Weekly AI Roundup: What Happened in AI This Past Week (October 31–November 6, 2026)

Welcome back. Dhaka in the first week of November is properly cool in the mornings now. The fog sits low over the river at dawn, the winter vegetables are starting to appear in the markets, and the city’s calendar is filling up with fairs, seminars and summits — this week, for the first time, one with AI in its name.

The AI world spent the week arguing about speed. One of the biggest AI companies told its future investors, in a formal filing, that its own AI may soon be speeding up its research dramatically, and an independent group that looked inside the company broadly agreed. At the same time, American voters went to the polls with data centres on the ballot, and the next day stock markets took a sharp fall partly on fears that the building boom behind AI would slow. In between, AI found hundreds of security holes in the software that runs most of the world’s servers, journalists found a network of harmless-looking apps hiding a tool for sexual deepfakes, and two large companies started building and warning about off-switches for AI agents. At home, Dhaka hosted its first AI Business Summit, and the most useful things said there came from the panels rather than the stage. Five global, five from home, and the same two questions on each — who benefits, who pays.

5 Big International Developments

1. Anthropic told investors it may soon reach its own threshold for AI-accelerated research — and an outside assessor says its pace has roughly doubled

On Monday, November 2, Anthropic filed an amended version of its IPO registration statement with the US Securities and Exchange Commission, and, as TechCrunch reported, the update says the company’s internal estimates suggest it “may imminently meet” the definition, in its own safety policy, of “dramatically accelerated progress attributable to AI.” The filing says no current model has yet been designated as crossing that line. It also says Anthropic does not expect to be able to safely let Claude fully and autonomously design and build successive generations of frontier models without “meaningful human supervision or understanding,” and that it is considering changes to its Responsible Scaling Policy to “keep humans in control of frontier model development at Anthropic.” According to The Neuron’s reading of the filing, Claude accounted for more than 90% of code contributions at the company by September, and average daily code contributions per engineer in the third quarter were more than 45 times their 2024 level.

The same amendment added the company’s third-quarter numbers. Let’s Data Science, summarising the filing, reports revenue of roughly $17 billion for the quarter and $33.24 billion for the first nine months of the year, and about $895.4 billion in aggregate purchase commitments and lease agreements, “substantially all of which is related to compute and infrastructure.”

Three days later, on November 5, METR — an independent research group that evaluates AI systems — published its assessment of exactly this question, based on access to Anthropic’s internal data between August 12 and October 29. Its central finding: as of October 2026, Anthropic is “likely improving the capabilities of their models on well-specified tasks at around twice the rate” it managed between March 2024 and October 2025, and METR says it would not be surprised by a figure as low as 1.3 times or as high as 3 times. METR’s own rough judgement is that about 80% of the recent speed-up is due to AI itself, though it calls that highly uncertain. It found no convincing bottleneck — in computing power, data or new ideas — that would obviously slow things down over the next 12 months, and it says that “the lack of observed bottlenecks or strong diminishing returns means there is a significant risk.” It also says something important about the limits of its own report: Anthropic had the right to redact, and removed a majority of the evidence METR found most informative. Anthropic said the redactions were “aimed at removing confidential information that could accelerate other AI developers.” METR said it was “extremely unlikely Anthropic was actively trying to mislead us,” but that the public now has to rely more heavily on its judgement.

(A plain-language note: “recursive self-improvement” is the idea of AI helping to build better AI, which then helps build better AI still, so that progress speeds up on its own. For years it was a thought experiment. What is new this week is a company describing, in a legal document, how close it thinks it is to the start of that loop.)

Who benefits, and who pays? Anthropic benefits from a story investors love: a company whose product is speeding up its own work. Its rivals benefit less, which is exactly why so much was redacted. The public benefits from the fact that an outside group was let in at all, and from a filing that has to be accurate under US securities law. Who pays, if things go wrong, is everyone else — because the safety argument depends on humans still understanding what is being built, and a faster loop leaves less time to check. It is worth saying plainly that both documents were careful: no threshold crossed yet, and lots of uncertainty. It is also worth noticing that the people closest to the work are the ones raising the alarm.

The Dhaka view is that this is not a distant debate. If the companies that build these systems are speeding up, the tools that reach Bangladeshi offices, classrooms and freelancers will change faster too — and our institutions already struggle to keep up with the current pace. Last week we suggested reading what AI companies tell their own investors. This week’s filing is the clearest example yet: it is a free, public warning, written by the people who know most, and our policymakers should read it before they write a national AI policy around today’s tools.

2. AI stocks fell hard after the US midterms, as data centres became a political issue — and OpenAI promised to stop using NDAs with officials

On Thursday, November 5, the Nasdaq Composite fell 5% and entered a correction, more than 11% below its October record, as investors questioned whether the AI building boom could keep its pace, CNBC reported. The iShares Semiconductor ETF dropped 11%, its steepest one-day decline since March 2020, and the Dow lost more than 1,600 points. CNBC’s experts pointed to two things: Democratic gains in Tuesday’s midterm elections, which raised fears of tighter limits on data-centre construction, and a 4.8% ten-year US Treasury yield that makes safer investments more attractive. “Right now, investors are questioning that $1.2 trillion of capex that we forecast for next year,” said Peter Boockvar. Not everyone agreed it was the end of anything: Rob Sechan pointed to Microsoft’s 48% growth in its Azure cloud business and said, “Nobody voted to make compute less scarce.”

Data centres were literally on the ballot. According to the Washington Examiner’s election coverage, 41 data-centre measures appeared across eight states, three of which would temporarily pause development; its tally on Thursday morning showed Republicans losing control of both houses of Congress. Two days before the vote, Semafor reported that roughly 300-plus state and local moratoriums were threatening the US build-out, citing The Information, and that the Democratic candidate for Texas governor had pledged a strict moratorium in a state that, according to SemiAnalysis, could account for about 29% of the net data-centre capacity planned nationally for 2028. The Washington Examiner reports that the Democrat, Gina Hinojosa, ousted Governor Greg Abbott. Abroad, Semafor noted local opposition to Singapore-linked data-centre projects in Indonesia, and a large Australian project whose developer dropped its construction timeline after the regulator extended its review.

On the same Thursday, OpenAI published what it called a National Community Compact and a Compute Campus Policy Blueprint. According to The Neuron, the compact says: “We will not use non-disclosure agreements (NDAs) with public officials” — a practice that has kept local residents in the dark about many US data-centre deals. OpenAI also commits to paying for the electricity its campuses use and the extra infrastructure needed to serve them, and to publish data on water, electricity, construction impacts and noise before operations start and every year afterwards. The Neuron notes the limits: the compact covers future campuses OpenAI owns and operates, not partner-run facilities that supply it with computing power, and publishing a promise is not the same as keeping it.

(A plain-language note: “capex” is capital expenditure — money spent on buildings and machines rather than salaries. Most of the AI boom’s capex goes on data centres: huge buildings full of computer chips that need enormous amounts of electricity and, often, water for cooling.)

Who benefits, and who pays? Communities near planned data centres benefit from having a vote, a voice and, now, at least one company promising openness and to pay its own power bills. Investors who bought at the top pay, and so does anyone whose pension is heavily invested in a handful of AI stocks. The deeper question is who pays for the electricity. For two years, the answer in many places has been “partly, everyone else on the grid,” through higher bills. The voters’ message this week was that they had noticed.

The Dhaka view is that Bangladesh wants data centres too — the government spoke this week of becoming “a regional hub for AI and data centres” (Bangladesh story 3). The American lesson is that data centres are not just technology projects; they are power, water and land projects, and the people who live next to them eventually get a say. Bangladesh, which already rations electricity in summer, should decide now who pays for the power a data centre uses, and should write “no secret deals” into its rules before the first big contract, not after.

3. OpenAI said an upcoming AI model found 948 ways to take over Linux machines — and released a cyber model it rates “Critical”

On November 4, OpenAI published findings showing that an unnamed upcoming cyber model had reproduced 1,724 Linux kernel vulnerabilities, including 948 distinct “local privilege escalation” flaws, at a cost of about $1.2 million in API credits, according to The Neuron’s report on the announcement. The workflow, The Neuron says, ran from discovery to checking that an exploit worked, writing a candidate patch, and human review. Hundreds of reports and patches have been submitted, though that does not mean all 948 fixes have reached users. The next day, November 5, OpenAI published the safety card for GPT-6 Cyber, a specialised model for authorised security researchers. It rates the model’s cybersecurity capability “Critical” under its Preparedness Framework, and limits access to approved members of its Daybreak Red programme: individuals must be invited and pass identity checks, and organisations must use strong login security and managed devices.

The bigger picture is the volume. The Neuron counted the Windows vulnerabilities in Microsoft’s monthly security updates: 240 in August, 714 in September and 1,105 in October — about 4.6 times as many in October as in August. Microsoft had warned in July that AI-assisted discovery would increase update volumes, calling it “evidence that defenders are getting better at identifying and addressing issues.” The bottleneck is no longer finding flaws. It is reviewing, testing and installing the fixes.

(A plain-language note: the “kernel” is the core of an operating system — the part that controls everything else. “Privilege escalation” means someone who already has limited access to a machine, for example an ordinary user account, can turn it into full control. Linux runs most of the world’s servers, cloud systems and Android phones.)

Who benefits, and who pays? Every organisation that runs Linux benefits from hundreds of flaws being found by a company that reports them, rather than by criminals who sell them. OpenAI benefits from showing what its models can do, and from selling access to the defenders who need it. Who pays is the people who maintain the software — many of them volunteers or small teams — who now face a flood of reports, and every IT department that has to install a much larger stream of updates. And there is a quieter cost: a model this good at finding flaws is, by OpenAI’s own rating, dangerous in the wrong hands, which is why access is tightly limited.

The Dhaka view is simple and a little uncomfortable. Two weeks ago we wrote about the PHP emergency and asked whether Bangladeshi website owners had been told to patch. This is the same lesson at a bigger scale. Banks, telecoms, government offices and universities here run Linux servers, and many of them run old versions. When AI makes finding holes cheap, the organisations that patch slowly become the easy targets. The cheapest security investment any Bangladeshi institution can make this month is a person whose job is to install updates promptly.

4. Researchers and WIRED exposed more than 400 harmless-looking apps and sites hiding a tool for sexual deepfakes

A WIRED investigation, reported on November 2 by Digital Trends, described more than 400 connected apps and web domains — presented as pet-photo editors, language tutors and AI workout coaches — that hid a tool for making explicit deepfakes from uploaded photos, a network uncovered by researchers at AI Forensics. According to Digital Trends, a server decided whether a user saw the harmless app or the hidden layer, using details such as location, language, device and how the user found the app. The hidden layer offered dozens of templates for putting a real person’s photo into sexual scenes, including ones involving threats and violence, and some apps accepted payment through Apple Pay and Google Pay. The apps were promoted through ads on Facebook and Instagram, sometimes under fake business names. WIRED traced much of the network to Chinese developers and a Hong Kong company, KeLing Technology Limited, which scrubbed its online presence after being contacted.

Apple removed 19 more apps after learning of the scheme, including a pet-portrait app called Viefur, and said the apps had passed its review before their developers changed how they behaved; it has since updated its rules against nudifying and bait-and-switch apps. According to Digital Trends, Meta has blocked 11,600 ad accounts and 209 apps since banning KeLing from advertising in 2025. One app had more than 100,000 downloads before Google removed it. Similar ads, Digital Trends reports, kept reappearing while WIRED was finishing its investigation.

(A plain-language note: a “bait-and-switch” app looks innocent when app-store reviewers check it, then changes what it shows to certain users afterwards. That is why app-store approval alone is not a guarantee of what an app will do.)

Who benefits, and who pays? The network’s operators made money; so, through payment fees and advertising, did the platforms that carried them, until they were told. Journalists and researchers who did the digging benefit the public. The people who pay are the victims — overwhelmingly women and girls — whose photos are turned into sexual images without their consent, and who rarely know it has happened until the damage is done.

The Dhaka view connects directly to last week’s edition, where we wrote about how AI-made abuse of women is shaping the draft Cyber Protection (Amendment) Act. This investigation shows why a law aimed only at the people who share such images is not enough: the tools are distributed through mainstream app stores and paid for through mainstream payment systems, by companies outside the country. Bangladesh’s regulators should be asking the app stores and Meta, in writing, how many of these apps and ads reached users here — and the cyber-law debate should focus on that, rather than on vague offences about “rumour.”

5. Apple warned investors that AI agents “may behave in unintended or unpredictable ways” — and CrowdStrike started selling an off-switch

On Tuesday, November 3, Apple’s annual filing to US regulators warned that “artificial intelligence agents and other technologies can take actions autonomously” and “may behave in unintended or unpredictable ways,” CNBC reported. The filing also flags that its ability “to oversee and control third-party artificial intelligence agents” may be limited, and that agents could expose it to lawsuits or harm its reputation. CNBC notes that Apple has no agentic AI product of its own yet, and that in October it changed how full disk access works on Macs to reduce the risks from AI agents installed on computers.

On the same day, at its Fal.Con Europe conference, the security company CrowdStrike launched what it calls an AI Kill Switch, part of its Falcon Guardian platform, which lets security teams stop misbehaving AI agents, block them from restarting and cut off their network access. “Everyone is building AI agents, but they never had the off button, and now they do,” CEO George Kurtz told CNBC. He was also frank about the business: “Enabling agents to work safely is a massive market opportunity for CrowdStrike.”

(A plain-language note: an AI “agent” is a program that does not just answer questions but takes actions — clicking, buying, sending, changing files — on its own. The more it can do, the more it can do wrong.)

Who benefits, and who pays? Apple’s shareholders benefit from an honest list of risks, and Apple benefits legally from having warned them. CrowdStrike benefits from a new product line. Ordinary users benefit if off-switches become standard. Who pays is anyone who trusts an agent with real accounts before the controls exist — and the bill for those controls, as Kurtz makes clear, will be paid by customers.

The Dhaka view is that Bangladesh is about to meet agents through phones, banks and mobile-money apps, not through security consoles. The question for any Bangladeshi company plugging an AI agent into payments or customer accounts is the one Apple just put in writing: if the agent does something nobody asked it to, who notices, who stops it, and who pays the customer back?

5 Stories That Hit Closer to Home (Bangladesh)

1. Bangladesh held its first AI Business Summit — the President promised a human-centred AI policy and Bangla-language AI

On Thursday, November 5, President Mirza Fakhrul Islam Alamgir inaugurated the Bangladesh AI Business Summit 2026 at the Bangladesh-China Friendship Conference Centre in Dhaka, the first event of its kind in the country, jointly organised by the Ministry of Science and Technology and the Dhaka Chamber of Commerce and Industry (DCCI) under the theme “AI-fication for Tomorrow’s Economy.” According to The Business Standard, the President said AI “must be used as a strategy for increasing production, modernising industries” and that “Technology must be for people, not people for technology.” He said the government has begun work on a human-centric national AI policy and on national AI capabilities based on the Bangla language, and that AI should not aggravate the country’s unemployment problem. The Bangladesh Today reports that he described the goal as a country that is “AI-capable, innovation-driven and oriented towards human welfare,” said the policy would focus on ethics, human rights and risk-based regulation, and warned that “Vigilance is necessary at every level.” He tied it all to a target of a trillion-dollar economy by 2034 and the challenges of graduating from least-developed-country status.

Science and Technology Minister Faqir Mahbub Anam attended, as did M Ruhul Quddus Talukder Dulu, chairman of the parliamentary standing committee on the ministry. DCCI President Taskeen Ahmed, according to The Business Standard, described AI as a possible “game changer” for cottage, micro, small and medium enterprises, and said Bangladesh should be “not only an adopter of AI” but a developer of it. Professor Alamgir Hossain, CEO of the UK-based Digital Readiness and Intelligent Systems Ltd, gave the keynote. An AI Business Expo with more than 150 stalls ran alongside.

(A plain-language note: “risk-based regulation” means rules that get stricter as the possible harm gets bigger — light rules for a chatbot that recommends recipes, much tighter ones for AI that decides on loans, medical care or policing. It is the approach the European Union has taken.)

Who benefits, and who pays? Businesses benefit from the government signalling, at the highest level, that AI is a priority, and from a policy that promises to be human-centric rather than a blank cheque. Young people benefit if the promised skills training arrives. Who pays depends on the details nobody has seen yet. Last week, at a different event, officials spoke of an AI policy within months and an AI Act within a year; this week the President spoke of a policy. A policy is a statement of intent, and intent is cheap. The test is whether the draft is published for public comment, and whether it says who is responsible when AI harms someone.

The Dhaka view is that the President’s most important sentence was the least glamorous: the worry that AI could add to unemployment among educated young people. Last week we reported IT firms telling The Daily Star that AI is already shrinking their teams. A policy that starts from that fact, rather than from a slogan, would be worth having.

2. At the same summit, the sharpest voices asked what value AI actually creates here — and where the electricity will come from

At the November 5 summit, speakers argued that adopting AI is not enough and that the real test is whether it creates economic value, The Daily Star reported on November 6. “The real question is how we can create value from it,” said Zulkarin Jahangir, an assistant professor at North South University, whose work involves around 600 farmers. Fahim Ahmed, chief executive of the ride-hailing and delivery company Pathao, said Bangladesh needs reliable electricity and a clear financing plan to build the data centres AI services depend on, and called for businesses and technology companies to be involved in writing the country’s AI policy. Mahdi Amin, adviser to the Prime Minister, said AI could replace some jobs but also create new ones. Simin Rahman, chief executive of Transcom Group, said wider adoption was needed to raise productivity. And Tasfia Tasbin, co-founder and CEO of Markopolo AI, put the skills problem neatly: “Being tech-savvy and knowing how to use technology to create something quirky are not the same thing.”

Who benefits, and who pays? Companies that build real products — Pathao, Markopolo AI and others — benefit if the government listens and gives them a seat at the policy table. Farmers like the 600 in Zulkarin Jahangir’s work benefit only if AI is designed around their actual problems. Who pays is less visible: if electricity and financing are not sorted out, Bangladeshi users will keep sending their data and their money to foreign data centres, and the value will be created somewhere else.

The Dhaka view is that this panel said the quiet part out loud. Bangladesh is good at using technology and at adopting it quickly — bKash and Pathao proved that. The harder step is making something that others want to buy. The international stories this week (story 2) show that even the US is now fighting over who pays for AI’s electricity. Fahim Ahmed’s point about power is not a detail; in a country that has load-shedding in summer, it is the whole question.

3. Bangladesh discussed a Bangla-language AI model with South Korea’s LG and AI cooperation with Singapore

During an official visit to South Korea led by Finance and Planning Minister Amir Khosru Mahmud Chowdhury, the Bangladesh delegation explored “opportunities to develop a large language model (LLM) tailored to Bangladesh’s linguistic and cultural context” with LG AI Research, according to a press release published by The Business Standard on November 4. The delegation also met the South Korean AI technology company FuriosaAI, in talks that focused on “strengthening Bangladesh’s higher-value AI ecosystem” and advanced computing, and the Export-Import Bank of Korea, which expressed interest in infrastructure projects that incorporate AI. The release says Bangladesh wants to become “a regional hub for AI and data centres” and is seeking technology and investment partners; the talks also covered a $3 billion Korean development cooperation framework for 2023–2027.

Two days earlier, on November 2, Minister Faqir Mahbub Anam met Singapore’s non-resident High Commissioner to Bangladesh, Derek Loh, The Business Standard reported. Anam said the government wants to promote artificial intelligence and use it in healthcare and public administration; Loh said Singapore sees opportunities in applying AI to healthcare, scientific research, biomedical technology and public services.

(A plain-language note: a “large language model,” or LLM, is the kind of AI behind chatbots like ChatGPT and Claude. Most are trained mainly on English text, which is why they are often weaker in Bangla, and why a model “tailored” to Bangla is attractive.)

Who benefits, and who pays? Bangla speakers would benefit from an AI that understands their language well — in government services, schools and health care. LG and FuriosaAI benefit from a new market. Who pays, and how, is the question these releases do not answer. An “exploration” is not a deal, and the terms of any deal matter: who owns the model, where the data is kept, and whether Bangladeshi researchers learn to build such systems or only to use them.

The Dhaka view is that it is healthy to see Bangladesh talking to Korea and Singapore, not only to the US and China. Last week we noted that a national AI policy was discussed at an event co-hosted by Huawei. Spreading partnerships is a reasonable hedge. But a Bangla-language model is also a matter of national data, and it should be built with public oversight and with the Bangla text it is trained on treated as a national resource, not a free input.

4. University students showed 16 AI projects in Dhaka — from dengue tracking to soil health

On November 5, students from Daffodil International University (DIU) presented 16 AI projects at the Bangladesh AI Business Summit, according to a DIU press release published by The Business Standard on November 6. The projects came from three departments — six from Computer Science and Engineering, three from Software Engineering and seven from Computing and Information Systems. They include DengueCluster AI, which monitors dengue cases and flags high-risk areas; NutriSoil, which assesses soil health from real-time sensor data; Shilpo AI ERP, a predictive management system for garment factories; NeuroCluster AI, for MRI analysis; and Arbour, a voice-based communication tool for mothers with speech disabilities. Daffodil Group also showed five of its own software products. The release says the President visited the DIU pavilion; DIU was a sponsor of the summit.

Who benefits, and who pays? The students benefit from exposure to business leaders and investors, and the university from publicity. The public benefits if any of these projects becomes a working service — a dengue warning system, for example, could save lives in a country where dengue season fills hospitals every year. The honest caveat is that this is a press release from a sponsor, describing student projects; we do not know which of them work outside a demonstration, and they deserve to be judged on that later.

The Dhaka view is that this is the most encouraging story of the week, and also the most fragile. The ideas are exactly right: they start from Bangladeshi problems — dengue, soil, garments, disability — rather than from copying foreign apps. What usually happens next is that the students graduate, the projects stop, and the best developers leave for jobs abroad. If the national AI policy wants a quick win, it could fund the best of these student projects for one more year, with real data and real users.

5. A Dhaka op-ed argued Bangladesh cannot afford to pause AI — but must protect its data and find the power

On November 2, The Daily Star published an opinion piece by Syed Almas Kabir, chairman of the Bangladesh ICT & Innovation Network (BIIN), arguing that Bangladesh is at a crossroads between data sovereignty, energy shortages and the race for AI. He argues that the country must keep developing AI rather than pause, because delay risks lost competitiveness and jobs, but that it needs stronger data-protection laws covering consent, storage, algorithmic transparency and penalties, citing Europe’s GDPR and Malaysia’s Personal Data Protection Act as references. He proposes starting with smaller, energy-efficient data centres powered by solar and wind, and suggests seawater cooling in coastal areas to reduce pressure on freshwater. He notes that internet penetration has crossed 40%, according to the Bangladesh Bureau of Statistics, and that mobile broadband users exceed 120 million. “In the end, the debate is not about whether Bangladesh should develop AI — it must,” he writes.

(A plain-language note: “data sovereignty” means that data about a country’s citizens is stored and governed under that country’s laws, rather than sitting on servers abroad under someone else’s rules.)

Who benefits, and who pays? Citizens benefit from a data-protection law that gives them rights over their own information — something Bangladesh still lacks in a strong, enforced form. The local data-centre and IT industry benefits from rules that favour keeping data at home. Who pays is the question of cost: sovereign infrastructure is expensive, and if it is slower or pricier than foreign cloud services, businesses and, eventually, consumers will pay the difference.

The Dhaka view is that this op-ed and this week’s international news belong side by side. Americans voted on data centres this week because they discovered too late who was paying for the power. Bangladesh can have that argument before the buildings go up. A small, solar-powered start, as Kabir suggests, is less glamorous than a “regional hub,” but it is the version the grid can actually support.

Wrapping up

Ten stories in seven days, and the international five were about who controls the pace. Anthropic told its future investors that its AI may soon be dramatically speeding up its own research, and METR, the outside group it let in, estimated that its progress has roughly doubled in pace — with most of the strongest evidence redacted. US voters put data centres on the ballot, and two days later the Nasdaq fell 5% as investors wondered whether the building boom would slow; OpenAI promised to stop using NDAs with public officials and to pay for its own power. OpenAI’s models found 948 ways to take over Linux machines, and the company released a cyber model it rates “Critical” to vetted defenders only. WIRED and AI Forensics researchers exposed more than 400 apps and sites hiding a deepfake tool behind pet photos and workout coaches. And Apple and CrowdStrike, in different ways, admitted that AI agents need an off button.

At home, the week belonged to the first Bangladesh AI Business Summit. The President promised a human-centred AI policy and Bangla-language AI, and warned about unemployment. The panels asked the harder questions about value and electricity. The government explored a Bangla language model with South Korea’s LG and AI cooperation with Singapore. University students showed 16 projects built around Bangladeshi problems. And an op-ed argued that Bangladesh must keep building AI while protecting its data and finding the power.

The thread across both halves is that speed has a bill attached. Abroad, the bill is arriving in the form of voters, regulators and stock prices asking who pays for the electricity, the security fixes and the harm. At home, the summit was full of ambition, and ambition is right. But the useful habit from this week’s international news is to ask, before every big promise, the boring questions first: where does the power come from, where does the data go, and who is accountable when it goes wrong?

If something here caught your interest, share it with someone who does not follow tech. The conversation only spreads if we keep starting it.

See you next week. Stay well.