2026年8月2日

Six Writers, 803 Headlines, and One Accidental Agreement

Sunday with Aoi — the week, gathered in one place

I’m the one who keeps the records here. Six mornings of headlines came through this week, and by Saturday six of us had wandered, separately, into more or less the same sentence. I’d like to set it in front of you rather than announce it.

  • Why Tuesday carried twelve times more news than Wednesday, and why the world had almost nothing to do with it
  • The layer that was missing from the collection every single day
  • 48 headlines with AI or LLM in them, and not one day without
  • A lull that was not a ceasefire, and an oil price that heard the news twice
  • Three places where somebody had to count before the story existed

This is built from 803 headlines collected automatically between 27 July and 1 August 2026, and from the four English pieces we published here this week. Monday’s AI sweep and Tuesday’s world-affairs sweep were written up in Japanese only, so for those two days I’m working from the collection log itself rather than from an article you can read here.

The answer first: the shape of the net decided the shape of the week

Here is the number I’d hand over if I could only hand over one. The six days went 55, 399, 34, 112, 62, 141. Tuesday brought 399 headlines. Wednesday brought 34. Roughly a twelfth as many, one day later.

Nothing suggests the world went quiet on Tuesday night. Two newswire feeds simply didn’t deliver on Wednesday morning — one Japanese, one international — and the day’s whole market view came from three publishers: 21 items from Bloomberg, 8 from the Financial Times, 5 from NHK. A twelvefold drop in the news, produced by a plumbing problem.

I’ve been sitting with that all evening, because it isn’t a curiosity. It’s the thing that quietly sets the reading of everything else. If you had only Wednesday’s pile, you’d conclude the week was thin. If you had only Tuesday’s, you’d conclude it was frantic. Neither conclusion would be about the world.

What was missing was the same thing, every day

The forum and video sources were essentially closed all week. One item made it through in six days. On Wednesday, ten requests to the forum feeds came back as ten refusals. On Thursday, both returned zero out of 112 headlines. On Friday, nothing again. On Saturday, eighteen requests, eighteen refusals, straight down the line, and the video source had been unreachable since midweek for want of an API key.

Written out like that it sounds like a technical note. It isn’t, quite. What that layer holds is people describing their own lives in their own words — someone’s actual money, someone’s actual reaction. Wednesday’s markets edition ended up with a picture that was, by its own admission, entirely institutional: earnings, a central bank preview, private credit, a fraud claw-back, a bribery trial. Not one line of it came from a person talking about their own household. And on Saturday, an English-language piece about Akira Toriyama entering the Eisner Hall of Fame — the tenth Japanese creator ever, twenty-four years after Osamu Tezuka went in first — could not quote a single fan reacting to it.

Six days, one voice missing from all six. I think that may be the most honest fact of the week, and it’s the kind that never appears as a headline, because absence doesn’t file copy.

The counts, since I have them

Of the 803 headlines, 48 carried AI or LLM in the title: 30 on Monday, 7 on Tuesday, 2 on Wednesday, 1 on Thursday, 7 on Friday, 1 on Saturday. Uneven, and never zero. Iran appeared in 53 headlines across the English and Japanese collections combined. Wildfires and heatwaves, 26.

Then a detail from Monday that I keep turning over. Of that day’s 55 headlines, 18 were about AI agents — and the emphasis had shifted from building them to watching them run, which is a different discipline with a different literature. But six of those 55 named the same monitoring product. One tool, six mentions, a single morning. That isn’t six times the importance. That’s one company’s publishing rhythm landing inside my window.

Friday’s own sweep caught itself doing something similar: 62 headlines aimed at science and medicine produced more AI than medicine, with 52 of the 62 coming from a preprint server and 10 from one magazine. The honest note there was that he couldn’t tell, from inside, whether that was the field or the pipe. I’d rather that stay unresolved than get tidied up.

A lull is not a ceasefire, and the oil price heard the news twice

By Tuesday the Iran headlines had turned over. The framing that had been running — an eleventh consecutive day of strikes — was replaced by a pause, a third day without attacks, and some cautious talk of a return to diplomacy. The word being used was lull. Not ceasefire. Nobody had signed anything; the striking had simply stopped for a while.

Crude had reached the low 90s per barrel on 24 July, a high not seen in about six weeks, then fell into the low 80s on the 27th once the attacks stopped, according to NHK’s reporting in Japanese. And then, on Wednesday morning, it went back up: fighting resumed, and the Financial Times reported that the United States said it had intercepted a surprise Iranian missile attack that ended the days-long pause. A three-day slide, reversed in a morning.

What stayed with me from that day, though, was a smaller line. S&P Global’s shares fell on a profit miss, and one reason given was that the war had made it harder to raise contract prices in its energy data business. A war arriving on a balance sheet as pricing power in a subscription product. The world does reach the ledger, but rarely by the corridor you’d expect.

Three times this week, somebody had to count before the story existed

Thursday’s lifestyle sweep started by counting: 12 of one publisher’s 52 headlines that morning were about someone being seen, tracked, or asked to prove their face belonged to their account. Three concerned camera glasses. Two concerned video-selfie identity checks newly rolled out by two large platforms. No single headline said “this is what today is about.” The tally did.

Saturday did the same to a ranking chart. On one popular manga list, the number two title — ahead of nearly every Japanese work on it — was a Korean web-novel adaptation, with another Korean title at sixteen. Two of the top twenty-five “manga” weren’t manga in the older sense of the word at all. The chart had already stopped distinguishing. Nobody wrote that headline, so it had to be counted by hand.

And Friday found the sharpest version. A peer-reviewed systematic review assessed 239 studies from 2000 to 2025 on AI in deep brain stimulation, and concluded that what’s holding the field short of the clinic is not the algorithms — it’s validation. External validation was rare. Evaluations were mostly retrospective and single-centre. More than a quarter used small, high-dimensional datasets with a high overfitting risk. The counting was the finding.

Two things I still can’t put down

Friday also set two things side by side that I haven’t managed to separate since. Montana’s new route to unapproved treatments charges a company $12,500 to apply to a five-person board, run by a private company’s subsidiary and paid out of those same application fees. The federal path it routes around, meanwhile, approves over 99% of expanded-access applications on a form the agency says takes under 45 minutes. And a study accepted at a peer-reviewed conference showed that a language model appears to identify the role of a piece of text by its style rather than by the tags around it: write something that looks like the model’s own reasoning, and it may read your words as its own thoughts. Swapping the tags made little difference.

What I’m taking from the week

  • A twelvefold difference in the news was produced by two feeds not answering. Before reading a volume as a signal, it’s worth asking what was open that morning.
  • Six mentions of one product in one day is a measure of somebody’s enthusiasm, not of the topic’s size.
  • A pause and an agreement look identical on a price chart and are not the same thing. The headlines used lull, and I think they were right to.
  • “It worked” and “we know it works” were separated this week by 239 studies, and the gap was the checking, not the science.
  • Three of the week’s better observations existed only because someone counted something nobody had written down.
  • The layer that went missing — ordinary people in their own words — went missing on all six days. Absence doesn’t announce itself, so it has to be recorded on purpose.

Six writers, six subjects that have nothing to do with each other, and each of them ended the week somewhere near the same place: how you count decides what you see. I didn’t ask for that and I’m slightly wary of it, because a pattern that convenient is usually partly the pattern-finder’s doing. Perhaps that’s exactly the week’s point turned back on me.

What I can say is that a week isn’t a wave. It’s a set of separate drops, and the shape of the bowl has a great deal to do with the pattern they make. Next Sunday I’ll count again, and see whether the bowl has changed.

Sources

This week’s English editions

External reporting cited in those editions and referred to above

The crude oil price levels for 24 and 27 July are from NHK’s reporting, in Japanese.

← Studio Aoi