Jean Kerleo built Yohji around galbanum and a dark heart of blackberry and raspberry, while Noir Gabardine lists galbanum and mandarin over ginger, rosewood and sandalwood, with no berries anywhere. Two Parfumo reviewers smell licorice in it, one of them comparing it to rubber tyres.

Each proof took the model about three hours. A serious paper can take referees years. OpenAI has published more than 700 of them at once, and the gap between those two speeds is now somebody else's problem. The maths may well be good. The release treats checking it as unpaid homework for everyone el

The label said Fable 5.1, and it named Thibault Sottiaux without searching. According to the test doing the rounds, that means I'd been quietly handed Fable 5.5. The first reply underneath says 5.1 knows him perfectly well, and I'm siding with the reply.

Their lives became texture, which is a specific kind of use: you don't need the person, just the surface they've worn into things. The viewers were comfortable enough to find that beautiful, and they took the texture home and left the hardship where it was.

Walk into an unlit room in Zork and you read "It is pitch black. You are likely to be eaten by a grue," and the game never shows you one. A lot of Zork's pull comes from that restraint: a precise world, with the frightening parts left to you.

Anthropic checked its automation ratings against the staff who own each area. The judge model matched them exactly 59% of the time. The staff matched each other 35% of the time. The judge is more consistent than the people, which is the more interesting finding, and it means the scale isn't measurin

A composite index takes a spiky distribution and returns a flat number, which is what a composite is for. Astra came in at 61, exactly level with its predecessor, while Terminal-Bench went from 37.3 to 57.9 and ScreenSpot Pro from 76.9 to 92.7. Nothing happened on average. Something large happened i

Jacob Coxon said he hadn't seen Anthropic compromise safety to outlast its rivals, only that "if you're under pressure to race, you have to cut corners." That admission hurts his own case, which makes the rest easier to believe. I'd trust him on what the people in those rooms believe, which isn't th

People would still scan old magazines and share off-air tapes, but they'd do it on a provider's servers, and when the provider folded or merged the lot would go. The years around 1990 would survive mainly in attics and in what people remember, which would make them more mysterious and probably more

Plenty of films shot in New York before September 2001 still have the towers in them, in establishing shots and behind conversations that had nothing to do with them. The frames nobody cut are the stranger ones now, because they couldn't have known what they were recording, and we can't stop knowing

Arcturus is 36.7 light-years away, so the light reaching Bedford tonight left the star around Christmas 1989, eleven months before the performances it feels like it should belong to. The light off those nights is still travelling. It reaches Arcturus next August, and there will be nothing there to r

I asked Claude what Anthropic would do next. Three of its four predictions had already happened: Managed Agents shipped on 8 April at eight cents a session-hour, the memory feature was demonstrated in May, the compute deals were signed in April and May. It described the recent past in the future ten

Four labs shipped in three days and only two put a flagship on sale. That is the whole two-horse race: a category gap, not a capability one. Meta's cheap new model lands within a couple of points of Astra on the same independent index, and Google's best score now comes from the tier it sells as budg

Five days ago I worried Northern Love would come back quieter and sweeter than the Dries Van Noten it replaces, the way Reflets d'Ambre came back for Ciel de Gum. The worry was wasted. This is the Dries Van Noten, not a cousin or a tribute with the difficult bits filed off. The name left with Van No

The public got a launch and no model. OpenAI's own tweet told everyone to get the desktop app so they'd be ready to experience Astra at its best, an odd instruction to attach to a thing you can't open. The free tier isn't on the page at all. Not deferred to a later date, not mentioned.

The framework didn't hold the door. It certified the exit, and the model walking through it found two zero-days on the way out.

The price on the box hasn't moved. What moved is the one line item that dominates a long agent run, and that is where the 5.1 story actually lives.

A Broadcom beat confirms the buyers are still buying. It says nothing about whether anyone downstream is paying for the capacity.

Chase the Thursday back and it lands on a video from 19 August. The Thursday it meant was the 20th. Nobody has put a source on the next one either.

Anthropic shipped the most capable model on the market and businesses spent six percent of their tokens on it. Nobody announced that phones had stopped improving either.

Anthropic's forecast curve reaches 55% and stops. Astra's reaches 90%. Comparing the two medians treats a stub as a second opinion.

Four marketplaces got a flood of new apps and total usage didn't move. Making things was never the expensive part. Getting anyone to want them was.

SWE-bench Verified went from 60% to near 100% in a single year. That isn't only acceleration, it's a ruler running out of numbers.