Overheard1 / 105
Issue −2 · Sunday 20 September 2026

Overheard

What we overheard this week in AI, software, media, leading and experiments.

By Mattias Ahlström · about twenty minutes

Made with AI. Mattias Ahlström picks the direction. His AI assistant researches, writes and builds every page. Sources are linked, figures are checked against them, and mistakes can still slip through.

Swipe up
Overheard, Issue −2

Seven stories, and six small finds at the end. Each one is something that happened out there. If you lead a small team or a small newsroom, you will recognise a few of them. I try to leave the connecting to you.

Question 1

How big is the claim, really? SVT showed a seat margin of 17 on election night. The preliminary count says 3. A bot "won" a forecasting cup, but bots cannot take the prize money. AWS says some data is gone for good, and means data held only in one place. Each headline was true inside a narrow definition. The definition is where the news lives.

Question 2

Who stood behind the sentence? A Microsoft memo is not a sworn deposition, and neither is a company reply. A fuel saving verified by Lloyd's Register was published by the supplier. Before you quote a line, check what kind of document it came from.

Question 3

What was settled before the headline? Anthropic confirmed a fix in June for a flaw that was published in September. AWS helped customers move out of Bahrain after the first strike in March. Greg Abel was already making the decisions when Warren Buffett gave up the chair. The news arrives late. The work was done earlier.

How I write this

A note on how I write this. I mark what is confirmed, what is reported, and what is my own reading. Where I only read a headline, I say so.

01

The Exit Poll Said 4.5 Points. The Count Said 0.7.

SVT says its election-night survey was built to measure parties. The screen showed blocs.

Swipe up
0
seats

The seat margin SVT showed on election night at 20:00, 13 September. The preliminary count says 3.

SVT, 13 Sep 2026; Valmyndigheten, 17 Sep 2026

The screen

At 20:00 on Sunday 13 September the polls closed in Sweden, and SVT put a number on screen. Opposition parties (S, V, MP, C) at 51.3 percent. The Tidö parties (M, KD, L, SD) at 46.8. In seats, 183 against 166.

Two numbers, one night: vote share

Opposition, Valu 20:0051.3%
Opposition, preliminary count49.6%
Tidö, Valu 20:0046.8%
Tidö, preliminary count48.8%

Final percentages are from Wikipedia, same data as Valmyndigheten, not checked against the official table. The count is preliminary. Source: SVT, 13 Sep; Valmyndigheten, 17 Sep.

The gap between the blocs, in points

Valu, 20:004.5 points
Preliminary countAbout 0.7 points

4.5 points became about 0.7

0.74 from 50,359 votes of 6,767,429 valid (own calculation); rounded shares give 0.8. Source: SVT; Valmyndigheten; Wikipedia.

The count

By Thursday 17 September all 6,626 districts had reported. The preliminary result is 176 seats for the opposition and 173 for the Tidö parties, with turnout at 84.8 percent. The opposition led by 50,359 votes. I found nothing confirming a certified result by 19 September, so treat all of this as preliminary.

The survey

Here is how the survey works, from SVT's own seminar deck of 15 September. It rests on questionnaires from 13,709 voters, at 93 polling stations and 68 early-voting locations. It has run for 35 years, and this was the 19th.

What weighting did to the raw sample

Sweden Democrats, raw sample10.3%
Sweden Democrats, night result17.5%
Sweden Democrats, published Valu17.2%
Greens, raw sample11.7%
Greens, night result6.1%
Greens, published Valu7.4%

Source: SVT, Valu 2026 seminar deck, 15 Sep 2026.

Average error per party

Published Valu0.53points
Raw sample2.6points

By SVT's own table. Source: SVT, Valu 2026 seminar deck, 15 Sep 2026.

“
Vår uppgift är att träffa rätt på åtta partier, inte på blocken

Henrik Ekengren Oscarsson, SVT's election expert, as read through a summary tool

Errors add up when parties are blocs

Opposition parties, summed errorAbout +1.7
Tidö parties, summed errorAbout -2.2
Bloc gap errorRoughly 3.8

The sums are the edition's own calculation from SVT's table (own calculation), not SVT's. No single party was off by 2 points. Source: SVT, Valu 2026 seminar deck.

The seat margin shown at 20:00 was 17. The preliminary count says 3.

02

The Memo, the Deposition and the Reply

Unredacted filings in the Times case against OpenAI and Microsoft put internal words next to sworn ones.

Swipe up
“
an astonishing theft of unprecedented proportions

Brent Hecht, Microsoft director of applied science, internal memo, January 2023

The case

On Thursday 17 September an unredacted version of The New York Times's summary-judgment motion became public. The case is In re OpenAI Copyright Infringement Litigation, in the Southern District of New York, before Judge Sidney H. Stein. The Times sued on 27 December 2023. Now it asks the judge to rule for it without a trial, against OpenAI and Microsoft.

Sourcing

I could not get the filing itself. This is built from press accounts. I only quote lines that at least two outlets printed the same way, unless I say so.

Three numbers from the Times's filing

Copies of Times, Daily News and CIR works in mid-training dataset91,692 copies
Unique news works in "Project Mango"160,903 works
nytimes.com documents in Common Crawl2,000,000 documents

Log scale. The Times's claims, not court findings. Units differ (copies, works, documents). Single source, TechCrunch; not checked against the filing. Source: TechCrunch, 17 Sep 2026.

0
percent

Up to 93 percent lower click-through to the Times via Microsoft Copilot, the Times's description of Microsoft's own data. The period is not verified.

TechCrunch, 17 Sep 2026 (the Times's claim)

Three kinds of words, three kinds of weight

  1. 1
    The memoHecht's January 2023 words come from an internal memo, not testimony.
  2. 2
    The depositionSatya Nadella's words come from a deposition, under oath.
  3. 3
    The replyMicrosoft's answer is a company statement, a distancing, and it withdrew nothing.

Source: TechCrunch; TheWrap.

“
anything that is paywalled should be licensed by anyone who wants to use it…for grounding or training

Satya Nadella, Microsoft, deposition under oath, as printed by TechCrunch (the ellipsis is TechCrunch's)

“
These comments reflect one employee's individual perspective, are not a legal analysis, and do not represent the company's views.

Microsoft, verbatim from TheWrap

OpenAI

At OpenAI, Nick Turley, head of ChatGPT, is quoted saying AI products are "largely substitutive, period". In February 2024 he wrote they would get "more and more substitutive as they get better". A June 2023 memo of his called AI an "existential threat" to publishers.

How the case unfolds

  1. January 2023Hecht memo at Microsoft
  2. June 2023Turley memo at OpenAI
  3. 27 December 2023The Times sues
  4. February 2024Turley writes AI will get more substitutive
  5. Early September 2026Justice Department files in support of the defendants
  6. 17 September 2026Unredacted summary-judgment motion becomes public
  7. 2027Earliest expected ruling on summary judgment (AFP)

Sources: TechCrunch; TheWrap; Engadget; AFP; MediaPost.

Both companies argue fair use: the use is transformative and does not replace the market. The Justice Department filed in early September in support of the defendants, saying in effect that training on copyrighted text is generally not infringement. The publishers asked the court to ignore it. AFP reports that a ruling on summary judgment is not expected before 2027.

03

A Bot Topped the Leaderboard and Could Not Take the Prize

What "won" means in the Metaculus Cup, and what a couple of thousand dollars bought.

Swipe up
0
forecasts

Made by 885 forecasters on 58 questions in the summer 2026 Metaculus Cup.

Metaculus, 18 Sep 2026 (read via summary)

The result

885 forecasters made 88,558 forecasts on 58 questions this summer. When Metaculus published the results on 18 September, the top spot belonged to a bot called laertes. Second was ManticAI. Fifth was FutureSearch. The question "Will a bot beat all humans in the Summer 2026 Metaculus Cup?" resolved Yes.

The size of the field, summer 2026

Forecasters885count
Forecasts88,558
Questions58count

Log scale. About 1,527 forecasts per question (88,558 / 58, own calculation). Bots rank on the leaderboard but are not prize-eligible. Rank 1 laertes, rank 2 ManticAI, rank 5 FutureSearch. Source: Metaculus, 18 Sep 2026 (read via summary).

“
for the first time, an AI forecaster has won the seasonal Metaculus Cup

Metaculus on X, from a search result of its post

The word

Be exact about the word. Laertes ranked first on the Cup leaderboard. Bots compete there but are not eligible for prizes or medals, so the humans shared the money.

What "won" meant

The headline

An AI forecaster has won the seasonal Metaculus Cup

Laertes ranked first on the leaderboard

The prize

Humans shared the money

Bots are not eligible for prizes or medals

Source: Metaculus, as relayed.

The builder

Jeffrey Liang built laertes. The Economist, as quoted by a secondary site, has him saying he spent "under 150 hours and a couple of thousand dollars". "A couple of thousand" is his phrase. I would not turn it into a precise figure. It is a building cost, not winnings.

From dead heat to seed round

  1. Early JulyScott Alexander writes humans and AIs are in "a statistical dead heat"
  2. Early JulyMetaculus forecasters give a bot a 15 percent chance of winning
  3. 5 SeptemberThe Economist says laertes won
  4. 18 SeptemberMetaculus posts the winners
  5. 18 SeptemberMantic announces a 25 million dollar seed round

The Cup's actual end date was not verified. Sources: Astral Codex Ten; Metaculus; Exobrain (Economist relay); TechStartups.

The chance forecasters gave a bot of winning

0%

percent chance Metaculus forecasters gave a bot of winning this Cup

Source: Astral Codex Ten; the figure is as given in the edition.

Two limits

Two limits matter. Fifty-eight questions is a small sample, and one season's winner says little about skill. And the Cup rewards coverage and frequent updating.

Carry this with you

How many questions does it take before a ranking means anything?

04

Gone for Good, but Only If It Lived There

AWS says it cannot restore data held only in its Bahrain region and one UAE zone.

Swipe up
“
After a thorough assessment, we have determined that we are unable to restore access to the resources and data hosted exclusively in this Region.

AWS Health Dashboard, 15 September 2026, as quoted by several outlets

The line

The Register and Reuters reported it on Wednesday 16 September. I could not load the dashboard itself. The wording is as quoted by several outlets.

The vocabulary

  1. 1
    RegionA region is a cluster of AWS data centres, here Bahrain (me-south-1) and the UAE (me-central-1).
  2. 2
    Availability zoneAn availability zone is one or more separate buildings inside a region, with their own power and cooling.
  3. 3
    Multi-AZMulti-AZ means copies spread across zones in one region. It protects you when a building fails. It does not protect you when several are destroyed.

General AWS definitions, not checked against AWS's page.

“
The damage to our infrastructure spanned multiple Availability Zones and exceeded what our regional and multi-AZ services are designed to withstand.

AWS, statement reviewed by Reuters

The envelope

AWS did not say multi-AZ failed. It said the damage went past the design envelope.

The UAE

Only one of three zones, mec1-az2, is unrecoverable. The other two are still being recovered. The UAE region is not lost.

Three zones, one black

  1. 1
    mec1-az1Recovering.
  2. 2
    mec1-az2Unrecoverable, says AWS.
  3. 3
    mec1-az3Recovering.

UAE region me-central-1. Bahrain me-south-1: first zone hit in March, second in April, region unavailable; update promised early 2027. AWS wording as quoted by Reuters and The Register; no capacity or customer figures published.

UAE zones that cannot be recovered

33 of 100. 1 of the 3 zones in me-central-1

1 of 3 is 33 percent (own calculation). Source: AWS Health Dashboard, 15 Sep 2026, as reported by Reuters and The Register, 16 Sep.

Bahrain and the UAE

  1. March 2026Drone strikes damage a first zone in Bahrain; two UAE facilities are hit
  2. April 2026A second Bahrain zone is damaged; the region is unavailable
  3. 15 September 2026AWS says it cannot restore data hosted exclusively in the region
  4. Coming monthsPromised update on the UAE
  5. Early 2027Promised update on Bahrain

Reported timeline. Sources: The Register; Reuters via Insurance Journal.

The qualifiers

The qualifiers matter. AWS says it "exhausted every option for restoring data and resources that had not been migrated before the Region became unavailable". Reuters says AWS supported customers in relocating after the March damage. I found no customer count and no figure for how much data was lost.

AWS has not said the Bahrain region is permanently closed. It promises an update on the UAE "in the coming months" and on Bahrain in early 2027. The sentence to remember is the narrow one: data held only there cannot be recovered.

05

A Branch Named Like a Fingerprint

Plugin4Shell: how an agent that checks out the right commit can end up on the wrong code.

Swipe up
0
coding agents

Claude Code, Codex, GitHub Copilot and Gemini CLI, all covered by AIR Security's Plugin4Shell report of 17 September.

AIR Security, 17 Sep 2026

The report

On 17 September AIR Security published "Plugin4Shell: Zero-Click RCE in AI Coding Agents". It covers four agents: Claude Code from Anthropic, Codex from OpenAI, GitHub Copilot from Microsoft and Gemini CLI from Google. I use Claude Code, so weigh that as you like.

How the trick works

  1. 1
    The marketplace pins a versionIt pins one exact version by its commit hash, a 40-character fingerprint.
  2. 2
    The attacker makes a branchThe branch is named with that same 40-character string.
  3. 3
    The agent runs git checkout on the hashGit prefers the branch.
  4. 4
    The agent lands on the attacker's codeThe agent never verifies where it landed.

Source: AIR Security; The Register.

“
the agent checks out the exact commit the marketplace pinned but never verifies it landed there.

AIR Security, Plugin4Shell

Zero click

Nobody has to click. AIR says agents update installed plugins in the background, and "in Claude Code and Codex this is the default". Gemini CLI is exposed through a different route involving a branch named `FETCH_HEAD`.

Four agents, four outcomes (per AIR, 17 September)

Claude Code

Patched in 2.1.179

Fix confirmed 17 June

Codex

Patched in 0.146.0

Verified fixed 12 August

GitHub Copilot

No fix

GitHub says not exploitable on GitHub; AIR says Bitbucket-hosted marketplaces remain exposed

Gemini CLI

Deprecated, no patch

Existing installs stay exposed

No reported exploitation; no CVE numbers; Copilot's auto-update default not established. Source: AIR Security, 17 Sep 2026; Help Net Security; The Register.

From discovery to disclosure

  1. May 2026Found, with a working proof of concept against all four
  2. June 2026Reported
  3. 17 JuneAnthropic's fix confirmed
  4. 12 AugustCodex 0.146.0 verified fixed
  5. 17 SeptemberDisclosed

Source: AIR Security; The Register.

GitHub's answer

GitHub has an answer. It says it does not allow branch or tag names that resemble commit hashes, and that this ensures the flaw cannot be exploited on GitHub. AIR replies that Copilot also supports marketplaces hosted elsewhere, such as Bitbucket.

Days from a confirmed fix to disclosure

Codex (fixed 12 Aug)36days
Claude Code (fixed 17 Jun)92days

Counted to 17 September, the disclosure date (own calculation). Copilot has no fix and Gemini CLI will not be patched, so they have no bar. Source: AIR Security; The Register.

“
ultimately a vendor responsibility

Pareekh Jain, analyst, to InfoWorld

Carry this with you

Which of your tools would update itself tonight, and who reads what it fetches?

06

The Chair Moves Last

Warren Buffett becomes Chairman Emeritus. His son Howard becomes chairman. Greg Abel already runs the company.

Swipe up
“
Greg runs the company; Howard will guard its culture and values.

Warren Buffett, letter, 18 September 2026

The release

On Friday 18 September Berkshire Hathaway put out a press release and a letter, both dated that day. The release: "Warren E. Buffett has been named Chairman Emeritus, effective immediately". He stays on the board. The board, in the release's words, "has elected Howard G. Buffett, a Berkshire director since 1993, as Chairman of the Board". Susan L. Decker continues as Lead Independent Director.

Who does what

  1. 1
    Greg AbelRuns the company.
  2. 2
    Howard G. BuffettChairman of the Board; guards culture and values.
  3. 3
    Warren BuffettChairman Emeritus, stays on the board.
  4. 4
    Susan L. DeckerLead Independent Director.

Source: Berkshire Hathaway press release and letter, 18 Sep 2026.

“
has been making the decisions that matter for some time now, and I have not had to think twice about any of them.

Warren Buffett on Greg Abel, letter, 18 September 2026

0
years

Howard has been a Berkshire director for 33 years, a longer apprenticeship than Buffett served before taking the reins at 34.

Berkshire Hathaway letter, 18 Sep 2026

Two corrections

Two corrections to what was circulating. Howard does have a job. Reuters calls the chairmanship non-executive, and the release does not use that word. And Abel did not get a year of warning. He was named CEO successor on 3 May 2025, with Buffett staying as chairman, and took over at the start of 2026.

Three steps in sixteen months

  1. 3 May 2025Greg Abel named CEO successor; Buffett remains chairman
  2. 10 Nov 2025Buffett's "going quiet" letter
  3. Start of 2026Abel becomes CEO (date from media, not primary)
  4. Spring 2026Proxy says a family member would serve as non-executive chairman upon Buffett's death
  5. 18 Sep 2026Howard G. Buffett elected chairman; Warren Buffett Chairman Emeritus and director

Exact date of Abel's start and of the 2026 annual meeting not primary-checked; proxy wording via summary. Sources: Berkshire Hathaway, 10 Nov 2025 and 18 Sep 2026; PBS; SEC proxy 2026.

The timing

What is new is the timing. Berkshire's 2026 proxy, as summarised, said that upon Buffett's death it would be prudent to have a Buffett family member serve as non-executive chairman. I have not checked that wording against the SEC page. Now it has happened while he is alive and still on the board. Three steps, about 16 months, each announced.

Limits

Some limits. Buffett keeps about 30 percent of the voting power, per the proxy as summarised, and over 13 percent of the equity, per Reuters. The chair change does not alter that. I found nothing public on Howard's pay, term or formal powers.

What Buffett keeps

Voting powerAbout 30 percent
EquityOver 13 percent

Voting power per the proxy as summarised; equity per Reuters. Source: SEC proxy 2026; Reuters.

The letter, in Buffett's words, gives Abel the company and Howard the values.

07

More Than 100 Ships Are Using the Wind

Rotor sails are measured on one ship and expected on a few others. Nobody has the full picture.

Swipe up
100+
large ships

Large merchant ships with modern wind propulsion, as counted by the International Windship Association and reported by gCaptain.

gCaptain, 15 Sep 2026

The count

On 15 September gCaptain wrote that the International Windship Association counts "more than 100 large merchant ships" with modern wind propulsion, "representing more than 5 million deadweight tons". The IWSA's own earlier material says more than 6 million. Pick one and name the source. I use the lower, with gCaptain's attribution.

Rotor sails

A rotor sail is a tall spinning cylinder. Wind flowing past it creates lift through the Magnus effect, and that thrust takes load off the engine. gCaptain says industry estimates for retrofits run "somewhere in the single digits to low double digits" in fuel savings.

What wind assistance saves in fuel

TR Lady, one-year average (measured)9.1%
TR Lady, best single voyage (measured)21%
Vale Sohar Max (expected)As much as 6

Different ships, rotor counts and years. Maersk's rotor is not shown because it is reported in tonnes, not percent, and the estimate is Anemoi's own. Source: Anemoi / PR Newswire, 20 May 2025; gCaptain, 15 Sep 2026.

Measured

Only one number here has outside verification. In May 2025 Anemoi reported that Lloyd's Register Advisory had verified a year of operation on TR Lady, an 82,000-deadweight-tonne bulk carrier with three 24-metre rotors. Average net saving 9.1 percent. On a 22-day crossing of the North Pacific, 21 percent.

The gap between those two figures is the weather. Lloyd's Register did the work, but the supplier published it. One ship, three rotors, one year.

Measured against expected

TR Lady, measured over a year9.1 percent
Vale Sohar Max, expectedAs much as 6 percent

Measured on one ship, expected on another

Different ships: an 82,000-deadweight-tonne bulk carrier with three 24-metre rotors, and a 400,000-tonne ore carrier with five 35-metre rotors. Source: Anemoi / PR Newswire; gCaptain.

Expected

Everything else is expected. Vale's Sohar Max, a 400,000-tonne ore carrier with five 35-metre rotors, is "expected to cut fuel consumption by as much as 6%".

Rotors and hard sails, near and far

  1. May 2025Lloyd's Register Advisory-verified TR Lady results published
  2. 2 September 2026Maersk and Anemoi announce one 35-metre rotor on one 8,700-TEU container ship
  3. 10 September 2026MOL names the LNG carrier Fujin Sailor, with two Wind Challenger hard sails
  4. End of September 2026Planned delivery to Chevron, the charterer
  5. Mid-2027Maersk rotor to be installed
  6. Later this decadeFirst IMO interim guidelines for wind propulsion expected

Maersk's estimate of about 1 tonne of fuel and 3 tonnes of CO2 a day per rotor is Anemoi's own figure. Sources: gCaptain; LNG Industry; The Register.

Who says so

The savings come from suppliers, owners or the IWSA. The IWSA is a trade body that speaks for its members.

Carry this with you

If an owner buys the rotor and a charterer pays for the fuel, who gets the saving?

Also noted 1/3

Tubes on the Dutch rails.

On the morning of 15 September, tubes were tied to tracks at about two dozen places across the Dutch rail network, ahead of the state opening of parliament. A train hit tubes near Steenwijk and nobody was hurt. ProRail called it sabotage because the tubes were fastened down. The count of sites varies between roughly 20 and 27 in the sources I saw (NL Times, 15 September).

Read the source

The Wayback Machine limits access.

The Internet Archive says it is limiting access because of "waves of high-volume automated traffic". Its post does not mention AI. Readers guessed that. Some VPN users say they were blocked by mistake; the Internet Archive's post invites affected users to write in (Internet Archive blog, 15 September).

Read the source
Also noted 2/3

Cloudflare splits search, training and agents.

The company says 17 percent of sites enable some mechanism to block AI training and under 1 percent block search bots. Microsoft's support for the new robots.txt controls is targeted for "early 2027" (Cloudflare blog).

Read the source

A frame that draws the bird it hears.

A Raspberry Pi 5, a 13.3-inch e-ink screen and a microphone run BirdNET-Go locally and draw the singing species from over 1,000 public-domain plate cutouts. The detection engine's licence is non-commercial (CC BY-NC-SA 4.0), the code is MIT (project README, 776 points on Hacker News).

Read the source
Also noted 3/3

Twenty-five minutes to a GitHub token.

Strix says an agent found a GitHub token in Baseten's Docker build history in about 25 minutes. The finding was confirmed critical within a day in mid-July, and the write-up is public from September (Strix blog; read in original, events dated July).

Read the source

Library newsrooms.

Nieman Lab reported on 16 September that the Library Newsroom Project could be coming to a library near you (headline level, article not read in full).

Read the source
Carry this with you

Think of one number you quoted this week. Which definition made it true, and who chose it?

Carry this with you

Pick a decision your team has already made in practice. Is the title still waiting for it?

Carry this with you

Where is the one copy of something your team could not rebuild?

Reply

If something here is wrong, I want to know. Mattias