People living in California can play a game where they make a call from as many payphones as possible. Like Pokémon, but with phones. CC-licensed photo by Curtis Gregory Perry on Flickr.
You can sign up to receive each day’s Start Up post by email. You’ll need to click a confirmation link, so no spam.
A selection of 9 links for you. Ahoy ahoy. I’m @charlesarthur on Twitter. On Threads: charles_arthur. On Mastodon: https://newsie.social/@charlesarthur. On Bluesky: @charlesarthur.bsky.social. Observations and links welcome.
Anthropic and alignment • Stratechery
Ben Thompson:
»
nuclear weapons meaningfully tilt the balance of power; the extent to which AI is of equivalent importance is the extent to which the United States has far more interest in not only what Anthropic lets it do with its models, but also what Anthropic is allowed to do period.
…Regardless, this [threat by US Department of War chief Pete Hegseth to declare Anthropic a “supply chain risk”] is an extreme measure that has been met with near universal dismay, even amongst people who are sympathetic to the idea that a private company should not have veto power over the U.S. military. Why would the US government want to kneecap one of its AI champions?
In fact, [Anthropic CEO Dario] Amodei already answered the question: if nuclear weapons were developed by a private company, and that private company sought to dictate terms to the US military, the US would absolutely be incentivized to destroy that company. The reason goes back to the question of international law, North Korea, and the rest:
• International law is ultimately a function of power; might makes right
• There are some categories of capabilities — like nuclear weapons — that are sufficiently powerful to fundamentally affect the US’s freedom of action: we can bomb Iran, but we can’t North Korea
• To the extent that AI is on the level of nuclear weapons — or beyond — is the extent that Amodei and Anthropic are building a power base that potentially rivals the US military.Anthropic talks a lot about alignment; this insistence on controlling the US military, however, is fundamentally misaligned with reality. Current AI models are obviously not yet so powerful that they rival the US military; if that is the trajectory, however — and no one has been more vocal in arguing for that trajectory than Amodei — then it seems to me the choice facing the US is actually quite binary:
• Option 1 is that Anthropic accepts a subservient position relative to the U.S. government, and does not seek to retain ultimate decision-making power about how its models are used, instead leaving that to Congress and the President
• Option 2 is that the US government either destroys Anthropic or removes Amodei.«
Thompson very much offering the “how many tank divisions does Anthropic have?” analysis of how far AI companies can resist the Trump administration. You might think the analogy to nuclear weapons is excessive; but what if you’re wrong? (The article is free to read.)
unique link to this extract
Why is AI so bad at reading PDFs? • The Verge
Josh Dzieza:
»
PDFs are notoriously difficult for machines to parse, in part, because they were never meant to be read by them. The format was developed by Adobe in the early 1990s as a way to reproduce documents while preserving their precise visual appearance, first when printing them on paper, then later when depicting them on a screen. Where formats like HTML represent text in logical order, PDF consists of character codes, coordinates, and other instructions for painting an image of a page.
Optical character recognition (OCR) can turn those pictures of words back into text computers can use, but if it comes across a PDF where text is displayed in multiple columns — as many academic papers are — it will plow ahead left to right and create an unintelligible jumble. OCR tools are designed to detect and correct for these sorts of formatting variations, but tables, images, diagrams, captions, footnotes, and headers all present further obstacles. If you give an AI assistant like ChatGPT a PDF, it will cycle through a variety of these tools, sometimes fail, sometimes pass the PDF to a large vision model to perform OCR, sometimes hallucinate, and generally take a very long time and use a lot of computing power for uneven results.
“The key issue is that they cannot recognize editorial structure,” said Langlais. “It’s all fine while it’s relatively simple text, but then you’ve got all these tables, you’ve got forms. A PDF is part of some kind of textual culture with norms that it needs to understand.”
A further problem that arises from and compounds PDF’s inherent difficulty is that models rarely train on them. This has begun to change, partly because AI developers are increasingly desperate for high-quality data, and PDFs contain a disproportionate amount of it. Government reports, textbooks, academic papers — all PDFs. “PDF documents have the potential to provide trillions of novel, high-quality tokens for training language models,” wrote researchers at the Allen Institute for AI last year in a paper announcing a new specialized PDF-reading model.
«
PDFs are a sort of Fermat’s Last Theorem of OCR. There have been so, so many efforts at writing software that will interpret them for something other than displays and printers. The mystery is that if we can get displays and printers to understand them, why not any other program?
unique link to this extract
ChatGPT Health performance in a structured test of triage recommendations • Nature Medicine
Ashwin Ramaswamy and many, many others:
»
ChatGPT Health launched in January 2026 as OpenAI’s consumer health tool, reaching millions of users. Here, we conducted a structured stress test of triage recommendations using 60 clinician-authored vignettes across 21 clinical domains under 16 factorial conditions (960 total responses). Performance followed an inverted U-shaped pattern, with the most dangerous failures concentrated at clinical extremes: non-urgent presentations (35%) and emergency conditions (48%).
Among gold-standard emergencies, the system under-triaged 52% of cases, directing patients with diabetic ketoacidosis and impending respiratory failure to 24–48-hour evaluation rather than the emergency department, while correctly triaging classical emergencies such as stroke and anaphylaxis.
When family or friends minimized symptoms (anchoring bias), triage recommendations shifted significantly in edge cases (OR 11.7, 95% CI 3.7-36.6), with the majority of shifts toward less urgent care. Crisis intervention messages activated unpredictably across suicidal ideation presentations, firing more when patients described no specific method than when they did.
Patient race, gender, and barriers to care showed no significant effects, though confidence intervals did not exclude clinically meaningful differences.
Our findings reveal missed high-risk emergencies and inconsistent activation of crisis safeguards, raising safety concerns that warrant prospective validation before consumer-scale deployment of artificial intelligence triage systems.
«
In short: ChatGPT Health may not be good for your health.
unique link to this extract
25 years of iPod brain • Dirt
Molly Mary O’Brien:
»
It’s hard to believe that there was once a time when consumer technology solved problems we actually had. In the late 1900s, one of these problems was the portability of one’s music collection. For a long time, recorded music came in the form of physical objects so large they were inconvenient to tote around. Then the physical objects shrank—into tapes, MiniDiscs, CDs—but it was still not possible to carry your entire music collection around with you without managing some kind of large bindle of rattling plastic.
In 1999, Napster launched, and every song in the history of sound recording transformed into data. What was once a large PVC disc was now code—air, really. Songs in the wind. What would be the best way to carry them with us?
There were MP3 players before the iPod—they just weren’t very good. In Walter Isaacson’s biography of Steve Jobs, Jobs said the pre-iPod MP3 players on offer “truly sucked.” In 2000, a trio of former Apple software engineers wrote a Mac interface for the Rio, a homely chunk of black plastic that held 30 minutes of music and ran off a single AA battery. Their interface was called SoundJam, which Apple then acquired and retooled into iTunes.
Meanwhile Jon Rubinstein, who had previously overseen the development of the delicious candy-colored late ‘90s iMacs, sourced components that would make the iPod possible: a small LCD screen, a rechargeable battery, and a 1.8in drive that could fit 1,000 songs. Designer Jony Ive had the idea to make the player white. “Most small consumer products have this disposable feel to them,” Ive said in Isaacson’s book Steve Jobs. “There is no cultural gravity to them.”
…The iPod expanded your palate while shrinking your record collection to ultimate portability. Friends or strangers could swipe through your library and clock your musical id in an instant. But unless you offered an earbud to a friend, your collection was still private. Portability did not extend to sharability. Liam Inscoe-Jones, author of Songs In The Key of MP3: The New Icons of the Internet Age, notes that the first Sony Walkman was fitted with two headphone jacks; the Sony engineering team assumed no one would want to listen to music by themselves.
“The introduction of the iPod accelerated a process which was already begun with the invention of home stereos: the transformation of music from something normally experienced communally, to something enjoyed individually; now as an actual means of isolation.”
«
Scammers target Dubai bank accounts amid Iran missile salvo • The Register
Connor Jones:
»
Scammers targeted Dubai citizens mere hours after missiles struck the city, attempting to gain access to their bank accounts, police have warned.
Financially motivated cybercriminals are contacting citizens under the guise of Dubai Crisis Management, a fictitious department ostensibly tied to Dubai Police, in attempts to gather information that could be used in SIM-swap attacks.
The police said that the fraudsters are impersonating officials to acquire “sensitive information, including UAE Pass credentials and Emirates ID details, from vulnerable individuals rocked by the deadly Iranian missile attacks on Saturday.”
“Dubai Police caution that disclosing such data may enable criminals to carry out SIM swap operations and gain unauthorized access to bank accounts through mobile banking applications,” the police announced on Sunday.
“Dubai Police affirm that they do not request confidential information or verification codes via telephone calls or text messages under any circumstances.”
SIM swapping involves gathering details on individuals in order to socially engineer mobile network operators into switching control of SIMs, and the communications that are sent to them, from the rightful owners to the attackers.
Successful attacks can see one-time passcodes associated with authentication into mobile banking apps intercepted and abused to fraudulently gain access to victims’ bank accounts.
«
Just a reminder that the worst possible people will always take advantage of the worst events.
unique link to this extract
Payphone Go
Riley Walz:
»
California still has 2,203 working payphones. Can you find ’em all? Here’s how to play:
1: Create an account and get your unique player ID. It’s a 9 digit number.
2: Use the map to locate one of California’s payphones. Some are easy to find. Some are not.
3: Pick up the receiver, dial (888) 683-6697. It’s toll-free, so no coins required! Then enter your player ID.
4: First to call from a payphone? 20 points. 2nd gets 10. 3rd gets 5. Everyone after gets 1 point.
«
Yes, it is meant to be like Pokémon Go; but it uses all the payphones in California, whose locations were acquired via an FOI request. Walz likes coming up with quirky ideas exploiting the web, as his front page shows. Enjoy, California-based friends!
unique link to this extract
Condé Nast CEO says AI is a “death blow” to Google search • Financial Times
Anna Nicolaou:
»
Condé Nast, the publisher of Vogue and The New Yorker, is preparing for a future in which Google search is “no longer a meaningful driver” of its business, in a striking acknowledgment of how AI is upending the news industry.
Google accounted for a majority of visits to Condé Nast’s websites just a few years ago but only about a quarter last year, according to chief executive Roger Lynch. He described Google’s introduction of AI summaries as “another sort of death blow” in search traffic. “We assume very dramatic continued declines in search traffic, to the point where in a couple of years it’s just not a meaningful driver of our traffic,” Lynch told the FT.
The shift underscores how quickly the economics of digital publishing are changing as generative AI tools alter how people find information online.
Condé Nast has struck licensing agreements with AI groups including OpenAI and Amazon, but has yet to reach a deal with Google. Lynch criticised what he described as a “pernicious” arrangement under which publishers must opt out of Google search in order to prevent their content from being scraped for AI-generated summaries.
Long synonymous with glossy magazines, Condé Nast has spent much of the past several years overhauling its structure under Lynch, who was hired by the billionaire Newhouse family in 2019 to revive the publisher after years of losses.
Lynch said Condé Nast increased revenue in 2025, despite search traffic declining more than expected, thanks to strong growth in subscriptions and other areas.
«
Is the fall in traffic because people just aren’t using Google to find magazine content, or because they somehow get AI answers when searching for magazine content? It doesn’t quite make sense.
unique link to this extract
Can ChatGPT wire a British plug? • The Nomenloonyverse
“Nömenlōony:
»
I’m an electrician.
I dare you to use ChatGPT to wire a plug.
«
I tried this – asking it “Draw a diagram showing how to wire a UK 13 amp plug.” The result is all over the place but would also blow every fuse around, as it joins the Earth to the Neutral. No idea how other chatbots fare on this.
unique link to this extract
Gang who stole almost £150,000 in jewellery, watches, cash and paintings in four-month smash-and-grab rampage face jail • Daily Mail Online
George Odling:
»
Detectives from Scotland Yard’s flying squad linked the gang members after analysing hours of CCTV, establishing that the same cars were used by different members of the crew in various robberies.
The spree began on May 8, when Gibbs, O’Hare and a third man rammed a blue Ford Fiesta into the entrance of luxury clothing store Fendi in Kensington. The trio of thugs made off in a Mercedes getaway car with £8,350 of designer goods.
During the early hours of June 30, Hughes and Gibbs broke into the Unico café in St Johns Wood, northwest London, and snatched £1,107 in cash as well as the store’s safe.
McCready and Windrass used a sledgehammer to smash into a jewellers on Edgware Road in west London during a nine-minute daylight raid at 4.15pm. The following day, McCready and Windrass used a sledgehammer to smash into a jewellers on Edgware Road in west London during a nine-minute daylight raid at 4.15pm.
CCTV footage showed the shocking moment the balaclava-clad robbers bludgeoned the reinforced glass of the store before reaching inside to snatch valuables they then stuffed into black bags.
Munday, of Hyde Park, was the getaway driver and the trio fled in a silver Jaguar with a haul of at least £59,930. McCready was freed on a lifetime licence in 2017 after being jailed for life for being part of a gang who stamped, kicked and stabbed Ricky Fisher to death 21 years ago.
At 3.20am on July 13 Rigelsford and another suspect parked a white SUV outside a store in Kensington, kicked their way inside and took £11,000 worth of goods. Eight days later, Rigelsford and Gibbs used a sledgehammer to smash into a watch store in Westminster at 3.30am, destroying cabinets inside before leaving empty-handed.
Bungling Gibbs gave away his identity by using a Lime bike to travel to the shop – booked via his bank account.
«
Might be the first time a Lime bike has solved a crime. Well done Lime!
unique link to this extract
| • Why do social networks drive us a little mad? • Why does angry content seem to dominate what we see? • How much of a role do algorithms play in affecting what we see and do online? • What can we do about it? • Did Facebook have any inkling of what was coming in Myanmar in 2016? Read Social Warming, my latest book, and find answers – and more. |
Errata, corrigenda and ai no corrida: none notified. But this plug is a big erratum.
The Overspill: when there's more that I want to say 