OpenAI's ChatGPT Wallet lets its AI agents pay for shopping, while Amazon and United Airlines keep agents out

The stories

32 on this beat today.

Every story told the same way: what happened, who is involved, why it matters, how it works, the truth and the good news.

1. 01

   OpenAI's ChatGPT Wallet lets its AI agents pay for shopping, while Amazon and United Airlines keep agents out

   1.2 billion

   people use ChatGPT each week

   nearly 100 million of them pay [2]

   42.8%

   of AI shopping attempts failed in Visa's test

   Visa tried 22 AI models as shoppers this month [4]

   1%

   of a US streetwear chain's web visits now come from AI agents

   up from nothing before, its head of growth says [6]

   What happened

   • OpenAI launched ChatGPT Wallet on Friday, which lets its AI helpers, called dots, buy things for people. [1]
   • OpenAI built it with American Express, Mastercard and Visa, the big card networks. [2]
   • At first it is only for ChatGPT Pro subscribers in the US aged 18 and older. [3]
   • Users add a card once, then check the items, the total and the card before each purchase. [4]
   • Every purchase needs a security check, such as Face ID or a fingerprint scan. [5]
   • Instead of the real card, the card network sends a payment token. [2] So the agent and the shop never see the card details. [4]
   • Some websites may block AI agents by mistake, and fraud checks can flag real agent purchases, OpenAI said. [4]
   • United Airlines said on Wednesday it will block AI agents from booking its tickets. [2]
   • Amazon stops outside agents buying on its site and steers shoppers to its own helper, Alexa. [1]
   • OpenAI published a playbook to help shops let its agents through their defences. [4]
   • OpenAI tried selling inside ChatGPT before, with Instant Checkout in autumn 2025. [2] It quickly moved away from it, and revamped shopping by March. [2]
   • Shopping has become a top activity on Meta's Muse agent, TechCrunch reports. [1]
   • DoorDash and Uber, by contrast, report few orders placed by AI agents so far. [1]
   • All Star Elite, a US streetwear chain, says about 1% of its website visits now come from AI agents. [6]
   • Adobe expects AI visits to US shops to double on Black Friday and Cyber Monday, compared with last year. [6]

   Who is involved

   • American Express card network, a partner
   • American Express, Mastercard and Visa the big card networks; they issue the stand-in card numbers Wallet pays with
   • Visa card network, a partner
   • Amazon the biggest online shop; it blocks outside AI agents and pushes its own, Alexa
   • Alexa Amazon's own helper
   • DoorDash few agent orders
   • Uber few agent orders
   • OpenAI the company behind ChatGPT; it launched Wallet so its AI agents can pay
   • United Airlines a big US airline; it said it will block AI agents from booking tickets

   How it unfolded

   1. Autumn 2025 OpenAI launches Instant Checkout, then moves away from it [2]
   2. March shopping revamped
   3. Friday ChatGPT Wallet
   4. Now approve each purchase
   5. Later limits set in advance
   6. March OpenAI revamps shopping in ChatGPT [2]
   7. Late September OpenAI introduces its dots agents [2]
   8. Wednesday United Airlines says it will block AI agents [2]
   9. Friday ChatGPT Wallet launches for US Pro users [2]

   Why it matters

   Shops now choose whether to let agents in, and United, Delta and Amazon have chosen to keep them out. [7]

   OpenAI says it will widen Wallet as it learns from early use with shops, banks and card networks. [8]

   Think it through

   1. 1

      Why would OpenAI take no cut of the sales?

      It is not taking a cut for now, while it builds ties with shops. [2] It wants ChatGPT to be a serious place to shop. [2]

      No cut for now→Ties with shops→A serious place to shop
   2. 2

      Who has more to bring right now?

      ChatGPT has 1.2 billion weekly users, against more than 10 million for Meta's Muse agent. [2,1]

      1.2 billion a weekChatGPT

      vsvs

      10 million+Meta's Muse
   3. 3

      So what is Amazon doing instead?

      An internal Amazon plan would index every product sold online, so agents might rely on Amazon's list, Reuters reported. [9]

   4. 4

      Can AI shoppers be trusted to get it right?

      Not yet: in Visa's test this month, 22 AI models acting as shoppers failed 42.8% of the time. [4]

      42.8%of AI shopping attempts failed in Visa's test

   5. 5

      Is anyone else handing agents people's money?

      Google let its Gemini app read people's Google Wallet spending on Thursday, but it cannot spend. [10]

   6. 6

      So what is really going on?

      OpenAI brings the shoppers, but shops keep control of prices, delivery and returns. [4] So Wallet works only where shops do not block agents. [3]

      OpenAIbrings shoppers

      vsvs

      Shopshold the checkout

   How it works

   Payments engineers call this tokenisation: a stand-in number replaces the real card number.

   Each token has its own card number, security code and expiry date, made for one purchase. [2]

   Seen it before

   A cloakroom ticket gets your coat back, but it cannot open your front door.

   The truth

   • What is really going on OpenAI and Google gave their AI helpers access to people's cards and spending records this week. [1,10]Shops still decide whether those agents get in, and Amazon and United Airlines are keeping them out. [2]
   • American Express says the small start is deliberate, to "start it in a contained environment". [2]
   • Who gains American Express, Mastercard and Visa — Every Wallet purchase still runs through their cards and the payment tokens they issue. [2]
   • Who pays Shoppers whose agent buys the wrong item — They fall back on disputing the charge with their card, a protection OpenAI says still applies. [1]
   • Nobody knows yet

     When OpenAI will start taking a cut of agent purchases.

     It takes none for now while it courts shops, CNBC reported, and it gave no date. [2]

   What comes next

   OpenAI plans to let people pre-approve purchases, with limits on spending, categories or time. [3]

1. 02

   Microsoft blames a network change for its outage

   What happened

   • Microsoft 365, the company's online email and office service, was down for more than 24 hours from Thursday. [12]
   • Outlook email, Teams calls and the Copilot AI helper failed in more than 70 countries. [12]
   • Microsoft said the fault came from an earlier network upgrade and surfaced after a later change. [12]
   • Its first fixes failed, and some servers got stuck in a bad state, its status updates said. [21]
   • In Japan, where 78% of companies rely on Microsoft, some offices went back to fax machines. [22]
   • Portugal's ministers lost their email and used WhatsApp and Signal to keep working. [23]

   Who is involved

   • Outlook email down
   • Teams calls down
   • Copilot AI helper down

   Why it matters

   Hospitals, airlines and police reported disruption, because so much daily work runs through Microsoft. [12]

   Think it through

   1. 1

      Was it a cyberattack?

      No, Microsoft said: it ruled out hackers and AI agents as the cause. [12]

   2. 2

      Why did it take so long to fix?

      Microsoft brought services back in stages and checked each was stable, which it says took longer. [24]

      Restore one part→Check it holds→Restore the next
   3. 3

      Who tries to profit from the confusion?

      Scammers do: NCC Group warned of fake IT staff asking people to reset passwords. [24]

   How it works

   Reliability engineers call this change management: every change to a working system is planned, tested and can be undone.

   Here a fault sat unseen after a network upgrade, until a later change set it off. [12]

   Seen it before

   A wobbly shelf can hold for weeks, until one more book makes it fall.

   The truth

   • Microsoft first blamed servers failing automatic health checks, then named the network upgrade. [21,12]
   • Nobody knows yet

     What the full cause of Microsoft's 24-hour outage was.

     Microsoft has promised a full post-incident report once its review is complete. [24]

   Good news

   • Microsoft's email and Teams came back after more than 24 hours. Microsoft stretched its email queue to 48 hours so messages sent during the outage still arrive.
2. 03

   Europe's leaders talk AI as the outage bites

   What happened

   • EU leaders held their first summit about AI in Dublin, Ireland's capital, on Friday. [25]
   • They were joined by Arthur Mensch, boss of the European AI firm Mistral, and Nobel winner Geoffrey Hinton. [25]
   • Mensch said only 20% of European digital services run on European infrastructure. [14]
   • Microsoft's outage knocked out email and Teams for many of the leaders gathered in Dublin. [14]
   • France's President Emmanuel Macron then wrote: "Our independence cannot be outsourced." [14]
   • The EU's plan for 7 AI gigafactories, huge data centres, is held up by a fight over its budget. [26]

   Who is involved

   • Arthur Mensch runs Mistral, Europe's top AI firm
   • Geoffrey Hinton Nobel winner, warns of AI risks

   How it unfolded

   1. Friday Dublin summit
   2. 17 November Apply AI Summit

   Why it matters

   A Finnish EU lawmaker, Aura Salla, said Europe has handed too much of its digital systems to outside firms. [14]

   Think it through

   1. 1

      Why can't Europe just switch?

      When everything hangs on one supplier, firms cannot switch quickly, Nextcloud's co-founder told The Register. [27]

      One supplier→Everything depends on it→No quick switch
   2. 2

      Who is pushing hardest for European tech?

      Mistral is: Mensch says money spent on US or Chinese tech could be reinvested in Europe. [14]

      Spent abroadUS or Chinese tech

      ↓→

      Reinvestedin Europe

   How it works

   Policy experts call this tech sovereignty: a country controls the computers and software its public services need.

   German lawmaker Johannes Schatzl said sovereignty means staying open for work through outages like this one. [14]

   Seen it before

   A family with one car is stuck when it breaks down, unless a bus runs nearby.

   What comes next

   The European Commission's next AI meeting, the Apply AI Summit, is in Brussels on 17 November. [26]

3. 04

   Workday outage locks staff out

   What happened

   • Workday, a service many employers use for pay and staff records, went down on the evening of 13 November. [13]
   • Employees could not clock in or out, see their pay details or ask for time off. [13]
   • Workday put some systems into maintenance mode to fix a problem it found during testing. [13]
   • Workday says 11,500 organisations use it, including Walmart, Netflix and NASA's Jet Propulsion Laboratory. [13]

   Why it matters

   Workers who clock in through Workday had no way to record their shifts. [13]

   Think it through

   1. 1

      When was it worst?

      Reports on Downdetector, an outage tracker, peaked around 5:19pm Eastern time, then tapered off. [13]

      5:19pmwhen outage reports peaked, Eastern time

   2. 2

      Was every customer hit?

      It is not clear whether the outage reached all of Workday's customers, Engadget reported. [13]

   How it works

   Engineers call this maintenance mode: a service shuts its doors to users while it repairs itself.

   The problem left some of Workday's data inconsistent, so letting people in could cause errors. [13]

   Seen it before

   A shop closes for a stocktake, because counting while shoppers move things gives the wrong total.

   The truth

   • Nobody knows yet

     Whether Workday's fault is fixed for good.

     Workday said it was still looking for the best way to fix the problem. [13]

4. 05

   Gemini can now read people's bank spending

   What happened

   • Google now lets every Gemini app user in the US connect Google Wallet, so the chatbot can see their spending. [10]
   • People can also link bank and card accounts through Plaid, a financial technology firm. [10,7]
   • Gemini can chart spending, find subscriptions that went up and suggest a holiday budget. [10]
   • It can only read the accounts: it cannot move, manage or invest money. [7]
   • OpenAI let some ChatGPT users link bank accounts in May, and Anthropic added a similar feature last week. [10]

   How it unfolded

   1. May ChatGPT links banks
   2. Last week Claude follows
   3. Thursday Gemini joins

   Why it matters

   Gemini users can now hand a chatbot their money records, and chatbots do make mistakes, Engadget notes. [7]

   Think it through

   1. 1

      Does Google learn from people's spending?

      Google says not from Wallet data directly. [10] But it may train on summaries of transactions if Gemini activity is left on. [10]

      Wallet datanot trained on

      vsvs

      Summariesmay be trained on
   2. 2

      Can people at Google read the chats?

      Some of those chats may be read by human reviewers, Google's help pages say. [10]

   3. 3

      What else does Google do with linked data?

      Google recently changed its privacy policy so connected Gmail and Photos data can shape ads. [10]

      Gmail and Photos linked→Data shapes ads

   How it works

   Privacy lawyers call this consent: an app may use personal data only in ways the person agreed to.

   Here the Wallet link is opt-in, read-only, and can be switched off at any time. [10]

   Seen it before

   Lending a friend your diary to check one date lets them read every page.

1. 06

   Shopify opens every store to AI agents

   What happened

   • Sierra, an AI customer-service firm, released version 1.0 of Poppy on Thursday. [11] Poppy is a standard for how AI agents deal with companies. [11]
   • Shopify, which hosts online shops, switched Poppy on for every store on its platform. [11]
   • Meta's Muse agent, set up by more than 10 million people, now looks for Poppy first on every site. [11]
   • Poppy tells a company who a visiting agent works for and what it may do, like checking an order. [11]
   • More than 100 companies have signed up, including Delta Air Lines, Intuit, Salesforce and Zoom. [11]

   Who is involved

   • Shopify turned Poppy on everywhere

   Why it matters

   Shopify stores need to do nothing, while Amazon wants shoppers on its own agent instead. [11]

   Think it through

   1. 1

      Who decides what an agent may do?

      The shopper does: they sign in and choose what their agent can reach, Sierra says. [11]

   2. 2

      Are the rules finished?

      No: the rules are still labelled a draft and could change in ways that break existing setups. [11]

      Version 1.0→Still a draft→Changes may break setups

   How it works

   Engineers call this a protocol: an agreed set of rules two machines follow to talk.

   Poppy handles the moment an agent arrives, and Shopify's own rules handle carts and checkout. [11]

   Seen it before

   Traffic lights work because every driver has agreed that red means stop.

   Why an AI shopper is no use until the shops let it in

   Economists call it a two-sided market: shoppers go where the shops are, shops go where the shoppers are, so a newcomer has to win both.

   The twist

   A newcomer pays to win the side it lacks. OpenAI already has 1.2 billion weekly users, so it is the shops it must win, and it charges them nothing for now.

   The picture

   ChatGPT users1.2 billion a week

   vsvs

   Shops that block agentsAmazon, United

   Step by step

   1. 1 A platform needs buyers and sellers at the same time
   2. 2 Each side joins only if the other is already there
   3. 3 So the platform makes the harder side's entry free or cheap
   4. 4 As that side fills up, the other side follows
   5. 5 Once both sides depend on it, leaving gets hard and charges can start

   The same force, elsewhere today

   Where this chain is also running, in today's other stories.

   • Shopify opens every store to AI agents

     Shopify switched Poppy on for all its stores at once, so agents can reach them without any shop doing work.

   • Amazon plans an index of every product online

     Amazon wants agents to depend on its product list, and blocks rivals' agents meanwhile, to keep both sides on its own site.

   • Schools face bills for Microsoft email

     Microsoft gave schools email for free to win them, and now that their work lives on its cloud, it is starting to charge.

   Where you've seen this

   Credit cards

   a card is useless until shops take it, and shops take it only when enough people carry it

   Games consoles

   a console sells only with games, and studios make games only for consoles people own

   Food delivery apps

   apps paid discounts to both diners and restaurants to grow each side at once

   The catch

   Paying for one side works only while the platform can afford it and the other side follows. OpenAI's earlier Instant Checkout faded within months.

   And the whole of it

   The shopper, the shop and the card company each see only their own step. Whichever side joins first helps set what the others pay later, and most people join without noticing they chose.

1. 07

   Amazon plans an index of every product online

   What happened

   • Amazon plans a data store on every product sold online, even ones it does not sell, Reuters reported. [9]
   • The project, called the World Product Graph, would show that information on Amazon's own site. [9]
   • Its aim is to stop shoppers leaving Amazon to research products on Reddit or Google. [9]
   • Amazon's AI shopping helper, now called Alexa for Shopping, used to be named Rufus. [9]
   • An internal document said: "Many Rufus answers are factually wrong or incomplete." [9]
   • The document gave a release date of 1 October 2027. [9]

   Who is involved

   • Rufus now Alexa for Shopping

   How it unfolded

   1. Now a plan on paper
   2. 1 October 2027 planned release

   Why it matters

   If agents rely on Amazon's index, people shopping with bots would still end up on Amazon's pages. [9]

   Think it through

   1. 1

      Does Amazon say it will happen?

      Amazon said: "Draft, outdated documents often paint an incomplete picture." [9]

   2. 2

      How would Amazon get other shops' data?

      The plan depends on learning how brands lay out products on their own websites, which could mean crawling them. [9]

   The truth

   • The plan comes from one internal document seen by Reuters, not from anything Amazon has announced. [9]
2. 08

   Chinese AI models show signs of banned chips

   What happened

   • Will DePue, a former OpenAI researcher, says top Chinese AI models were trained on Nvidia Blackwell chips. [16]
   • The US bans exports of Blackwell, Nvidia's most powerful AI chips, to China. [28]
   • He named models from DeepSeek, Moonshot AI, MiniMax and Ant Group, all Chinese firms. [16]
   • Epoch AI, a research group, estimates about 2 million advanced chips reached China in 2024 and 2025 by hidden routes. [28]

   Who is involved

   • Will DePue ex-OpenAI researcher
   • DeepSeek Chinese AI lab
   • Moonshot AI Chinese AI lab
   • MiniMax Chinese AI lab
   • Ant Group Chinese tech firm

   Why it matters

   At least 4 bills in the US Congress aim to limit China's access to chips like these. [16]

   Think it through

   1. 1

      Does this prove the law was broken?

      No: the chips may have sat in a data centre outside China, a popular route for Chinese firms. [28]

      Chips abroadallowed there

      ↓→

      Trained from Chinathe model
   2. 2

      Where do such chips sit?

      Epoch AI estimates 275,000 Nvidia AI chips went into Malaysia in 2024 and 2025. [28]

      275,000Nvidia AI chips shipped into Malaysia

   3. 3

      So where is the gap in the rules?

      The rules cover who owns chips and where they sit, and critics want them to cover who uses them. [28]

      Rules: owner and place→Gap: the user

   How it works

   Computer scientists call this a fingerprint: a trace a machine leaves in what it makes.

   A model's first settings are laid out in a pattern that depends on the exact chip, and traces survive training. [16]

   Seen it before

   A typewriter with a worn letter leaves the same flaw on every page it types.

   The truth

   • The finding is one researcher's analysis of published models. [16] A White House adviser said earlier this year that Moonshot used banned chips. [28]
   • Nobody knows yet

     How many banned AI chips are inside China now.

     Epoch AI's figure of about 2 million is an estimate, because the routes are built to hide where chips end up. [28]

3. 09

   Singapore files 41 more chip charges

   What happened

   • Singapore charged Shi Bingzhe, a 34-year-old Chinese citizen, with 41 crimes over servers holding Nvidia chips. [17]
   • Two firms he ran bought 917 servers with restricted Nvidia hardware, prosecutors say. [17]
   • Prosecutors say the firms had reason to believe the servers were linked to crime by their seller. [17]
   • The case now involves 5 people and 6 companies. [17]
   • Malaysia is also checking whether Speedmatrix, a local arm of Singapore's Megaspeed, sent AI chips abroad. [17]

   Who is involved

   • Speedmatrix Megaspeed's Malaysian arm, under check

   How it unfolded

   1. Now 41 charges
   2. 4 January 2027 next court date

   Why it matters

   If smuggling is proven, Nvidia could face blame under US export rules, TechCrunch reports. [28]

   Think it through

   1. 1

      Did the chips actually reach China?

      Neither country's investigation has found that they did, Tom's Hardware reports. [17]

   2. 2

      What has Megaspeed done?

      It hired a US law firm to count all its AI chips and check it followed US rules. [17]

      Hire a law firm→Count every chip→Check the rules

   How it works

   Customs officers call this trans-shipment: goods pass through one country on their way to another.

   Malaysia has required permits since July 2025 to send US AI chips on from its ports. [17]

   Seen it before

   A parcel forwarded from a friend's house hides where it was really going.

   What comes next

   Shi's next court date is 4 January 2027. [17]

4. 10

   Schools face bills for Microsoft email

   What happened

   • Microsoft will end free student email and teamwork tools at schools' first renewal from 1 January 2028. [29]
   • Analyst Mary Jo Foley put rough student prices at $2, $4 or $6 a year, plus $1 for Teams. [29]
   • The new free plan leaves out email, file sharing and Teams. [29]
   • The UK's Department for Education said it was "disappointed" by the news. [29]
   • Rival Nextcloud says some Swiss universities moved 39,500 users to it over one August weekend. [29]

   How it unfolded

   1. Last week plan announced
   2. 1 January 2028 changes start

   Why it matters

   A few dollars a student becomes a yearly bill across a whole school system. [29]

   Think it through

   1. 1

      Why was it free in the first place?

      "Free was the bait on the hook," says Michael Meeks of Collabora, a rival office-software firm. [29]

   2. 2

      Can schools really leave?

      Meeks says a year can be enough to move, as France's University of Lille did. [29]

   3. 3

      So who holds the power here?

      The supplier does once switching is hard, says Element's Patrick Alberts: it can raise prices at will. [29]

   How it works

   Economists call this lock-in: once work lives on one system, leaving costs more than paying.

   Microsoft moved schools' email, files and logins onto its cloud, so switching means moving all of them. [29]

   Seen it before

   A phone plan that was free for a year is hard to drop once all your friends have that number.

   Microsoft gave schools email for free to win them, and now that their work lives on its cloud, it is starting to charge.

   The idea, in story 06: Why an AI shopper is no use until the shops let it in →

   The truth

   • Who gains Rival office-software firms like Nextcloud and Collabora — Microsoft's new school prices hand them an opening pitch to schools, The Register says. [29]
5. 11

   Microsoft's November updates fix old bugs

   What happened

   • Microsoft's November security updates fix 1,240 flaws, 64 of them rated critical, by Computerworld's count. [30]
   • Attackers are already using one of them, a flaw that makes fake messages look real in Outlook. [30]
   • Adding a printer listed in a company directory works again, after failing since October. [30]
   • Some USB audio devices that stopped working on 8 September should also start again. [30]
   • The November updates for Windows 11 list no new known problems. [30]

   How it unfolded

   1. 8 September stopped working
   2. November working again

   Why it matters

   IT teams are told to install the Windows, Office and Exchange fixes first. [30]

   Think it through

   1. 1

      Is anything still broken?

      Yes: on some Windows servers, the update tool still hides why a sync failed. [30]

   2. 2

      Why would Microsoft hide an error?

      It hid those details on purpose to close a security flaw, and there is no published workaround. [30]

      Security flaw→Error details hidden→No workaround

   How it works

   Engineers call this a regression: a change fixes one thing and breaks another that worked.

   October's server updates broke web sign-in on some Exchange email servers, and November's updates fix it. [30]

   Seen it before

   Oiling a squeaky door can leave it swinging open on its own.

   Good news

   • Adding a company printer in Windows works again after Microsoft's November update, after failing since October.
6. 12

   Word drops old layout settings

   What happened

   • Microsoft is retiring some old layout settings in Word, its word-processing app. [31]
   • Documents made or last saved in Word 2010 or earlier may change their layout. [31]
   • Text can reflow, page counts can change, and pictures can move to another page. [31]
   • Microsoft's advice is to export important documents to PDF before the change. [31]
   • The change reaches Word on Windows from December 2026 to January 2027, and Macs in early 2027. [31]

   How it unfolded

   1. December 2026 Windows rollout starts
   2. January 2027 Windows done
   3. Early 2027 Macs

   Why it matters

   Offices with old forms and templates whose pages must stay put will need to check them. [31]

   Think it through

   1. 1

      Why change it at all?

      Microsoft says it wants simpler layout handling and a more consistent experience across devices. [31]

      Old layout options→Retired→More consistent layout
   2. 2

      Which settings are going?

      Microsoft has not named the settings it is retiring. [31]

   How it works

   Software engineers call this backward compatibility: a new version keeps old files working as before.

   Word's Compatibility Mode opened an old file with newer features off, so its layout stayed fixed. [31]

   Seen it before

   A new games console that still plays your old discs is keeping them working.

7. 13

   C++26 stalls over one disputed feature

   What happened

   • C++, a language used to write fast software, may miss this year's new version over one disputed feature. [32]
   • The draft of C++26 failed a vote of national standards bodies in October: 13 for, 9 against, 12 abstaining. [32]
   • The dispute is over Contracts, checks that make code test its own conditions while it runs. [32]
   • Bjarne Stroustrup, who created C++, says the feature was rushed and adds complexity and bloat. [32]
   • Committee member Andreas Fertig says vetoing it this late damages C++ itself. [32]

   Who is involved

   • Bjarne Stroustrup says it was rushed
   • Andreas Fertig says the veto does damage

   How it unfolded

   1. October vote fails
   2. Next week meeting in Brazil

   Why it matters

   Many projects have moved to Rust because C++ makes it easy to create security holes by accident. [32]

   Think it through

   1. 1

      Does a failed vote mean the system is broken?

      No, says Harry Bott of the C++ Alliance: "it's the process working as intended". [32]

   2. 2

      What can the committee do now?

      It can cut Contracts out, move them to a separate document, or revise and vote again. [32]

      Cut themship the rest

      ++

      Separate paperkeep working

      ++

      Revisevote again

   How it works

   Standards bodies work by consensus: a rule passes only when few enough members object.

   ISO, the body that approves C++, lets no more than 25% of votes be against, so the draft failed. [32]

   Seen it before

   A class trip needs nearly everyone to agree, so a few noes can stop it.

   What comes next

   The committee, WG21, meets next week in Buzios, Brazil, to decide. [32]

8. 14

   Mozilla shuts its Solo website builder

   What happened

   • Mozilla is shutting Solo, its AI website builder and web host for small traders, on 30 November. [33]
   • All sites and data on Solo will be deleted that day. [33]
   • About 54,000 people used it to build pages, many with their own web addresses. [33]
   • Mozilla says it told users on 1 October, giving them 60 days to move. [33]
   • Users can download their pages as a zip file, but must save their images separately. [33]

   How it unfolded

   1. 1 October users told
   2. 30 November sites deleted

   Why it matters

   Sole traders who picked Solo to avoid fiddly tech must now move their sites themselves. [33]

   Think it through

   1. 1

      Why is Mozilla closing it?

      Mozilla says experimenting sometimes means "making the decision to bring a product to an end". [33]

   2. 2

      What happens to their web addresses?

      Domains bought through Solo must move to Name.com, a domain seller. [33]

      Solo domain→Moved to Name.com

   How it works

   Analysts call this platform risk: a firm built on someone else's service can lose it at short notice.

   Solo hosted the sites itself, so when Mozilla stopped, its 54,000 users had to find new homes. [33]

   Seen it before

   A market stall on rented ground must pack up when the owner sells the field.

9. 15

   Palantir wins UK police deal as NHS app lags

   What happened

   • Three police forces in the east of England gave Palantir, a US data-analysis firm, a 10.1 million pound contract. [15,34]
   • The five-year deal covers Bedfordshire, Cambridgeshire and Hertfordshire police and runs to October 2031. [15]
   • London's mayor, Sadiq Khan, had blocked a separate 50 million pound Palantir deal with London's police. [15]
   • In the NHS, Palantir's Cancer 360 app is used by only 5 or 6 hospital trusts, official figures show. [34]
   • It runs on a 330 million pound NHS data platform that Palantir operates. [34]

   How it unfolded

   1. Now contract awarded
   2. October 2031 contract ends

   Why it matters

   MPs warned that Palantir's growing role in UK public services is "an unacceptable point of weakness". [15]

   Think it through

   1. 1

      Who is advising Palantir?

      Former London police chief Cressida Dick has been a paid Palantir adviser since September 2025. [15] She says she never lobbied for it. [15]

   2. 2

      Does the cancer app work?

      Two trusts praised for their pilots logged zero users in a year, NHS England's own reply shows. [34]

   3. 3

      What will Britain do with the NHS contract?

      Reports say it may renew it after February 2027 while looking for an alternative to Palantir. [34]

   How it works

   Statisticians call this attribution: credit for a result needs evidence the tool caused it.

   Cancer waits in England improved overall, but with so few trusts using Cancer 360, the app cannot take the credit. [34]

   Seen it before

   A team wins after buying new boots, but the boots may not be why.

10. 16

    UK experts say tech firms do the minimum

    What happened

    • Experts told UK lawmakers that tech firms do only the minimum under the Online Safety Act. [35]
    • The Act is the UK's law requiring platforms to protect users, especially children, from harmful posts and images. [35]
    • Law professor Clare McGlynn tested searches for extreme porn. [35] One banned term showed results only from page 10, a close variant by page 3. [35]
    • The charity Refuge says its reports of abusive posts succeeded 53.7% of the time before November 2025. [35]
    • In the past 12 months that fell to 18.2%, and none of its reports to Meta succeeded. [35]

    Who is involved

    • The UK set the law
    • Platforms must protect users

    Why it matters

    Abuse victims who report harmful posts are finding it harder to get them taken down. [35]

    Think it through

    1. 1

       What do the companies say?

       Meta did not reply on its takedown rate, and Google, Microsoft and TikTok did not comment. [35]

    2. 2

       Does any country do better?

       McGlynn pointed to British Columbia in Canada, where victims can get fast orders to remove images. [35]

       Victim applies→Fast order→Image removed

    How it works

    Legal scholars call this creative compliance: following a rule's words while missing its purpose.

    Search engines push down exact banned terms but not close variants, so the material stays easy to find. [35]

    Seen it before

    A pupil who tidies only the desk the teacher checks leaves the rest of the room messy.

11. 17

    HPE mixes in hard drives as flash costs soar

    What happened

    • HPE, a big maker of business computers, will sell supercomputer storage that mixes hard drives with flash drives. [20]
    • It says the mix gives up to 41% more data speed per storage rack. [20]
    • Prices of large flash drives, called SSDs, have soared as AI data centres buy up parts. [20]
    • Flash drive prices are rising faster than those of old-style hard drives. [20]

    Why it matters

    Research labs can spend less on storage and more on the computers that do the work, HPE says. [20]

    Think it through

    1. 1

       Why are flash drives so dear?

       Demand for AI equipment is pushing up the price of memory and many other parts. [20]

    2. 2

       When can labs get it?

       The new setups go on sale in early 2027. [20]

    How it works

    Engineers call this tiered storage: fast, dear drives hold busy data and slow, cheap ones hold the rest.

    HPE's new racks pair fast SSDs with new 32 TB hard drives to cut what labs spend. [20]

    Seen it before

    You carry today's books in your bag and leave the rest on a shelf at home.

12. 18

    Cornelis takes on Nvidia's networking

    What happened

    • Cornelis, a networking firm spun out of Intel in 2020, launched new 800G network cards and switches for AI. [36]
    • It pitches them against InfiniBand, Nvidia's networking for big AI computers. [36]
    • Cornelis says its card is 21% faster than InfiniBand at passing a small message, in its own tests. [36]
    • The cards work with chips from AMD, Intel, Nvidia and Qualcomm, with systems due in early 2027. [36]

    Who is involved

    • Cornelis the challenger
    • Nvidia InfiniBand the one to beat

    How it unfolded

    1. Now cards launched
    2. Early 2027 systems from Dell, Lenovo, Supermicro

    Why it matters

    Firms building AI clusters get another choice besides Nvidia for the network that links their chips. [36]

    Think it through

    1. 1

       Why does the network matter so much?

       Thousands of chips swap data constantly, so a slow network leaves expensive chips sitting idle. [36]

       Slow network→Chips wait→Paid-for power wasted
    2. 2

       Does 21% faster mean jobs finish 21% sooner?

       No: how much sooner a job ends depends on the work, TNW notes. [36]

    How it works

    Computer engineers call this latency: the delay before a message arrives.

    Lower latency means each chip waits less for data from the others, Cornelis says. [36]

    Seen it before

    A slow waiter leaves the cooks' dishes going cold on the counter.

13. 19

    Hospital AI reads 16 years of heart records

    What happened

    • Dyania Health, a US start-up, used AI to read up to 16 years of records for over 7,800 heart patients. [37]
    • It ran entirely on Cleveland Clinic's own Nvidia servers, so records never left the hospital. [37]
    • It answered 88 questions about each patient who had a heart valve replaced between 2005 and 2021. [37]
    • The data built a model that predicts the risk of a new valve failing after 3, 5 and 10 years. [37]
    • Doctors still check the AI's evidence and make every decision. [37]

    Who is involved

    • Doctors make every decision

    Why it matters

    Much of a patient's history sits in doctors' notes, and Nvidia says reading them by hand would take clinicians years. [37]

    Think it through

    1. 1

       Does it make care better?

       Not shown yet: the study shows it can build data and predict risk, and needs outside checks. [37]

       Builds datashown

       vsvs

       Better carenot yet shown
    2. 2

       Who stands to gain?

       Cleveland Clinic is also an investor in Dyania, which it disclosed last year. [37]

    How it works

    Doctors call this a chart review: reading patients' records to answer a research question.

    Synapsis links each answer to the record it came from, so a doctor can check it. [37]

    Seen it before

    A friend who marks the page number beside every quote lets you check them fast.

14. 20

    LangChain doubles its value to $3 billion

    What happened

    • LangChain, which makes software for building and running AI agents, raised $300 million at a $3 billion value. [38]
    • That is more than double its $1.25 billion value about a year ago. [38]
    • The round was led by CapitalG, with Singapore's GIC also investing. [38]
    • It started in 2022 as a free open-source project for tasks early AI models found hard. [38]
    • It now claims 7,000 customers for its LangSmith tool, including Nvidia, Toyota and Salesforce. [38]

    Who is involved

    • CapitalG led the round
    • GIC Singapore investor

    How it unfolded

    1. 2022 open-source project
    2. Now agent management

    Why it matters

    Its pitch is to firms that want agents not tied to one AI model, cloud or company. [38]

    Think it through

    1. 1

       Why are investors paying so much?

       LangChain had a head start in managing agents, a crowded but busy field, TechCrunch says. [38]

    2. 2

       Who are its rivals?

       The big clouds and AI labs push their own agent tools, and start-ups like Mintive pitch the same. [38]

       Big clouds→AI labs→Start-ups

    How it works

    Investors call this a valuation: the price paid for a slice sets the whole company's worth.

    Investors paid $300 million for a stake priced to value LangChain at $3 billion. [38]

    Seen it before

    If one slice of a cake sells for a lot, people say the whole cake is worth a lot.

15. 21

    Start-up plans data centres powered by the sea

    What happened

    • Drogue, a six-month-old US start-up, raised $20 million to run floating data centres on ocean heat. [19]
    • Its system uses the gap between warm surface water and cold deep water, at about 4C, to make electricity. [19]
    • Warm water boils ammonia, the gas spins a turbine, and cold water turns it back to liquid. [19]
    • The idea, ocean thermal energy conversion, was first developed in the 1950s. [19]
    • Drogue plans to put a prototype on a ship at sea in early 2028. [19]

    How it unfolded

    1. Now $20 million raised
    2. Early 2028 prototype at sea

    Why it matters

    Its boss says data centres take 8 to 10 years to plan with grids, far too slow for AI. [19]

    Think it through

    1. 1

       Why put the servers at sea?

       Sending power to shore needs long cables, but sending data off a ship is cheap. [19]

       Power to shorecostly cables

       vsvs

       Data to shorecheap
    2. 2

       Has this been tried before?

       It was tested in the 1980s and dropped, but its boss says pipes and parts are better made now. [19]

    How it works

    Physicists call this a heat engine: it turns a difference in temperature into useful work.

    Drogue's loop runs on the gap between surface water and water drawn from 1,000 metres down. [19]

    Seen it before

    Steam from a boiling kettle can spin a little paper windmill held above it.

16. 22

    GridAI buys a gas producer to power AI

    What happened

    • GridAI Technologies agreed to buy Pronghorn Resources, a gas producer in Wyoming, to power AI data centres. [18]
    • Pronghorn pumps nearly 100 million cubic feet of gas a day from about 3,200 wells. [18]
    • GridAI aims for its first 300 megawatts of power in 2027, but has named no customers. [18]
    • The deal is paid in shares, and GridAI was worth about $31 million in October. [18]

    How it unfolded

    1. Now deal agreed
    2. 2027 first 300 megawatts

    Why it matters

    Wyoming is drawing AI money: Google plans a $17 billion data centre there that will run mainly on gas. [18]

    Think it through

    1. 1

       Can a small firm build all this?

       Not yet: the plants still need to be built and paid for, TNW notes. [18]

    2. 2

       Has this gone wrong for anyone?

       Last week Crusoe gave up its Wyoming site to Google, reportedly after failing to find customers. [18]

       Crusoe's site→No customers→Handed to Google

    How it works

    Energy economists call this behind-the-meter power: a plant feeds one customer directly instead of the grid.

    GridAI's plants would feed data centres directly, so they need not wait years for a grid connection. [18]

    Seen it before

    A shop with its own generator keeps its lights on without waiting for the power company.

17. 23

    Drone strikes push up Russian data-centre insurance

    What happened

    • Insuring data centres in Russia has become far dearer since Ukrainian drones began hitting them in October. [39]
    • Cover cost 0.03% to 0.06% of a building's value in the first half of 2026, Kommersant reported. [39]
    • Rates have since jumped to 1% to 5%, more than 100 times higher in some cases. [39]
    • Ukrainian drones hit data centres of Yandex, Sberbank and VK, three big Russian firms, in October. [39]
    • Russia's central-bank-owned reinsurer set a base rate of 1% and now decides some sites one by one. [39]

    Who is involved

    • Yandex hit
    • Sberbank hit
    • VK hit

    Why it matters

    Russian firms are cutting their cover or leaving some risks out of their policies, experts say. [39]

    Think it through

    1. 1

       Who struck first?

       Russian troops had earlier attacked data centres in Ukraine, Meduza notes. [39]

    2. 2

       So who carries the risk now?

       With less cover, more of the cost of a strike falls on the data-centre owners themselves. [39]

    How it works

    Insurers call this risk-based pricing: the premium rises with how likely a loss is.

    Drone strikes made a loss far more likely, so the top rate rose from 0.06% to 5%. [39]

    Seen it before

    Bike insurance costs more in a town where bikes are stolen every week.

18. 24

    Russia finishes its self-driving law

    What happened

    • Russia's government approved a final bill to regulate self-driving cars, trucks, taxis and trams. [40]
    • Russia's Transport Ministry began drafting it in 2021, and ministers approved it on 12 November. [40]
    • It sets rules on blame in crashes, on protecting the vehicles' software and on the systems they must carry. [40]
    • Self-driving vehicles have run in Russia since 2018, usually with a backup operator on board. [40]
    • Moscow's first self-driving tram has run on route 10 since September 2025. [40]

    Who is involved

    • Backup operator usually on board

    How it unfolded

    1. 2021 drafting starts
    2. 12 November final bill approved

    Why it matters

    The bill decides who is liable when a self-driving vehicle crashes. [40]

    Think it through

    1. 1

       Is it law yet?

       No: it still has to go to the State Duma, the lower house of Russia's parliament. [40]

       Government approves→State Duma votes→Law
    2. 2

       How many driverless trams will Moscow run?

       Its mayor promised 15 by the end of 2026 and 25 by 2027. [40]

       End of 202615 trams

       202725 trams

    How it works

    Lawyers call this liability: deciding who pays when something goes wrong.

    With no driver at the wheel, the bill must set who answers for a crash. [40]

    Seen it before

    When a school bus crashes, rules already say whether the driver or the bus company pays.

19. 25

    Apple's HomeView and first home devices arrive

    What happened

    • Apple's HomeView smart display went on sale this week. [41]
    • Apple's first Home Collection devices, made with LG and Schneider Electric, are open for pre-order, starting with LG's. [41]
    • LG cameras at $199.95 and $299.95 and a $249.95 doorbell are due to ship from 11 to 17 December. [41]
    • HomeView runs no apps of its own: it mirrors the ones on a nearby iPhone or iPad. [42]
    • The devices use Thread, a home network, so they stay connected if the internet drops. [41]

    Who is involved

    • LG cameras and doorbell
    • Schneider Electric thermostat and sensor

    Why it matters

    Buyers need an iPhone or iPad close by for HomeView to work. [42]

    Think it through

    1. 1

       Why tie it to the iPhone?

       Apple says its home products must be "intelligent, personal and private", and the phone holds personal data. [42]

    2. 2

       Do the devices do anything special?

       No exclusive features, The Verge says: Apple mainly makes setup a tap with an iPhone. [41]

    How it works

    Economists call this an ecosystem: products that work best together keep buyers with one brand.

    HomeView borrows its apps from an iPhone, so it needs an Apple phone or tablet nearby. [42]

    Seen it before

    A games controller that only fits one console makes you stick with that console.

20. 26

    Disney turns to the AI it once fought

    What happened

    • Disney will put Ozzy Fox, a children's series animated with AI tools, on Disney+ in December, Bloomberg reported. [43]
    • It was made with Animaj, a French start-up whose tools turn drawings into 3D animation. [43]
    • When it first appeared on YouTube in July, some viewers called it "AI slop". [43]
    • Disney is also building AI for budget reports and testing AI viewing suggestions on Disney+. [43]
    • Disney once sued AI firms over its characters, and still guards them closely. [43]

    How it unfolded

    1. Before sued AI firms
    2. Now uses AI itself

    Why it matters

    The push is led by Josh D'Amaro, Disney's boss since this year. [43] He says Disney is being "very aggressive" on new tech. [43]

    Think it through

    1. 1

       Why the change of heart?

       OpenAI shut its Sora video app in D'Amaro's first week, ending a planned deal to license Disney characters to it. [43]

       Sora shut down→Licence deal ends→Disney builds its own
    2. 2

       So who gets to use Disney's characters?

       Adobe can build AI models on them under Disney's control, rather than outside AI firms. [43]

    How it works

    Lawyers call characters intellectual property: a creation the law lets its owner control.

    Disney lets Adobe build AI on its characters, but keeps the say over how they are used. [43]

    Seen it before

    Lending your football to one friend, on your rules, is not giving it to the whole street.

21. 27

    Anthropic shows how much safety scores can swing

    What happened

    • Anthropic, maker of the Claude AI, published advice on building AI safety tests, with some unflattering numbers of its own. [44]
    • In self-harm test chats, Claude Fable 5.1 behaved properly 94% of the time in the Claude app, after a fix. [44]
    • Through Anthropic's API, which developers use without that fix, the figure was 60%. [44]
    • Its simple one-question tests are now too easy: every recent model scores 94% to 98%. [44]
    • Anthropic says it stopped many tests giving models internet access after its models breached 3 companies in cyber tests. [44]

    Why it matters

    A claim that one model is safer is only as good as the test behind it, the guide says. [44]

    Think it through

    1. 1

       Why did the same model score so differently?

       The app adds a hidden instruction that stops the bad replies, and the API does not have it. [44]

       Apphidden instruction

       vsvs

       APIno instruction
    2. 2

       Was Opus 5.5 safer than Opus 5?

       Opus 5 scored 98% and Opus 5.5 scored 94% on a scam test, too close to call, Anthropic says. [44]

       % on a scam test

       Opus 598

       Opus 5.594

    How it works

    Statisticians call this uncertainty: every score comes with a range it could really fall in.

    Anthropic says the gap between 98% and 94% was too small to show either model was safer. [44]

    Seen it before

    Two runners timed by hand a moment apart may really have tied.

22. 28

    AI unlocks Godel's lost notebooks

    What happened

    • An AI helped a Finnish logician read 1,009 pages of notebooks by Kurt Godel, a great 20th-century mathematician. [45]
    • Godel wrote them in the early 1940s, much of it in an old German shorthand that few can read. [45]
    • Jan von Plato's team read about a page a week, until he turned to Claude, Anthropic's AI, in September. [45]
    • Within days, Claude produced a 650-page explanation that von Plato checked page by page. [45]

    Why it matters

    Godel had found far more than anyone knew, including results not known until the 1970s and 1980s. [45]

    Think it through

    1. 1

       Had Godel beaten a famous later result?

       Nearly: he was one step from a proof that helped win Paul Cohen maths' top prize in 1966. [45]

    2. 2

       Can the AI's work be trusted?

       Von Plato ran the rebuilt proof through Coq, a program that checks proofs, and it passed. [45]

       Claude rebuilds the proof→Coq checks it→It passes

    How it works

    Mathematicians call this formal verification: a program checks every step of a proof.

    Of 663 theorems Claude found, about 400 held as written, about 200 needed repair and about 50 were false. [45]

    Seen it before

    A calculator checks your sums, so a slip in your head gets caught.

    Good news

    • The notebooks also hold 14 themes in logic that nobody has pursued yet. [45]
    • An AI helped a logician read 1,009 pages of Kurt Godel's notebooks within days, work that had gone at a page a week. It showed Godel had proved hundreds of theorems nobody knew about.
23. 29

    Universities drop Flock's plate cameras

    What happened

    • At least 10 university campuses in 7 US states have dropped Flock, a maker of number-plate cameras, since April. [46]
    • They include Stanford, Cornell, Penn State and Emory. [46]
    • The cameras record where a car was at a given time. [46]
    • Flock says its cameras help solve crimes, and denies they feed mass surveillance. [46]

    Who is involved

    • Stanford dropped Flock
    • Cornell dropped Flock
    • Penn State dropped Flock
    • Emory dropped Flock

    Why it matters

    Students fear the records can be used to target vulnerable people, says Stanford student Angela Nguyen. [46]

    Think it through

    1. 1

       Are the cameras actually gone?

       Often not: Stanford and Binghamton are switching to plate readers from Genetec and Axon instead. [46]

    2. 2

       What do students want now?

       Many now push their universities to drop every plate-reading camera, not just Flock's. [46]

       Drop Flock→Drop every plate reader

    How it works

    Privacy experts call this function creep: data gathered for one job gets used for others.

    Houston police searched UNC Charlotte's campus cameras 183,000 times between March and May. [46]

    Seen it before

    A class photo taken for the yearbook ends up on a poster nobody agreed to.

24. 30

    Gen Z has not made up its mind on AI

    What happened

    • 73% of Gen Z adults have not made up their minds about AI, a Milltown Partners survey found. [47]
    • Half expect AI to harm society, close to baby boomers at 50% and Gen X at 48%. [47]
    • Millennials are the most upbeat, with 39% expecting harm. [47]
    • 46% of Gen Z use AI every day, second only to millennials. [47]
    • 30% sometimes feel guilty using it, more than double the rate among people over 30. [47]

    Why it matters

    Firms and politicians want young people's views, after AI and data centres shaped the US midterms. [47]

    Think it through

    1. 1

       Why use it every day and still feel bad?

       Many say AI makes them worse at thinking for themselves and makes jobs harder to find. [47]

       Daily use→Worry about thinking and jobs→Guilt
    2. 2

       Are they the most afraid of AI wiping us out?

       No: they worry less about that than older generations do. [47]

    How it works

    Psychologists call this ambivalence: holding good and bad feelings about the same thing at once.

    Gen Z leans on AI daily, while many fear it weakens their own thinking. [47]

    Seen it before

    Loving sweets while knowing they are bad for your teeth pulls you two ways.

25. 31

    AI helps a 1994 Jeep pass its smog test

    What happened

    • A California man used two AI tools, Claude and Codex, to get his 1994 Jeep Cherokee through a smog test. [48]
    • The 32-year-old car kept failing for making too much NOx, a polluting gas, at test speeds. [48]
    • Each test cost nearly $100, and his own guesses first made it fail at more speeds. [48]
    • Claude told him the engine computer talks on a Chrysler data line an Arduino Nano board can read. [48]
    • The parts cost no more than $15 in all. [48]

    Who is involved

    • Claude gave guidance
    • Codex wrote the code

    Why it matters

    Muhammad Zulali could tune the engine from real readings, instead of adjusting blind. [48]

    Think it through

    1. 1

       What did the AI actually do?

       Claude explained the old engine's wiring and signals, and Codex wrote the code he needed. [48]

    2. 2

       Did it work?

       Yes: he fixed the emission problems through the laptop link, Tom's Hardware reports. [48]

       Laptop link→Emissions fixed→Test passed

    How it works

    Engineers call this reverse engineering: working out how a machine works from the outside.

    Zulali read the old engine computer's signals to learn its settings, then changed them. [48]

    Seen it before

    Working out a recipe by tasting the dish and trying again until it matches.

26. 32

    A new project runs macOS on PCs

    What happened

    • A new free project called Stela lets some PCs run macOS, Apple's computer system, again. [49]
    • Apple moved its Macs from Intel chips to its own Arm-based chips from 2020. [49]
    • That ended Hackintoshes, ordinary PCs set up to run macOS, until Stela began translating Apple-chip code. [49]
    • It boots macOS 27 on a PC by presenting a virtual M2 Mac Mini. [49]
    • Developer Tyler Smith warns it is "slow and not very stable". [49]

    How it unfolded

    1. Before 2020 Intel chips in Macs
    2. 2020 on Apple's own chips

    Why it matters

    PC owners can try macOS without buying a Mac, though the tool is rough for now. [49]

    Think it through

    1. 1

       Does it change macOS itself?

       No: Smith says the macOS image is not modified at all. [49]

    2. 2

       How does Apple's code run on a PC chip?

       Stela translates Arm instructions into x86-64 ones, the kind Intel and AMD chips use. [49]

       Arm codeApple

       ↓→

       x86-64 codeIntel and AMD

    How it works

    Computer scientists call this a translation layer: software that turns one chip's instructions into another's.

    Stela translates some code before it runs and some as it runs, so macOS acts as if on a Mac. [49]

    Seen it before

    A live interpreter lets two people talk though neither speaks the other's language.