Britain's competition watchdog moves to give Microsoft a special status that would let it set rules on how Windows, Office and Teams are sold

The rest of the day

38 more stories on this beat.

Each with its own sources. None of these is a link to the story above.

1. 02

   China wants licences to export its biggest AI

   China's commerce and science ministries proposed on Friday that 23 kinds of AI technology need a licence to leave the country. [5][6] They include the inside workings of large models, AI agents, chip designs and big computer clusters. [6] The model rules would cover models with at least 2 trillion parameters, or ones ranked in the world's top ten, and Moonshot's Kimi K3 and Alibaba's Qwen3.8-Max are above that size. [5] Open-source code is exempt, and the public can comment until 15 November. [5][7]

   Why it matters — Firms that build on Chinese models now face a new question over whether those licences can be pulled, a Council on Foreign Relations fellow said. [15] The draft comes a month after the US and China agreed to cut tariffs and start talks on AI. [7]

2. 03

   Taiwan to make AI chip sales to China a crime

   Taiwan will add AI servers and cutting-edge chips to its list of controlled exports, its economy minister Kung Ming-hsin told Bloomberg. [8] Once the rule takes effect, shipping that hardware to China will be a crime under Taiwan's Foreign Trade Act. [8] Taiwan is a centre of world chipmaking, and the rule brings it closer to US limits on sales to China. [8][9] Semafor reports the move is meant to win favour with the US government. [9]

   Why it matters — The rule aims at the black market that sends banned chips to Chinese buyers through middlemen. [9] Chinese chipmakers such as Huawei are meanwhile winning more of China's own AI chip market from Nvidia, Reuters reported. [9]

3. 04

   Anthropic drafts its own AI safety law

   Anthropic, the company behind the Claude chatbot, has quietly shown US lawmakers a draft law on AI safety, Politico reported on Friday. [21] It would let the US government stop a model only if it posed an imminent risk of more than $1 billion in damage, or of killing or badly injuring more than 50 people. [21] Outside testers registered with the US Commerce Department would check each lab's models, and the labs could pay them. [21] Only a federal judge in Washington, DC could order a model halted. [21]

   Why it matters — The draft would bar every other US agency from suspending a model, which leaves the US Commerce Department and one court as the only route. [21] Anthropic did not immediately respond to Reuters. [21]

4. 05

   Open AI models now carry most of one gateway's work

   Open-weight models, which anyone can download and run, handled more than seven in ten tokens, the small chunks of text AI reads and writes, on Vercel's AI Gateway in September. [15] Vercel's gateway lets companies buy AI from many makers through one connection. [15] Open models' share of spending there rose from one dollar in twenty in June to nearly one in four. [15] A small new model called Jev reached 35% of paid teams within two weeks, the fastest Vercel has seen. [16]

   1 in 20June

   ↓→

   nearly 1 in 4September

   Why it matters — The figures cover only one gateway, and they miss companies that buy straight from OpenAI or Anthropic. [16] Chinese open models from DeepSeek and others are the strongest open ones, researchers told Rest of World. [17]

5. 06

   New York council questions SpaceX over Grok

   SpaceX's AI policy chief, Gabriela Castillo Madrid, answered New York City Council members by video on Friday. [26] SpaceX owns Grok, the chatbot on X, since it merged with Elon Musk's AI company in February. [26] The council speaker, Julie Menin, said she had found a sexual image of herself made with Grok that morning. [26] Castillo Madrid said the company has 'strict controls' and would look into it. [26] SpaceX had at first avoided a subpoena to appear. [26]

   Why it matters — A former employee, Devin Kim, told the council only three staff worked on AI safety in his time there. [26] The council says it will use what it hears to write new rules. [26]

6. 07

   Poland summons platforms over school threats

   Poland summoned Google, Meta, TikTok, Discord and Telegram on Friday after knife attacks at schools last week. [42] A school employee was killed and about 15 people were hurt, and threats against schools spread online afterwards. [42] Polish officials flagged 89 posts as illegal between 7 and 12 October, and only 30 were taken down, the digital affairs minister Krzysztof Gawkowski said. [42] X was invited and did not attend. [42]

   30 of 89

   posts flagged, and the ones taken down

   Why it matters — A law that would let Poland order illegal posts removed at once, and fine platforms, now waits for President Karol Nawrocki's signature. [42] He vetoed an earlier version in January over free speech. [42]

7. 08

   Waymo's robotaxis head to Toronto

   Waymo, the robotaxi company owned by Google's parent Alphabet, said on Friday that its cars will start driving Toronto's streets, with a person at the wheel making maps first. [27][28] It wants its first paying riders there in 2028. [28] Waymo now runs robotaxi services in 15 US cities and is testing in London and Tokyo. [27] Toronto will be its most wintry city yet. [27]

   Why it matters — Ontario's only rules for self-driving tests run out in October 2027 and allow no paid service, so the 2028 date depends on the province writing new ones. [28] Toronto's mayor has said she will back Waymo only if it does not cost jobs. [28]

8. 09

   London drivers push back on robotaxis

   Ride-hail drivers plan to protest outside the London base of Wayve, a British self-driving company that supplies Uber, from Saturday. [29] A London Assembly committee report this week asked that Transport for London, the city's transport body, get full power to allow or stop robotaxis. [29] Today the UK government grants those permits, and Transport for London can only agree or refuse. [29] More than 100,000 people are on Uber's London waiting list for a ride. [29]

   Why it matters — A think tank says more than 400,000 UK drivers, 121,000 of them in London, could find much of their work gone. [29] Uber says drivers and robotaxis will grow side by side for some time. [29]

9. 10

   Nvidia buys stakes in five UK start-ups

   Nvidia, which makes the chips most AI runs on, said in London on Wednesday that it is taking stakes in five British start-ups. [30] They are the biotech firms Prima Mente and Boltz, Foresight, Hiverge and Riverlane, a quantum computing company. [30] Those five and others Nvidia did not name will share $1.16 billion. [30] Nvidia has joined 116 funding rounds this year, up from 12 in 2022, according to Dealroom, a data firm. [30]

   Why it matters — Nvidia has also put money into OpenAI, Anthropic, Intel and xAI. [30] Its boss Jensen Huang told the audience that the UK finds too many reasons to fail. [30]

10. 11

    iPhone Duo pre-orders sell out on day one

    Apple's first folding iPhone, the iPhone Duo, went on pre-order on Friday, and the phones set aside for launch-day pickup sold out quickly. [52] Some online delivery dates slipped to late January, and Apple called the demand 'nothing short of extraordinary'. [52] The Duo starts at $1,999 and passes $3,000 with the most storage. [52][53] It reaches shops on 23 October, and T-Mobile is offering $1,000 off with a trade-in or a new line. [54]

    Why it matters — Folding phones are about 2% of phones sold this year but nearly 7% of the money spent, IDC, a market research firm, estimates. [52] Apple is betting that its biggest spenders will pay more, analysts told CNBC. [52]

11. 12

    China's CXMT makes nearly $20 billion

    CXMT, a Chinese maker of memory chips, expects a profit of 131 to 132.5 billion yuan, about $19.7 billion, for January to September. [11] A year earlier it lost 5.3 billion yuan in the same months. [11] Its sales rose more than eightfold, to 265 to 270 billion yuan. [11] AI computers need so much memory that chips for phones and PCs are short and prices have jumped. [11]

    Jan-Sep 202532.1 bn yuan

    Jan-Sep 2026265 bn yuan

    Why it matters — Its chairman wants the company to buy back up to 2 billion yuan of its own shares. [11] Outside testers have not checked CXMT's claim that its newest memory is in mass production. [11]

12. 13

    Taiwan raises $4 billion for a Texas supplier park

    Taiwan is raising at least $4 billion through an electronics makers' group to build an industrial park in Texas. [10] It is meant for the small firms that big chipmakers such as TSMC and Foxconn depend on. [10] Taiwan's government will also guarantee loans for those firms and help them with US visas and permits. [10] The plan may cover more than one site, at $2 billion to $4 billion. [10]

    Why it matters — TSMC has spent $165 billion on its Arizona site, but its smaller suppliers cannot afford to move alone. [10] The US government has not commented on the plan. [10]

13. 14

    A Chinese start-up shows a key chip-machine part

    Langdao Silex, a Chinese start-up founded in 2024, showed a prototype light source for extreme ultraviolet (EUV) machines at a Shanghai fair. [12] EUV machines print the finest patterns on the most advanced chips, and China cannot buy them from ASML, the Dutch company that makes them. [12] Langdao says its source gives over 15 watts and that 90% of its parts are made in China. [12] It needs three to five more years to sell it. [12]

    Why it matters — ASML's production machines use sources of about 250 watts or more, so the gap is large. [12] Langdao has not said where it measured its 15 watts, which affects how it compares. [12]

14. 15

    Thai police cut a cable feeding a scam compound

    Thai authorities found a 350-metre fibre-optic cable strung across the border from a village in Kanchanaburi province into Myanmar. [49] It carried about 6 terabytes of Thai internet data a day, enough for 300 to 400 people. [49] Officials say it kept a suspected scam compound in Myanmar online, along with an illegal casino. [49] A man was paid about $20 a day to let the cable run through his house. [49]

    Why it matters — The report calls Myanmar the centre of the world's online scam trade, and many people working in the compounds are trafficked there, groups that help them say. [49] Scam bosses, mainly Chinese, now run smaller operations in hotels and villas instead of big compounds, the report says. [49]

15. 16

    CERN and Mistral plan an AI model for physics

    CERN, Europe's particle physics laboratory, and Mistral, a French AI company, agreed on Thursday to build a model called Flare. [13][14] It will learn from billions of simulated particle collisions in the Large Hadron Collider, the world's largest particle accelerator. [13] It is meant to replace a chain of separate tools that turn detector signals into data physicists can use. [14] Neither side gave a budget or a date. [13][14]

    Why it matters — CERN already holds about one million terabytes of collider data, and the number of collisions will rise tenfold after an upgrade due around 2030. [14] The deal gives Mistral one of the world's largest science datasets. [14]

16. 17

    Banks and shops block outside AI agents

    Charles Schwab, a big US investment firm, is blocking outside AI agents such as Meta's Muse from moving clients' money, its chief executive told Bloomberg Television. [40] An AI agent is a program that acts for a person, here using the client's password. [40] Schwab will let outside AI only check balances, and wants clients to use its own assistant, Charley. [40] JPMorgan Chase's boss said the same this week, and Amazon and eBay have blocked shopping agents. [40][41]

    Why it matters — Nobody yet knows who pays if an agent moves money wrongly, Schwab's chief said. [40] Cheq, a firm that checks shopping agents, counted 100 times more agent visits to shops in 45 days. [41]

17. 18

    Replit says it dropped Anthropic

    Replit, an app that lets people build software in their web browser, no longer works with Anthropic, its chief executive Amjad Masad said on Thursday. [18] He said OpenAI's models now do better on Replit's tests and cost less. [18] Replit announced a deal with OpenAI in August. [18] A person who knows the deal told TechCrunch that the Anthropic partnership is still in place. [18]

    Why it matters — Replit used Anthropic's Claude models as recently as late September, its own blog shows. [18] Masad said 15% of Replit's revenue now comes from business contracts. [18]

18. 19

    AI reviewers delay the next PostgreSQL

    The team behind PostgreSQL, a popular free database, has pushed its next version back to around 19 November. [33] It has also pulled eight to ten big features. [33] Peter Eisentraut, a lead developer, said AI tools that review code found a long list of small flaws, which would once have been found over five years. [33] Even a three-line fix takes time to handle, and with a long list the team ran out of time, he said. [33]

    Why it matters — The pulled features will likely go into version 20 next year, which had been planned as a release of fixes only. [33] Eisentraut said the extra checks will make version 19 very robust. [33]

19. 20

    487 researchers want hacking AIs kept offline

    An open letter signed by 487 AI researchers, among them Geoffrey Hinton and Yoshua Bengio, says AI that can carry out cyberattacks should be trained only on computers cut off from the internet. [22] They want a new international body to set safety standards and to receive reports when AI escapes its controls. [22] Their demands go further than some recent calls to regulate AI. [22]

    Why it matters — In October last year the two signed a letter warning against building super-intelligent AI. [22] Earlier this week Microsoft's president said AI is less regulated than washing machines or milk. [22]

20. 21

    MI5 sets up a centre to judge AI threats

    MI5, Britain's domestic security service, is setting up an AI Threat Assessment Centre, its head Ken McCallum announced in his yearly speech on 13 October. [55] It brings in experts from the AI Security Institute, the National Crime Agency and the National Cyber Security Centre. [55] Its job is to give the UK government one agreed view of AI risks, he said. [55] McCallum said MI5 has caught terrorists trying to use AI to plan attacks. [55]

    Why it matters — The misuse he fears most, AI helping make biological weapons, has not happened yet, he said. [55] He told tech start-ups to expect foreign spies and to tighten their security. [55]

21. 22

    A conservative group sets out its AI plan

    Humans First, a conservative group that wants tighter limits on AI, published an 'America First AI Policy' on Friday. [23] It calls for tougher export controls on China, independent testing of big AI companies and a ban on AI that improves itself. [23] It also wants to stop the US government using Flock cameras and to pause H-1B visas for foreign workers. [23] Its executive director advised Trump's 2020 campaign. [23]

    Why it matters — Its plan goes much further than the Trump administration's own AI rules. [23] Its chair has been touring the US against new data centres. [23]

22. 23

    Anthropic's staff expect AI to do their jobs

    Anthropic published a survey of 1,334 of its own staff beside surveys of US adults on Friday. [24] Most of its staff said AI will do at least 80% of their work by 2030, against about 5% of US workers. [24] Most of them expect US unemployment above 12% in 2030; it is 4.2% now. [24] Their middle guess for US economic growth in 2030 was 17.1%, against 5.4% for the public and 2.2% growth last year. [24]

    Last year, actual2.2%

    US adults' guess for 20305.4%

    Anthropic staff's guess for 203017.1%

    Why it matters — Anthropic's chief economist says the company is about 18 months ahead of other firms in using AI. [24] A Yale study found AI has so far changed jobs without any link to unemployment. [24]

23. 24

    A tracker counts AI-linked job cuts

    The Center for AI Safety, a research group, runs a dashboard of US job cuts that companies linked to AI. [44] It counts just over 117,000 such cuts this year, worth about $8.5 billion a year in lost wages. [44] Its estimate for California is just over 27,000. [44] The counts come from layoff announcements that mention AI, which does not prove AI replaced each worker. [44]

    Why it matters — In 2025 about one layoff in 20 was publicly linked to AI. [44] The dashboard does not count jobs AI creates. [44]

24. 25

    Irish tech workers join a union

    Ireland's union for tech workers has grown fivefold since it started in 2023, to more than 1,350 members. [48] Ireland holds the European headquarters of most big tech firms, and more than 170,000 people work in its tech sector. [48] Meta has roughly halved its Irish staff since the pandemic, and TikTok said it would cut about 300 Irish jobs, mostly in safety teams. [48]

    Why it matters — Ireland's AI summit this week had 36 executives and industry bosses as speakers and one union speaker, the union says. [48] Total tech jobs in Ireland still rose, from about 144,000 in 2021 to over 177,000. [48]

25. 26

    Microsoft's president hits back on green cards

    Brad Smith, Microsoft's president, said a US crackdown on green cards for skilled foreign workers could cripple US tech for more than a generation. [47] Last week the US suspended Microsoft and seven other firms from PERM, the scheme employers use to sponsor workers to stay for good. [47] Vice President JD Vance said Microsoft cut 6,000 US jobs last year while getting 6,300 H-1B visas. [47] Smith called that 'completely off base'. [47]

    Why it matters — Smith says Microsoft hired 9,200 Americans in its last financial year and pays visa holders the same as other staff. [47] The suspension still lets it hire on H-1B visas but closes the main path to staying permanently. [47]

26. 27

    A UK boss gives AI only to his best staff

    Ben Chisell, chief executive of Paysend, a London payments firm, gives general AI tools only to staff who show what they add. [45] He says AI lets the firm do about 25% more engineering work with the same team. [45] Other firms do the opposite: Lloyds Banking Group has handed out 40,000 Copilot licences. [45] A UK government study in January found 75% of firms using AI saw better productivity, but 77% saw no change in sales. [45]

    Why it matters — Unily, a British software firm, says many companies now hand out AI mainly as a way to cut costs and staff. [45]

27. 28

    OpenAI gives $5 million after a maths row

    OpenAI said on Thursday it will give $5 million to a new Simons Foundation programme to help mathematicians check results made by AI. [25] The Simons Foundation funds maths and science research. [25] Days earlier, the Association for Human Mathematics accused OpenAI of plagiarism over a result it claimed on equations about how fluids move. [25] Two mathematicians say they were close to the same result; OpenAI denies seeing their work. [25]

    Why it matters — The whole programme is worth $15 million with money and computing from the foundation, Anthropic and OpenAI. [25] OpenAI says its model has put out more than 700 maths papers. [25]

28. 29

    Microsoft brings Defender to Linux desktops

    Microsoft released a test version of Defender, its security software, for computers running Linux, the free operating system many developers use. [35] It needs Defender for Endpoint Plan 2 or a Microsoft 365 E5 or E7 licence. [35] It works the same way as Defender already does on Linux servers. [35] Ubuntu, a version of Linux, is used by about 28% of professional developers, a 2025 survey found. [35]

    Why it matters — Companies that already pay for those Microsoft plans can now watch their Linux machines with the same tool. [35] Some Linux users dislike being watched by Microsoft software at all. [35]

29. 30

    Munich and Berlin move to free software

    Munich is replacing paid software with free open-source programs, its chief digital officer Laura Dornheim said at a conference in Prague. [36] It is swapping VMware for Proxmox and Oracle databases for PostgreSQL, saving about 1.5 million euros a year. [36] Munich dropped Linux desktops in 2017 after a new council brought back Windows. [36] Berlin aims for 70% of the software on its office computers to be open source by 2032. [32]

    Why it matters — Dornheim says open-source projects in government fail for political reasons, not technical ones. [36] Schleswig-Holstein, a German state, is also moving its offices away from US cloud services. [32]

30. 31

    Capita set to miss a pension deadline

    Capita, a British outsourcing company, will miss a legal deadline to correct about 93,000 civil service pensions, the PCS union says. [31] The UK government confirmed the March 2027 deadline would be missed, the union said. [31] Capita took over the scheme for 1.7 million members in December 2025, and its website and service have struggled since. [31] The corrections follow a court ruling that pension changes treated younger staff unfairly. [31]

    Why it matters — About 73,000 of those waiting are retired or have died. [31] The UK government is paying Capita an extra 45 million pounds for this work on top of its 239 million pound contract. [31]

31. 32

    Gmail answers spoken questions for paying users

    Google released Gmail Live, which lets people search their inbox by asking questions out loud. [37] It is only for people paying for Google AI Plus, Pro or Ultra. [37] Similar tools are coming to Docs and to Keep, Google's notes app. [37] Business customers on Google Workspace will get them later. [37]

    Why it matters — Google is adding AI features to its paid plans, so people already paying get them with no new bill. [37]

32. 33

    AI chats nobody else can read

    Brave, which makes a privacy-focused web browser, now encrypts questions to its AI helper Leo before they leave the user's computer. [20] The question is only unlocked inside a sealed part of a chip, so Brave cannot read it. [20] A start-up called TrustedRouter raised $6 million to offer the same kind of protection for open models. [19] OpenAI and Anthropic do not offer such guarantees, TrustedRouter's boss says. [19]

    Why it matters — AI companies often use people's chats to train new models, and encryption is meant to stop that. [19][20] Brave says some checks still have to move into the browser itself. [20]

33. 34

    MapQuest shows where cameras watch

    MapQuest, a map app, added a layer that shows known outdoor cameras and number-plate readers. [38][39] It includes cameras made by Flock Safety, one of the biggest makers of plate readers in the US. [38] Users can add cameras they spot themselves. [38] Downloads of MapQuest jumped in September after it refused to rename Lake Ontario as Lake America. [38]

    Why it matters — Plate readers log every passing car's plate, time and place, and police can search those records. [38] A US judge ruled one police Flock search unconstitutional. [38]

34. 35

    An AI designs a homemade gun

    Cody Wilson, who made the first 3D-printed gun more than a decade ago, showed a new one he says an AI designed. [43] He trained the AI, called GatGPT, on a large online library of gun files. [43] The single-shot gun fires real bullets and has no serial number. [43] Experts criticised its odd design and made-up details. [43]

    Why it matters — Mainstream AI tools refuse to help design guns, so a custom AI could lower the skill needed to print one. [43] US states are already in court over guns made at home. [43]

35. 36

    Police use an AI-made child to catch a suspect

    Police in Livonia, a US city, arrested a school maintenance worker who arranged to meet a 15-year-old for sex. [51] The boy did not exist: an officer made his profile picture with AI. [51] The officer also used face-recognition software and Flock plate readers to find and follow the man. [51] He was arrested when he arrived and later admitted it. [51]

    Why it matters — Child-abuse images made with AI are a fast-growing problem, the US centre for missing children reports. [51] US border police also use AI-made online characters to watch protesters, documents seen by 404 Media show. [51]

36. 37

    Fake experts trouble a research firm

    GLG, a firm that connects investors with experts for paid calls, is adding checks against fake experts. [46] Its chief executive Gemma Postlethwaite says demand for anyone who knows AI has drawn people using deepfakes and chatbots. [46] GLG now checks identity on camera with questions meant to trip up AI. [46]

    Why it matters — GLG also finds experts such as child psychiatrists to help AI labs train their models on hard subjects. [46]

37. 38

    A tool checks lawyers' AI citations

    CiteProof, a new online tool, checks legal briefs for court cases that AI made up. [34] It searches public court records directly and uses no AI itself. [34] A check costs $49 for one brief, or $499 a month for 200. [34] Last year a Georgia court decision was thrown out after its order was found to cite cases that did not exist. [34]

    Why it matters — Lawyers keep using AI to draft, and judges have fined and rebuked those who filed invented cases. [34]

38. 39

    Elorian's first model looks twice

    Elorian, a start-up led by a former Google DeepMind researcher, released its first AI model, Glimpse, on Friday. [50] Most AI models look at a picture once, then reason in words; Glimpse can zoom in and look again while it thinks. [50] It was trained on a few hundred chips and costs $3 per million tokens of output. [50] Elorian says it beats OpenAI's GPT-6 Astra on reading science charts. [50]

    Why it matters — The claims are Elorian's own tests, not independent ones. [50] Its first targets are firms with lots of video, such as robot makers and labs. [50]