Weekly AI Roundup: What Happened in AI This Past Week (May 23–29, 2026)

Come in, sit down, kick the slippers off. The kettle is on and the fan is turning slowly against another sticky pre-monsoon evening in Dhaka. Another seven days of AI news has gone by, and this week the headlines were quieter than last week’s billion-dollar compute contracts — but in a way that made them harder to shake. A machine disproved a mathematical conjecture that the best human minds had circled for eighty years. A pope wrote his first encyclical and chose artificial intelligence as the subject. The world’s largest accountancy network handed Claude to a quarter of a million staff. And in among it all, the most honest stories were not about how clever the models had become, but about what they cost — in electricity, in water, in human dignity, in who gets left out.

Like the last few weeks I have split the roundup into two parts. Five big global stories first — the ones that move money, shape policy, and reshape who has leverage over whom. Then five Bangladesh stories that show how the same wave is breaking on our own shore, sometimes with quiet ambition and sometimes with hard truths. I will keep asking the same two questions on every story: who actually benefits, and who actually pays. And then, where it makes sense: what does this look like from Dhaka, Chattogram, or Sylhet rather than from Mountain View or the Vatican. Pour yourself something warm and let us begin.

5 Big International Developments

1. An OpenAI model disproved an 80-year-old maths conjecture on its own

This is the story that made working mathematicians sit up this week. An OpenAI research model independently disproved the Erdős unit-distance conjecture — a problem in discrete geometry that has sat open since the 1940s, posed by Paul Erdős himself and resistant to every human attack since. The detail that matters is the word independently. This was not a researcher using AI as a fancy calculator or a literature-search tool. The model constructed the counterexample, reasoned through the geometry, and produced a result that closes a question humans could not close. It is, as far as anyone can tell, the first time an AI system has functioned as an autonomous mathematical discoverer at the frontier of a genuinely hard, long-standing open problem rather than as an assistant to one.

Step back from the technical achievement and the significance is about a line being crossed. For the entire history of the field, the implicit deal was that AI helped humans think — it autocompleted the code, summarised the paper, suggested the next move. The human stayed in the chair where the insight happened. Disproving an Erdős conjecture is insight. It is the kind of result that earns a human a permanent place in the mathematical literature. To have it arrive from a model, unprompted by a specific human chain of reasoning, moves AI from the passenger seat to the driver’s seat in at least one rarefied domain. The mathematicians I read this week were split between awe and a quiet unease that is hard to name.

Who benefits, and who pays? The obvious winner is OpenAI, which gets a proof point money cannot buy heading into its IPO — “our model does original mathematics” is a sentence that reprices a company. The deeper winners are the sciences that depend on hard combinatorial and geometric reasoning: cryptography, materials design, logistics, parts of physics. If a model can close an eighty-year-old conjecture, it can attack the smaller open problems that clog up real research programmes by the thousand. The quieter losers are the mid-career researchers whose value rested on being the cleverest person in the room on a narrow problem. That value just got more contestable. And there is a subtler cost: when the insight comes from a system whose reasoning is not fully legible even to its makers, the human community has to decide how much it trusts a proof it did not build. Mathematics has always been the one field where you could check every step yourself. That contract is now under quiet renegotiation.

The Dhaka view is unexpectedly hopeful here, and worth dwelling on. Frontier mathematics has historically been one of the most capital-light sciences — you need a brain, a blackboard, and access to the literature, not a billion-dollar data centre. Bangladesh has produced genuinely world-class mathematical talent on exactly those terms, from the maths-olympiad pipeline upward. If the frontier of mathematical discovery is shifting toward posing the right problems to capable models and verifying what they return, then a sharp 22-year-old at Dhaka University or BUET with API access is not categorically locked out the way they are from training a frontier model. The constraint flips from compute to taste — knowing which problem is worth asking. That is a kind of constraint Bangladesh can actually compete on, if the universities treat it as a real opportunity rather than a curiosity.

Looking out to 2028–2030, the structural question this opens is governance of machine-generated knowledge. When a model produces a result no human fully derived, who is the author, who is accountable if it is wrong, and how does a discipline maintain its standards of proof? These are not abstract questions for long. The next government, the next syllabus committee, and the next research-funding body in Bangladesh will all have to decide whether AI-assisted and AI-led discovery counts, how it is credited, and how students are taught to verify rather than merely trust. The countries that build those norms early will set them for everyone else.

2. Pope Leo XIV issued the first papal encyclical on AI: ‘Magnifica Humanitas’

In a move with almost no precedent, Pope Leo XIV promulgated the Vatican’s first encyclical devoted to artificial intelligence, titled Magnifica Humanitas — roughly, “the magnificence of humanity.” An encyclical is the highest and most considered form of papal teaching, the kind of document that is meant to last for decades and bind the conscience of more than a billion Catholics worldwide. That the Vatican chose AI as the subject of one tells you how seriously the institution now takes the technology as a moral and not merely technical matter. The document’s core argument is stated plainly: technology is never neutral. It rejects, in particular, any use of AI that seeks to “perfect or surpass” human dignity — the transhumanist and superintelligence framings that treat the human as a thing to be optimised or exceeded — and calls instead for guardrails centred on the common good.

It is easy, in a tech roundup, to wave past a religious document as background noise. That would be a mistake this week. The Vatican is one of the very few institutions on earth that operates on a timescale measured in centuries, has no equity in any AI lab, faces no quarterly earnings call, and speaks to a constituency larger than any single nation. When that institution says “technology is never neutral,” it is contradicting — deliberately and at the highest level — the foundational marketing claim of Silicon Valley, which is that the tools are neutral and only the use is good or bad. Magnifica Humanitas says no: the tool carries the values of its makers in its grain, and a technology built to surpass the human is built on a moral error before anyone uses it.

Who benefits, and who pays? The beneficiaries are, in the encyclical’s framing, ordinary people — the workers, the poor, the elderly, the disabled, the people most likely to be optimised away by a system designed for efficiency rather than dignity. The document hands them a moral vocabulary and an institutional defender. The ones who “pay,” in the sense of being put on the back foot, are the labs and ideologues building toward superintelligence as an explicit goal; they now have to argue against a billion-person moral framework, not just against safety researchers. There is also a quieter beneficiary: every secular policymaker who wanted to regulate AI on human-dignity grounds but lacked the cultural authority to do it. The pope just lent them a great deal of that authority.

The Dhaka view is layered, because Bangladesh is a Muslim-majority country and this is a Catholic document — but the framing travels further than the denomination. The encyclical’s central claim, that a technology must serve human dignity and the common good rather than surpass the human, maps very cleanly onto the language already appearing in Bangladesh’s own National AI Policy draft, which leads with citizen dignity and explicitly prohibits social scoring. The deeper resonance is that the global south, including Bangladesh, has a long and bitter experience of being told that a powerful new technology is “neutral” while it reshaped local economies on terms set elsewhere. Magnifica Humanitas is, among other things, a statement that the people on the receiving end of a technology have a moral standing to judge it. That is a useful thing for a country like Bangladesh to be able to point to.

Out to 2028–2030, the structural significance is that the moral and religious institutions are now entering the AI-governance conversation as serious players, alongside governments and corporations. For most of the modern era, the major faiths sat out the technology debates or arrived late. They are not arriving late this time. In a country where religion is woven through public life, that matters: the Bangladeshi conversation about AI ethics will not be conducted only in the secular language of Stanford’s AI Index. It will also be conducted in the language of dignity, of the common good, and of what a technology is for. The encyclical makes that conversation more legitimate, not less.

3. KPMG put Claude in front of its entire 276,000-person workforce

The single largest enterprise-adoption story of the week was the global alliance between Anthropic and KPMG, under which the accountancy-and-advisory giant has standardised on Claude across its 276,000-strong global workforce. This is not a pilot, not a department-level trial, not a “we are exploring AI” press release. It is one of the Big Four professional-services firms making Claude part of the core operating layer for every auditor, tax adviser, and consultant on its payroll, in every country it works in. The significance is in the totality. When a firm that sells judgement — audit opinions, tax advice, transaction diligence — wires a single AI model into the daily work of a quarter of a million professionals, it is making a bet that AI changes the unit economics of judgement itself.

Pause on what KPMG actually does, because it sharpens the story. Professional-services firms have spent a century selling the leveraged pyramid: a small number of senior partners whose expertise is monetised by a large base of junior staff doing the document review, the reconciliation, the first-draft analysis. That pyramid is precisely the shape that a capable AI model attacks most efficiently, because the base of the pyramid is the most automatable layer. KPMG standardising on Claude is therefore not just a productivity story; it is the leading edge of a structural change in how one of the world’s largest white-collar employers is shaped.

Who benefits, and who pays? KPMG’s partners benefit first and most — every hour of junior labour that Claude absorbs is margin that flows up the pyramid. Anthropic benefits enormously: a 276,000-seat reference deployment at a marquee firm is the kind of credential that wins the next ten enterprise deals, and it deepens Anthropic’s pivot from a research lab into the operative layer of multinational business. The clients of KPMG may benefit too, if the work gets faster and cheaper. The ones who pay are the people at the base of the pyramid — the graduate-entry auditors and analysts whose roles were the training ground for the next generation of partners. If the bottom of the pyramid thins out, the path to the top narrows with it, and the profession has to invent a new way of growing senior people. Nobody has solved that yet.

The Dhaka view here is direct and it is double-edged. KPMG, like the other global firms, runs significant delivery and shared-services operations across South Asia, and Bangladesh has spent years trying to capture more of that back-office and assurance-support work on the strength of its English-capable accounting graduates. The country produces tens of thousands of commerce and accountancy graduates a year, and the chartered-accountancy pipeline is real. The uncomfortable arithmetic is that the exact tasks Bangladesh hoped to win — the reconciliation, the document review, the first-pass analysis — are the tasks Claude now does inside KPMG. The opportunity flips: the Bangladeshi accountant who learns to supervise and direct Claude on complex assurance work is more valuable than ever; the one who only does the manual review is competing with a model that does not sleep. The local professional bodies — ICAB and the rest — should be treating AI fluency as a core competency in the syllabus right now, not in three years.

Looking out to 2028–2030, the structural pattern is that AI adoption at the frontier of enterprise is no longer happening at the team level or the tool level. It is happening at the firm level, all at once, across hundreds of thousands of people. That changes the speed at which white-collar work gets reshaped — not gradually, department by department, but in a single procurement decision. For a country whose services-export strategy rests on doing professional work for Western firms more cheaply, the message is that the firms are changing what “the work” even is. Bangladesh has to chase the new shape of it, not the old one.

4. The AI energy-and-water crunch stopped being a forecast and started being a fight

The least glamorous but arguably most consequential global story this week was infrastructural: a fresh wave of reports putting hard numbers on what AI’s compute build-out is doing to the electrical grid and the water supply. The figure that travelled fastest was the projection that data centres could consume around 12 per cent of the US electrical grid by 2028 — a share that would have sounded absurd two years ago and is now a planning assumption. Alongside the power numbers came the water numbers: the enormous quantities used to cool the hyperscale data centres that train and serve frontier models. The framing in the reporting had shifted notably. This was no longer “AI uses a lot of energy” as a curiosity. It was “the socio-economic reality of AI is now colliding with planetary limits,” with rising and increasingly organised calls for regulation of the utility infrastructure itself.

The thing to understand here is that the AI industry has, over the past year, run into a constraint that money alone cannot dissolve. You can raise a hundred billion dollars in a quarter. You cannot conjure a hundred billion dollars of grid capacity, transmission lines, transformers, and cooling water in the same quarter — those take years and they compete directly with the power and water that households, farms, and existing industry already use. When a data centre and a town are bidding for the same megawatts and the same aquifer, that is no longer an engineering problem. It is a political one, and politics is where this story now lives.

Who benefits, and who pays? The beneficiaries are the AI labs and hyperscalers who got their data centres sited, powered, and watered early, locking in capacity ahead of the squeeze — and the utilities and independent power producers who get to sell electricity at premium, AI-inflated rates. The ones who pay are starkly identifiable: ordinary ratepayers, whose electricity bills rise as data-centre demand outbids them for generation; communities whose water tables drop to cool somebody else’s GPUs; and the climate, if the new load is met by firing up fossil generation that was meant to be retired. This is the clearest “who benefits / who pays” split in the entire roundup, because the benefit accrues to a global industry and the cost lands on a specific town’s power bill and a specific region’s water.

The Dhaka view is uncomfortable and important. Bangladesh’s power grid is already the binding constraint on almost every industrial ambition the country has — load-shedding is a lived reality, not a historical memory, and the energy mix is stretched. The honest implication is that the data-centre dreams in this month’s ministerial speeches collide head-on with this week’s global lesson: AI infrastructure is, first and foremost, an energy-and-water project, and Bangladesh’s binding constraint is exactly the resource the rich countries are now fighting over. There is a version of this that is bad news — the country cannot build sovereign compute it cannot power. But there is a more useful reading: Bangladesh has the chance to plan its compute build-out around its energy reality from day one, siting modest, efficient, renewable-fed facilities rather than chasing hyperscale vanity projects it cannot keep cool or keep lit. The countries arriving late to data centres can learn from the water fights happening now in Arizona and Virginia, instead of repeating them.

Out to 2028–2030, the structural shift is that utility policy is becoming AI policy. The most important AI decisions in many countries over the next five years will be made not by technology ministries but by energy regulators and water boards. For Bangladesh, the practical ask is that the AI-policy conversation and the power-and-water-planning conversation — which currently happen in completely separate rooms — start happening in the same room. Nobody in the country’s energy policy has yet priced AI as a serious future load. They will have to, and the sooner the better, because the global queue for both GPUs and the power to run them only gets longer.

5. Microsoft moved its Copilot Studio computer-using agents to General Availability

The week’s clearest sign that the agentic era is now shipping, not just demoing, came from Microsoft: its Copilot Studio “computer-using agents” reached General Availability. GA is the unglamorous word that matters — it means these are no longer previews or limited betas, but supported, billed, production-grade products that any enterprise customer can switch on. The capability is significant. These agents can navigate ordinary business software and interact with graphical user interfaces in real time — clicking through the same screens, forms, and dashboards that a human administrator uses, rather than relying on clean APIs built for machines. In other words, they can operate the software that businesses already run, the way a person operates it.

That GUI detail is the whole story. The enterprise world is full of software that has no clean programmatic interface — legacy ERP systems, government portals, supplier dashboards, the long tail of line-of-business tools that a company depends on but that nobody will ever rebuild for an API. Until now, automating work inside those systems meant either brittle screen-scraping scripts or a human doing the clicking. An agent that can reliably see a screen, understand it, and act on it dissolves that barrier. It is the difference between automating the 20 per cent of business processes that happen to have nice APIs and reaching for the other 80 per cent that do not.

Who benefits, and who pays? The beneficiaries are enterprises drowning in repetitive screen-based administrative work — the data entry, the cross-system reconciliation, the form-filling, the status-checking that consumes a vast fraction of office hours globally. Microsoft benefits by deepening the moat around its already dominant enterprise stack: if your agents live inside Copilot Studio and operate your Microsoft 365 and Dynamics estate, you are not leaving. The ones who pay are the workers whose jobs are that screen-based administrative work — the back-office clerks, the operations coordinators, the entry-level administrators whose daily task list is exactly what a computer-using agent does best. This is the most direct labour-substitution product to reach GA so far, because it targets not a specialised skill but the generic act of operating a computer.

The Dhaka view is sharp and it cuts to the heart of the country’s services-export model. A very large share of Bangladesh’s business-process-outsourcing industry — the BPO work that the government keeps citing as a growth pillar — is, at bottom, people operating screens on behalf of foreign clients: data entry, form processing, content moderation, back-office administration. A GA-grade computer-using agent is, with uncomfortable precision, aimed at that work. The optimistic reading is that Bangladeshi BPO firms can deploy these agents themselves — selling clients “we will run your screen-based processes with a supervised agent fleet, cheaper and faster” — and move up from doing the clicking to orchestrating it. The pessimistic reading is that the Western client buys Copilot Studio directly and skips the Dhaka BPO firm entirely. Which of those two futures arrives depends on whether the local industry pivots from labour arbitrage to agent orchestration in the next 18 months. The window is open now. It does not stay open.

Looking out to 2028–2030, the structural point is that agentic AI has crossed from research into supported product across the whole industry at once — Google’s Antigravity, Anthropic’s Claude agents, and now Microsoft’s GA computer-using agents are all converging on the same thesis in the same quarter. The competition is no longer about whose model scores highest on a benchmark; it is about whose agents do real, messy, screen-based business work most reliably. For an economy built on doing exactly that work for others, this is the single most important industry shift to track. Bangladesh either builds the agent-orchestration layer on top of its existing BPO base, or watches the base get automated out from under it.

5 Stories That Hit Closer to Home (Bangladesh)

1. The National AI Policy 2026–2030 opened for public review — and it leads with dignity

The most consequential domestic development this week was governance, not gadgetry. The draft of Bangladesh’s National AI Policy 2026–2030 (Draft v2.0) is now open for public review, and the text is genuinely worth reading rather than skimming. It proposes a strict, risk-based regulatory framework — the same architectural choice the EU made with its AI Act — under which AI systems are governed in proportion to the harm they could cause. The two features that stand out are mandatory algorithmic impact assessments for higher-risk systems, and explicit prohibitions on mass surveillance and social-scoring systems. That second point is not a throwaway line. A country writing into its foundational AI document that it will not build a social-credit system is making a deliberate statement about the kind of state it intends to be in the AI era.

Step back and consider how unusual this framing is for the region. The default trajectory for a developing-country government acquiring powerful new surveillance-capable technology is to acquire it first and worry about limits later, if at all. Bangladesh’s draft does the opposite: it writes the limits into the founding document, and it centres the whole framework on citizen dignity. Whether that survives contact with implementation is the open question — policy text is cheap and enforcement is expensive — but the choice of values expressed in the draft is a meaningful signal, and it lands in the same week that Pope Leo XIV’s encyclical made dignity the centre of the global moral conversation about AI. That is not a coincidence so much as a convergence: serious people thinking about AI from very different traditions are arriving at the same place.

Who benefits, and who pays? If the policy is implemented with teeth, the beneficiaries are ordinary citizens — protected, at least on paper, from the most invasive uses of AI by the state and by large companies. Civil-society groups and rights organisations benefit from having a legal framework to point to. The ones who “pay,” in the sense of facing new constraints, are agencies and firms that might have wanted to deploy facial recognition at scale, build scoring systems, or run high-risk AI without assessment. And there is a real cost to watch: a strict risk-based framework with mandatory impact assessments imposes a compliance burden that large firms can absorb and small Bangladeshi startups may struggle with. If the assessment regime is designed clumsily, it could entrench incumbents and crush the very local AI industry the government says it wants. Good implementation means making compliance proportionate and cheap for small players — that detail will decide whether the policy protects citizens without smothering builders.

The Dhaka view is that this is the document where the country’s stated values meet its capacity to enforce them. A prohibition on mass surveillance is only as real as the institution that can detect and punish a violation. The honest assessment in May 2026 is that Bangladesh has the policy text and the right intentions, and does not yet have the regulatory muscle — the trained assessors, the oversight body, the technical capacity to audit an algorithm — to make the prohibitions bite. The public-review period is the moment for civil society, academics, and the technical community to push hard on exactly that gap: not whether the values are right, but whether the enforcement machinery will exist. Everyone who cares about this should actually file a comment. These windows close.

Out to 2028–2030, the structural significance is that Bangladesh is choosing, early and on the record, to be a rights-centred AI jurisdiction rather than a surveillance-first one. That is a brand, and it is a valuable one — it is the thing that lets a Bangladeshi firm credibly sell AI services to European clients who care about governance, and it is the thing that distinguishes the country from regional neighbours moving the other way. The policy is the floor. The next five years are about building the house on it.

2. Activate Rights documented how generative AI was weaponised in February’s election

The hardest story to read this week was the Activate Rights report documenting the extensive weaponisation of generative AI during Bangladesh’s February 12 election. The findings are not abstract. The report catalogues the use of deepfakes and synthetic media to manipulate the campaign — and, most disturbingly, the targeted use of deepfakes against women, weaponising fabricated intimate or compromising imagery as a tool of political intimidation and character assassination. This is the dark mirror of every hopeful story in this roundup. The same generative technology that lets a Dhaka startup build a Bangla-language tutor lets a bad actor manufacture a convincing video of a female candidate or activist saying or doing something she never did, and push it across the country in hours.

There is a reason to dwell on the gendered dimension specifically, because it is where the harm concentrates. Across the world, the first and most damaging mass-market use of deepfake technology has been the creation of non-consensual sexual imagery, and it overwhelmingly targets women. In a political context, and in a society where a woman’s reputation can be weaponised against her with brutal efficiency, a synthetic compromising video is not just disinformation — it is a weapon designed to drive women out of public life altogether. The Activate Rights report names this for what it is, and that naming is the first step toward any defence against it.

Who benefits, and who pays? The beneficiaries of election deepfakes are the actors willing to use them — the campaigns, operatives, and foreign or domestic interests who gain from a manipulated information environment and from intimidating opponents out of the race. The ones who pay are, first, the women directly targeted, who bear a personal and often irreversible cost; and second, the entire electorate, whose ability to make an informed choice is corroded by a media environment in which any video might be fake. There is a deeper, structural victim too: trust itself. Once voters learn that any damaging clip might be synthetic, bad actors gain a second weapon — the “liar’s dividend,” where a politician caught on a genuine recording can simply claim it was a deepfake. A society that cannot tell real from fake loses the ability to hold anyone accountable for anything.

The Dhaka view is that this report should be read alongside the National AI Policy draft as two halves of the same conversation. The policy talks about prohibiting harmful AI; the Activate Rights report documents the harm that already happened, in the country’s most recent election. The most useful thing the report does is convert an abstract global worry — “deepfakes might affect elections” — into concrete, documented, Bangladeshi evidence that they already have. That evidence is exactly what the public-review process for the AI policy needs, because it answers the question every sceptical official asks: “show me the actual harm here, not in America.” Here it is. The practical asks that follow are unglamorous but real: rapid-response verification capacity for journalists and the Election Commission, legal remedies specifically for image-based abuse, and public-awareness campaigns so voters approach synthetic media with informed scepticism rather than either credulity or universal doubt.

Looking out to 2028–2030, the structural reality is that every future Bangladeshi election will be contested partly in synthetic media, and the defences have to be built before the next one, not after. The technology to generate convincing fakes is now cheap, fast, and widely available; the technology and institutions to detect, debunk, and remedy them are not. Closing that gap — and protecting women in public life specifically — is one of the most urgent items on the country’s AI-governance agenda. It is also a test of whether the dignity-centred language in the policy draft means anything when the dignity under attack belongs to a woman running for office.

3. AIUB switched on a sovereign supercomputing lab — Bangladesh’s own compute, at last

After weeks of essays lamenting that Bangladesh has no sovereign AI compute, this week brought the first real answer. The American International University-Bangladesh (AIUB) inaugurated a sovereign supercomputing facility built on NVIDIA DGX systems — domestically operated, high-performance AI compute on Bangladeshi soil. The stated mission is pointed and local: to build locally relevant, culturally sensitive AI for the country’s real priorities — health, agriculture, and textiles — rather than to chase generic frontier-model bragging rights. The framing matters as much as the hardware. AIUB is not positioning this as “Bangladesh’s answer to OpenAI.” It is positioning it as the capacity for Bangladesh to own its own AI research agenda — to fine-tune and train models on Bangla data, for Bangla problems, without exporting the data and the dependence to a foreign cloud.

If you read last week’s roundup, you will remember the essay arguing that Bangladesh did not appear in Stanford’s 2026 AI Index even once, and that South Asia — two billion people — had reached a grand total of two state-backed supercomputing clusters while China built eighty-five. The AIUB lab does not close that gap. One university DGX cluster is not a national compute strategy, and it would be dishonest to pretend otherwise. But it changes the conversation from “we have nothing” to “we have a first node,” and first nodes matter more than their raw size, because they are where the operational know-how lives — the people who learn to run the cooling, schedule the jobs, keep the cluster fed and healthy. That expertise is the genuinely scarce thing, and you can only build it by operating real hardware.

Who benefits, and who pays? The immediate beneficiaries are AIUB’s researchers and students, who get hands-on access to compute that previously meant renting from AWS in dollars or borrowing a single GPU from a lab. The broader beneficiaries are the three sectors named — health, agriculture, textiles — which are precisely the domains where Bangla-specific, locally-grounded models could create real value that no Silicon Valley model will ever bother to build, because the market is invisible from Mountain View. The country benefits from the sovereignty angle: data on Bangladeshi patients, farms, and factories can be worked on without leaving the jurisdiction. As for who pays — the honest answer is that a single facility serves a tiny fraction of the national research community, and there is a real risk that “Bangladesh has a supercomputer now” becomes a complacent headline that takes pressure off the government to fund compute at national scale. One private university’s cluster cannot substitute for the National AI Compute Strategy the policy draft promises.

The Dhaka view is genuine, measured optimism. This is the kind of story that does not move markets but quietly determines whether the next generation of Bangladeshi AI researchers trains at home or abroad. A student who can run real experiments on a DGX cluster in Dhaka is a student who has a reason to stay — and the talent-retention problem this column keeps returning to is, at root, a problem of giving talented people something worth staying for. The smart move now is for AIUB to open the facility to researchers beyond its own walls, to partner with the public universities that lack the capital to build their own, and to treat the lab as shared national infrastructure rather than an institutional trophy. The know-how compounds fastest when it is shared.

Out to 2028–2030, the structural significance is that sovereign compute has gone from aspiration to fact in Bangladesh, even if at small scale. The question for the next five years is whether this first node becomes the seed of a real national network — public universities, a government-shared facility, private clusters like AIUB’s, all interoperating — or whether it stays a one-off. The policy draft’s compute commitments and this week’s actual hardware now need to find each other. The clock that last week’s essay said you could hear from Dhanmondi is still running; this week somebody, at least, started building.

4. 650,000 freelancers are pivoting from solo gigs to AI-specialised agencies

A quieter but structurally fascinating story this week: reports that over 650,000 Bangladeshi freelancers are shifting away from individual, task-based gig work and toward forming AI-specialised agencies. This is a big number and a meaningful change in shape, not just scale. The Bangladeshi freelancing economy was built on individuals — a person with a laptop in Jashore or Rangpur bidding for discrete tasks on Upwork and Fiverr, paid per job, competing globally on price. The pivot described here is toward collective, capability-building structures: freelancers banding together into agencies that offer AI-specialised services as a team, pursuing higher-value, more resilient work than any of them could win alone.

The reason this matters is that the individual gig model was always fragile, and AI made it more so. A lone freelancer selling “I will write your copy” or “I will do your data entry” was already competing on thin margins, and is now competing directly with the AI tools that do those exact tasks. The agency pivot is, in effect, the market’s own answer to that squeeze: you cannot out-cheap a model on a commodity task, but a team that knows how to deploy AI tools to deliver complex, supervised, multi-step outcomes can sell something a model alone cannot — reliability, accountability, relationship, and judgement. The freelancers are, collectively and without anyone directing them, climbing the same value-chain ladder this column has argued the whole country needs to climb.

Who benefits, and who pays? The freelancers who make the transition successfully benefit substantially — agency work is higher-value, more stable, and less exposed to the race-to-the-bottom of per-task bidding. Clients benefit from being able to buy a capable team rather than assemble one freelancer at a time. The country benefits from a freelancing sector that is more resilient to AI disruption and earns more foreign exchange per worker. The ones who pay are the freelancers who cannot make the jump — the older, less-connected, or single-skill workers for whom forming or joining an agency is not realistic, and who remain exposed to exactly the commodity-task automation that is eating their niche. A shift like this almost always leaves a tail behind, and that tail is real people with real income at stake.

The Dhaka view is that this is one of the more genuinely encouraging Bangladesh stories in months, precisely because it is bottom-up. Nobody legislated this. No ministry announced it. Hundreds of thousands of workers, reading the market correctly, are reorganising themselves into more resilient structures — and that self-organising instinct is the same one that built the garments industry and the mobile-financial-services boom. It is the demand-side and supply-side igniting together. The constructive role for policy here is light-touch and supportive: make it easy and cheap to register a small agency, provide AI-skilling at scale (not LinkedIn certificates but real, project-based training), and help these new agencies access the legal and payment infrastructure to contract directly with Western clients. The worst thing the state could do is over-formalise it and tax the life out of it before it matures.

Looking out to 2028–2030, the structural pattern is that Bangladesh’s enormous freelancing workforce is the country’s most adaptable AI asset — far more nimble than its universities or its bureaucracy. If even a meaningful fraction of these 650,000 successfully become AI-deployment agencies rather than gig-task individuals, the country’s services-export base shifts from “cheap labour” to “cheap capability,” which is exactly the pivot every other story in this roundup says is necessary. The risk is the tail left behind; the opportunity is a workforce that reinvented itself faster than the institutions meant to lead it. Watch whether the agencies formed this year are still standing — and exporting — next spring.

5. The BBS digital-divide report: near-universal phones, a deep and unequal gap underneath

The week’s most sobering domestic data came from the Bangladesh Bureau of Statistics (BBS), whose latest figures show a familiar but worsening pattern: mobile access is now nearly universal, while a deep digital divide persists along lines of gender and region. The headline number — almost everyone has a phone — masks the more important truth underneath it. Having a mobile phone is not the same as having meaningful digital access, and the gap between having a device and being able to use the digital economy falls hardest on women and on rural communities. The report frames the stakes precisely for this moment: as AI integration accelerates, that existing divide threatens to marginalise rural communities further, layering a new and steeper inequality on top of the old one.

This is the structural counterweight to every optimistic story above, and it deserves to be read as such. The AI Olympiad finds brilliant teenagers — mostly in elite urban English-medium schools. The freelancer-agency pivot lifts the connected and the skilled. The AIUB supercomputer serves a Dhaka university. Every one of those is real and good, and every one of those, left to itself, widens the divide the BBS just documented, because the benefits of AI flow first and fastest to the people who already have connectivity, skills, English, and capital. A rural woman with a basic phone, intermittent data, limited literacy in the digital interfaces, and no access to AI tools is not lifted by any of this week’s hopeful stories. She is left further behind by them.

Who benefits, and who pays? The brutal honesty of a digital divide is that the benefit and the cost fall on different people. The beneficiaries of AI-era growth are the digitally included — urban, often male, English-capable, connected. The ones who pay are the digitally excluded, who not only miss the upside but face a relative decline as the economy reorganises around tools they cannot reach. The gendered dimension compounds it: a divide that runs along gender lines means AI-era prosperity is, by default, distributed unequally between men and women, entrenching an inequality the country has spent decades trying to narrow. And the regional dimension means the gap between Dhaka and the districts — already wide — gets wider unless something deliberate counteracts it.

The Dhaka view, fittingly, is the view from outside Dhaka, and that is the whole point of the report. The country’s AI conversation — including, honestly, this column — is overwhelmingly a Dhaka conversation, conducted among the connected about the connected. The BBS data is a reminder that the majority of Bangladeshis live on the other side of the divide it measures. A National AI Policy centred on citizen dignity has to mean the dignity of the rural woman with a basic handset as much as the dignity of the Gulshan professional with Claude on her laptop. That means concrete, unglamorous things: Bangla-language and voice-first interfaces so literacy is not a barrier, rural connectivity investment, female-targeted digital-skilling, and AI services designed for the actually-existing phone in a Kurigram farmer’s hand rather than the flagship device in a Dhaka boardroom. Equitable AI integration is not a slogan in the BBS report; it is the difference between AI narrowing Bangladesh’s inequalities or carving them deeper.

Out to 2028–2030, the structural question this report poses is the most important one in the entire roundup: for whom is Bangladesh building its AI future? If the answer, in practice, is “the already-connected,” then the demographic dividend everyone forecasts accrues to a thin urban slice and the country’s inequality hardens into something AI-shaped and permanent. If the answer is genuinely “all citizens,” then closing the gender and regional divide has to be treated not as a charity add-on but as the central design constraint of every AI programme the country runs. The BBS just put the number on the table. What the country does with it is a choice, not a forecast.

Wrapping up

Ten stories in seven days. Globally: an OpenAI model independently disproved the eighty-year-old Erdős unit-distance conjecture, crossing the line from assistant to autonomous discoverer; Pope Leo XIV issued the first papal encyclical on AI, Magnifica Humanitas, arguing that technology is never neutral and must serve human dignity rather than seek to surpass it; KPMG standardised on Claude across its 276,000-person workforce, pushing AI from pilots into the operative core of global professional services; fresh reports turned AI’s energy-and-water appetite from a forecast into a political fight, with data centres projected to draw 12 per cent of the US grid by 2028; and Microsoft moved its computer-using Copilot Studio agents to General Availability, putting production-grade screen-operating agents into the hands of every enterprise customer.

At home: Bangladesh’s National AI Policy 2026–2030 opened for public review, leading with citizen dignity and explicit bans on mass surveillance and social scoring; an Activate Rights report documented how generative AI — including deepfakes targeting women — was weaponised in February’s election; AIUB switched on the country’s first sovereign NVIDIA DGX supercomputing lab, aimed squarely at health, agriculture, and textiles; over 650,000 freelancers began reorganising themselves from solo gig-workers into AI-specialised agencies; and the BBS laid bare a digital divide along gender and regional lines that AI threatens to widen unless something deliberate counteracts it.

If there is one thread running through both halves of this week, it is this: the frontier question this week was not “how smart” but “for whom, and at what cost to the human.” The maths model that disproved an Erdős conjecture is dazzlingly smart — but the question that followed was about authorship, trust, and what happens to the humans whose insight it displaced. The pope’s encyclical did not dispute that AI is powerful; it asked who it serves. KPMG’s deployment is not in doubt as a capability; the open question is what happens to the bottom of the pyramid. The energy reports are not about model quality at all — they are about whose power bill and whose water table pays for the intelligence. And at home, every story bent the same way: a policy that asks whose dignity, a report that asks who was harmed, a supercomputer that asks for which problems, a workforce asking who gets to climb, and a divide that asks, most sharply of all, for whom is any of this being built? Smartness was never the scarce thing this week. The scarce thing was an honest answer to for whom, and at what cost. Bangladesh, for once, is asking that question in its founding documents before the technology arrives at full force. Whether it answers honestly over the next five years is the open question.

If something in this post caught your interest, share it with one person who does not follow tech. The conversation only spreads if we keep starting it.

See you next week. Stay well.