia included. Set against the chiffon that came later, the violet and the gilt read less like an early draft than a different house altogether.

Two greens, and they are not meant to match. The storm coat is coated to a wet shine, cut wide enough through the sleeve to swallow whatever sits underneath, and fastened with gold snaps instead of buttons so the front stays flat. The double-breasted jacket beneath it lands in the same pistachio register but matte, tailored close, four gold domed buttons running down it. Wearing the two together is the proposition: one surface throws the studio lights back and the other absorbs them.

Under the greens Gail Elliott has powder blue on powder blue, a silk shirt with a white contrast collar and flapped patch pockets buttoned to the throat, over trousers in the same blue, dry and heavy. No print anywhere on the body. The pattern is quarantined into two scarves, one worn as a headband, one knotted at the neck, both in the same blue-green-orange floral. Six months later the winter book would put her in tartan and carriage-print silk from the collar down.

Then the gold: a plaque buckle with two hearts flanking a capital E, baroque earrings set with blue stones. Escada spent the eighties making gold the argument rather than the trim, and on this page that is where the volume has gone. The fabric behaves. The hardware does not.

In April 1993 the chief designer Michael Stolzenburg told the Los Angeles Times that "Escada will do no grunge", and sold retailers an easier, more youthful autumn, on the grounds that younger customers no longer felt it was the right time to be showy. Some of that easing had already happened in the cloth before he described it, and the autumn book then arrived in tartan and gold medallions anyway. Nobody was going to mistake this coat for a flannel shirt. But the restraint the eighties never managed is in the plain blue trousers, not in anything the house said about them.

In August 1993 I took a coach to Bristol to see Jurassic Park. I bought the fragrance first, a 50ml refillable bottle of Pasha de Cartier from the counter in Debenhams, then carried it in its bag to the long queue on Frogmore Street, where the cinema was a Cannon in those days. The bottle was expensive in the way that turns a purchase into an event. Silver metal over thick glass, gold fittings, a blue cabochon at the cap, and the promise folded into the word refillable: buy the vessel once, renew the contents for years.

I tried it in the shop and thought I liked it. At about eight that evening I smelled it again in the kitchen, decided I didn't care for it at all, and put the bottle and the box in the bin.

Cartier had launched Pasha the year before, in 1992, with the refill built into the design, and had the shape registered as property: US design patent D348,333, filed in April 1992 by Alain-Dominique Perrin and Jacques Diltoer. A design patent guards ornament against copying rather than promising durability, but nobody files one unless they expect the object to matter for years. Mine lasted about nine hours.

The money isn't the hard part. The bottle was destroyed while the world it came from still looked permanent. Debenhams was simply Debenhams, Pasha was a current fragrance rather than a vintage one, and August 1993 wasn't August 1993 yet. It was now, and nothing around me announced itself as temporary.

Twenty years later, in 2013, I found another full refillable bottle on eBay in Florida and bought it. It arrived across the Atlantic improbably intact, and for a moment the logic was seductive: the mistake corrected, the lost thing returned. Same architecture, same weight, the same gadrooned silver and gold. But it wasn't on the counter in Debenhams that day, it didn't go into the queue on Frogmore Street, and it has no memory of August 1993. It carries instead twenty years of unknown existence in Florida, a parallel biography, while the bottle I actually bought went into a kitchen bin.

Both buildings have stopped being what they were. Debenhams closed its last 49 stores on 15 May 2021, Bristol among them after fifty years on Broadmead, and the site is becoming flats, which is at least a definite ending rather than a slow one. The Cannon stopped showing films before the decade was out, so the room where I queued is an O2 Academy now.

I remember it being dark outside when I dropped the bottle in the bin, and it cannot have been, which unsettles the whole account. Bristol doesn't lose the light in early August until around a quarter to nine, and I was finished by eight. The dark is something the memory added later, and once I noticed that I stopped trusting the rest of the scene. Had I known what the bottle would come to mean I'd have kept it obsessively, but holding that knowledge would have meant not living in 1993 at all, only visiting.

I still don't like the fragrance. What I bought back from Florida was never the smell, it was the afternoon, and the afternoon isn't in there. So the second bottle sits in its sealed box, an object I don't want to use and can't throw away. The one promise the design kept is the one I'll never take it up on.

Wednesday morning, July 1994. The Saint Laurent couture show opened to Berlioz and Wagner, arranged by Pierre Bergé, with a printed warning against plagiarism on the invitations. Six months earlier he had shown sixty-one numbers, one of the shortest collections of his career, assembled while recovering from double pneumonia, and taken his customary walk down the runway with Lucie de la Falaise. The July show was therefore being read for signs of life before a single exit had left the runway.

The waist was the news. After several seasons of women dressed out of the men's wardrobe, the torso came back as the thing every garment got built around, and the trade press ran the show under "Yves leads the Fit Parade". Jackets nipped in and stayed narrow at the shoulder. It opened on sass rather than severity, though: minidresses, belts slung wide across the hips, hats WWD could only call wacky, thigh boots pulled so high they read as tights.

The same discipline runs through the outerwear at a different scale. Silver fox over black suede, the fur running the full length of the front so the coat reads as two grey columns with a dark seam between them, and enough of it at the cuff to double the width of the sleeve. Gail Elliott, who walked this collection, wears it with her hands buried in the pockets and a beret cocked hard to one side, finished with a black fabric flower. Nothing about it is casual.

Then he gave the room a jolt he hadn't managed in years: three black tuxedo suits cut over the shortest skirts imaginable. Hemlines elsewhere went up and down with no governing rule, knee-skimming, short and belled, long and columnar, because he liked going back and forth to satisfy his clients. That is a different instinct from the Valentino anniversary collection three years earlier, where every proportion was argued from a single position. Saint Laurent was offering a wardrobe, not a thesis.

Evening headed East, or at least toward what Paris in 1994 meant by East. Coats in electric colours that WWD thought fit for a Mandarin empress, worn over straight satin and velvet gowns that do nothing underneath but hold still. He took the ovation at the end looking heavier and steadier than he had in some time.

If Astra lands next week, my guess is that Anthropic answers by not answering. No counter-post, no teaser, no rival list of solved problems. The habit is to name a model and make it available the same day: Claude Opus 5 on 24 July at $5 and $25 per million tokens, live on every platform immediately, and Fable 5 and Mythos 5 generally available in the API on the June day they were named.

That habit argues against teasing, though, not against announcing something else loudly. Anthropic has run the same theatre before: it called Mythos too dangerous to release, and Mashable was right that the warning doubled as marketing. A safety-framed capability post would fit that pattern well, and would land in the same news cycle without naming Astra once.

I still don't think it needs one. Astra has no published price, no benchmark table and no outside testing, and even the launch date is a leak rather than an announcement. The comparison is a shipped product against a research claim, and that one makes itself.