Weekly AI Roundup: What Happened in AI This Past Week (June 27–July 3, 2026)

Welcome back. It is the start of July, the rain is taking the occasional breather now, and the week just gone had a satisfying shape to it: several of the arcs this newsletter has been tracking for a month actually closed. The most powerful models that the U.S. government switched off in mid-June came back on. Anthropic, having spent weeks warning about the dangers of its own technology, pointed that technology at something unambiguously good — the search for drugs for neglected diseases. Europe finalised the biggest revision to its AI rulebook. And in Dhaka, the country’s central bank did something quietly significant: it told its own staff to stop pasting confidential financial data into ChatGPT. This was a week less about new capability and more about institutions deciding what they will and will not allow AI to do — which is, in the end, the more important story.

Two halves as always. Five global, five from home. The same two questions on each — who benefits, who pays — and the same effort to see each story from Dhaka as well as from the coasts. Let us begin.

5 Big International Developments

1. Anthropic launched Claude Science — and pointed its own models at drug discovery

The most interesting global story of the week was about turning AI toward science. On June 30, Anthropic launched Claude Science, a research “workbench” designed to do for scientists roughly what Claude Code did for engineers — running agentic research tasks, spinning up on-demand compute environments, tracing results back to the code that produced them, and offering more than sixty curated scientific skills and data connectors across biology, chemistry, and adjacent fields. It is available to all paid Claude subscribers. Alongside it, Anthropic said it would use the tool internally to pursue its own drug-discovery research aimed at rare and neglected diseases — the conditions that commercial pharmaceutical pipelines routinely ignore because the patient populations are too small to be profitable. The same day, the company also rolled out Claude Sonnet 5, a cheaper everyday model pitched for coding, research, and long-running agents.

This is the constructive counterweight to a month of alarming AI news. For weeks this newsletter has covered export controls, national-security shutdowns, and warnings about recursive self-improvement. Claude Science is the other face of the same technology: agentic AI aimed at the parts of science that markets under-serve. Rare diseases are the textbook case of market failure in medicine — real suffering, no commercial incentive — and a lab using its most capable models to work on them, at genuine cost to itself, is exactly the kind of thing the technology’s optimists have been promising.

Who benefits, and who pays? Scientists — especially those without the resources of a major lab — benefit from a research tool that compresses the grunt work of computational science. Patients with neglected diseases benefit if the drug-discovery push produces anything real. Anthropic benefits reputationally, and commercially, by demonstrating a high-value scientific use for its models. The honest caveat is that “AI drug discovery” has a long history of overpromising; the benefit is real potential, not yet a delivered cure, and it should be held as such.

The Dhaka view is genuinely hopeful, because Bangladesh’s disease burden is heavy on exactly the conditions that global pharma neglects. Neglected tropical diseases, drug-resistant infections, and conditions concentrated in poorer countries are precisely the “unprofitable” targets Claude Science is being pointed at. A tool that lowers the cost of serious computational biology also, in principle, lowers the barrier for Bangladeshi researchers — at ICDDR,B, at the public medical universities, at the pharmaceutical companies that already make Bangladesh a generics powerhouse — to do frontier-adjacent research on the diseases that actually afflict this region. Whether that potential is realised depends on whether those institutions have the researchers, the training, and the compute to use tools like this. It is a real opportunity, and it will be missed by default unless someone deliberately reaches for it.

2. The U.S. lifted the export controls on Claude Fable 5 and Mythos 5

The arc that opened this newsletter’s coverage in mid-June closed this week. On June 30, the U.S. Department of Commerce lifted the export controls that had forced Anthropic to disable Claude Fable 5 and Mythos 5, ending the standoff. Fable 5 was set to return for global users, and Mythos 5 access had been restored for some vetted U.S. organisations after government approval a few days earlier. The most capable models on the market, switched off by government order on national-security grounds two and a half weeks ago, were switched back on.

The whole episode — shutdown to restoration in under three weeks — is the most instructive governance story of the summer, precisely because it happened fast and then reversed. It established, in practice, that the U.S. government can disable a frontier model on short notice; and then that it can restore it just as quickly, presumably once the specific concern (a jailbreak that could weaponise the model) was judged managed. What it did not establish is any transparent, predictable process — the whole thing ran on executive discretion, opaque from the outside. That opacity is the real story: frontier-model availability is now subject to fast, discretionary state control, with no public rulebook governing when it is exercised.

Who benefits, and who pays? Anthropic benefits from getting its flagship products back. Users worldwide benefit from restored access. The one who “pays,” in a structural sense, is predictability itself: every business and developer now knows that access to the frontier can be revoked and restored by executive decision, on a timeline and by a logic they cannot see. That uncertainty is a cost even when the models are on.

The Dhaka view closes the loop from three weeks ago. Bangladeshi developers who lost Fable 5 access overnight have it back — but the lesson of the round trip should not fade with the inconvenience. Access to the frontier, for everyone outside the approving jurisdiction, is now a variable, not a constant. The right response is not relief that it came back; it is architecture that assumes it might go again. Multi-provider design, open-weight fallbacks, and portability are the concrete form of that lesson, and this month proved why they matter.

3. The EU finalised its AI Act revision — and banned AI-generated intimate imagery

Europe moved on the rulebook. On June 29, the Council of the EU gave final approval to a regulation streamlining parts of the AI Act — the “Digital Omnibus” package — which, among other things, delays the application of the toughest high-risk rules to December 2027 for stand-alone high-risk systems and August 2028 for high-risk AI embedded in products. Crucially for a theme this newsletter keeps returning to, the package also added a new prohibition on AI-generated non-consensual intimate imagery and child sexual abuse material — banning tools that generate nude images of real people or “undress” existing photos, with that prohibition set to take effect from December 2026. The AI Act’s transparency rules, separately, take effect in August 2026.

The revision is a study in the trade-offs of AI regulation. On one hand, Europe blinked on the hardest parts, delaying high-risk enforcement under industry pressure and fears of falling behind in competitiveness. On the other, it drew one of the clearest and most morally unambiguous lines yet: a hard ban on the AI tools used to manufacture non-consensual intimate imagery. That second part is the piece that matters most for the world beyond Europe, because the “undress” and deepfake-porn tools it targets are precisely the ones fuelling the abuse crisis in countries — Bangladesh very much included — that have no such ban of their own.

Who benefits, and who pays? European AI companies benefit from the delayed high-risk timelines. Potential victims of AI-generated intimate-image abuse benefit, in Europe, from a clear prohibition on the tools that produce it. The ones who pay are the advocates who wanted the high-risk rules enforced sooner, and — outside Europe — everyone in jurisdictions where these same tools remain legal and unregulated.

The Dhaka view writes itself, and it is a direct challenge to Bangladesh’s lawmakers. The EU just banned the exact category of tool that is being used to destroy Bangladeshi women’s lives — the “undress” apps and deepfake generators at the centre of the abuse crisis this newsletter documents every week. Bangladesh does not need to invent a legal theory to act; it can look at what Europe just did and adapt it. A clear, enforceable prohibition on the creation and distribution of non-consensual synthetic intimate imagery, paired with the enforcement capacity to make it real, is now not just a good idea but a demonstrated policy template. The question is whether Dhaka will use it.

4. Together AI raised $800 million as AI-infrastructure funding surged

The capital kept flowing to the layer beneath the models. Together AI, which builds infrastructure for running open-source AI models, raised an $800 million Series C led by Aramco Ventures at a reported $8.3 billion valuation. It landed in a heavy week for AI-infrastructure financing, part of a broader pattern of enormous sums flowing not just to the frontier labs but to the companies building the picks-and-shovels layer — the serving infrastructure, the optimisation, the tooling that makes open-source models usable at scale.

The significance is where the money is going. Together AI’s whole proposition is helping companies run open-weight models efficiently — the alternative to depending on a single closed frontier lab. An $800 million round for that proposition is a market vote that the open-source path has real commercial legs, not just idealistic appeal. In a month defined by governments switching closed models on and off, capital is quietly backing the infrastructure that makes the open alternative viable.

Who benefits, and who pays? Together AI and its backers benefit; so, indirectly, does the entire open-weight ecosystem that depends on serviceable infrastructure. The ones who “pay” are the closed labs whose pricing power the open path is designed to undercut.

The Dhaka view is that this is quietly good news for cost-constrained ecosystems. The healthier the open-model serving ecosystem, the cheaper and more portable frontier-adjacent AI becomes for exactly the developers — in Dhaka, in Chattogram — who cannot absorb dollar-denominated closed-API bills. Bangladesh benefits passively from a well-funded open-infrastructure layer, and could benefit actively if local cloud providers and businesses learn to deploy open models on it rather than defaulting to the closed APIs whose availability, as this month showed, is not guaranteed.

5. Meta launched a cloud business to sell its spare AI compute

A strategic pivot worth noting. Around July 1, Meta announced “Meta Compute,” a new cloud business to sell its excess AI infrastructure — a notable move for a company that has spent tens of billions building out compute for its own AI ambitions and is now looking to monetise the surplus. The context is that Meta’s infrastructure build-out has run ahead of its own product needs, leaving a surplus to sell — a move rivals like SpaceX have also made — amid competitive pressure as rival consumer AI products cross the billion-user mark.

The significance is that it signals a maturing, and slightly nervous, compute market. When a company that built enormous infrastructure for itself starts renting out the excess, it is both a sign of scale and a hedge against having overbuilt. Meta joining the cloud-compute business also intensifies competition in a market dominated by the established hyperscalers, which over time tends to push prices down — the thing that matters most to everyone who rents rather than owns compute.

Who benefits, and who pays? Meta benefits from monetising idle infrastructure; customers benefit if more competition lowers compute prices. The ones who pay are the incumbent cloud providers facing a new large competitor, and — in the longer run — the question of whether the industry has collectively overbuilt compute in a way that will eventually need to be absorbed.

The Dhaka view is, again, from the consumer end of the stack. Bangladesh rents compute; it does not own hyperscale infrastructure. More competition among the giants that sell it is, marginally, good for anyone in Bangladesh paying cloud bills. The more strategic point is that the country’s own ambitions for local data-centre capacity — floated in the “Smart AI City” language of the FY27 budget — sit in a global market where the giants are now fighting over exactly this business. Building sovereign compute in that environment is possible but hard, and it should be approached with clear eyes about the economics rather than as a matter of national pride.

5 Stories That Hit Closer to Home (Bangladesh)

1. Bangladesh Bank restricted staff use of AI tools over data-leak fears

The most significant home-front story was a directive, not a launch. On June 28, Bangladesh Bank issued an order requiring its officials to obtain prior approval before using AI tools — ChatGPT, Gemini, Claude, Grok, DeepSeek — for official work such as drafting office notes, preparing policy documents, analysing data, or writing reports. The stated concern was concrete and correct: that sensitive, confidential financial data could be transferred outside the country through these tools, creating data-leakage, policy-breach, and security risks. The order noted the rising informal use of these tools among staff — which is to say, people were already pasting bank data into chatbots, and the central bank moved to stop it.

This is exactly the right instinct, and it is worth praising without irony. Every organisation in the world is discovering that its employees are quietly feeding confidential material into AI tools whose data-handling they do not control, and most have not yet reacted. Bangladesh’s central bank identifying the risk — the offshore transfer of confidential national financial data — and drawing a clear line is a genuinely mature piece of institutional governance. It is the “amanah” principle this lab talks about, applied correctly: sensitive data is a trust, and you do not hand a trust to a system whose data flows you cannot see.

Who benefits, and who pays? The public benefits from a central bank that protects confidential financial data rather than leaking it into foreign AI systems. The institution benefits from a defensible governance posture. The ones who “pay” are the staff who lose a convenient productivity tool — but that is the correct trade when the data at stake is the country’s financial system.

The Dhaka view is that this should be a template, not an outlier. Every Bangladeshi ministry, bank, hospital, and law firm has the same problem Bangladesh Bank just addressed — employees using foreign AI tools on confidential data, with no policy governing it. The central bank’s directive is a model other institutions should copy immediately. And it points to a real opportunity: the demand for locally hostable, data-sovereign AI tools — models that run inside the country’s own infrastructure so that confidential data never leaves — is exactly the kind of need a serious domestic AI industry could meet. The restriction is a defensive move; the offensive move is to build the sovereign alternative it implies.

2. Parliament passed the Cyber Security (Amendment) Bill, 2026

The legal machinery moved. On June 30, the Jatiya Sangsad passed the Cyber Security (Amendment) Bill, 2026, which — among other provisions — targets the swift removal of harmful AI-generated content and strengthens the regulator’s powers to compel takedowns, including obligations on large platforms. It is the clearest legislative response yet to the AI-content and deepfake harms this newsletter tracks weekly.

The significance is that Bangladesh is legislating on AI harms at all, and specifically on the fast removal of synthetic content. That is real progress against the enforcement gap I keep naming. The caveat — and it is a serious one, given the country’s history with cyber legislation — is that broad content-removal powers are a double-edged instrument. The same authority that can pull down a non-consensual deepfake can, in the wrong hands, pull down legitimate speech. Bangladesh’s previous cyber laws have a well-documented record of being used against journalists and critics, and the value of this amendment depends entirely on whether it is applied to protect victims or to police dissent.

Who benefits, and who pays? Victims of AI-generated abuse benefit if the takedown powers are used to protect them. The regulator benefits from expanded authority. The ones who could pay are journalists, activists, and ordinary users, if the broad powers are turned toward speech rather than harm — the recurring risk with this category of law.

The Dhaka view is cautious support with a demand for safeguards. The country needs the power to remove non-consensual synthetic content quickly; it also needs firm guarantees that the power will not be repurposed against legitimate expression. The right posture is to welcome the tool and to watch its use like a hawk — because the difference between a law that protects women from deepfakes and a law that silences critics is entirely in the enforcement, and Bangladesh’s track record demands vigilance.

3. A report warned that 40 per cent of Bangladesh’s workforce is exposed to automation

The future-of-work question got a hard number. Reporting around the end of the week highlighted a warning that roughly 40 per cent of Bangladesh’s workforce could face disruption from automation and AI, with particular pressure on the export-led, labour-intensive ready-made-garment model that anchors the economy. Some estimates cited in the coverage suggested that a large share of female garment jobs could disappear on current trends over the coming decades. The reporting flagged weak links between the country’s technical-vocational training system and actual industry needs, and the disproportionate vulnerability of youth, women, and persons with disabilities.

This is the story that matters most for the largest number of Bangladeshis, and it is the one that gets the least glamorous coverage. Bangladesh’s economic miracle was built on labour-intensive manufacturing — precisely the category most exposed to automation. The garment sector employs millions, disproportionately women, and its vulnerability is not a distant abstraction; it is the central economic question of the country’s next two decades. The “physical AI” that new ventures such as Jeff Bezos’s Prometheus are building is exactly the kind of technology that eventually reaches a factory floor.

Who benefits, and who pays? The firms that automate successfully benefit from higher productivity and competitiveness. The ones who pay are the workers at the base of the manufacturing stack — disproportionately women — for whom “disruption” means a lost livelihood without an obvious next rung. That is not a marginal tail; in Bangladesh’s economy, it is a huge population.

The Dhaka view is that this is where the country’s AI conversation is dangerously lopsided. The headlines go to the Tk 500 crore fund, the olympiad medals, and the startup ecosystem — the upside, concentrated among the educated and connected. The 40-per-cent-exposure story is the downside, concentrated among the workers who built the country’s prosperity and have the least cushion against its disruption. A serious national AI strategy has to fund the reskilling and social-protection infrastructure for that population with at least the energy it brings to the startup fund. The augmentation-over-replacement slogan that gets repeated in university seminars and industry panels is tested here, in the garment districts, not in the seminar hall.

4. A KOICA-backed national AI hub moved forward

An infrastructure story with an international partner. Bangladesh advanced a project to establish a national AI hub, backed by a KOICA (Korea International Cooperation Agency) grant — a roughly Tk 192 crore project, with the bulk funded by a Korean grant, slated to run from mid-2026 through 2029, pending final Planning Commission approval. The hub is intended to build national capacity in AI research, training, and application.

The significance is that it is capacity-building with committed external funding, which addresses one of the real constraints on Bangladesh’s AI ambitions: money for the unglamorous infrastructure of research and training. A well-run national AI hub could be the institutional home that the olympiad pipeline, the civil-servant training, and the research ambitions all currently lack — a place where the scattered efforts consolidate into something durable.

Who benefits, and who pays? Researchers, students, and government all benefit from a funded national AI institution, especially one substantially paid for by an external grant. Korea benefits from the soft-power and partnership relationship. The risk — the way this “pays” — is the familiar one: that the hub becomes a building and a budget line rather than a functioning institution, which depends entirely on the quality of its leadership and its independence from politics.

The Dhaka view is cautious optimism conditioned on execution. Bangladesh has a pattern of announcing institutions that underdeliver. A KOICA-funded AI hub is a real opportunity to consolidate the country’s fragmented AI capacity-building — but only if it is staffed by serious people, connected to the actual talent pipeline (the olympiad students, the university researchers), and protected from becoming a patronage vehicle. The funding is the easy part; the institution-building is the test.

5. The National AI Policy stayed unfinished as the deployments multiplied

The backdrop, still moving slowly. The National AI Policy 2026–2030 remained in finalization this week, its risk-based framework and its provisions on surveillance limits and algorithmic impact assessments still short of cabinet approval or gazette notification. Meanwhile, as recent weeks show, the deployments and directives kept coming — the central bank’s restriction and the cyber amendment this week, the traffic cameras before them — each of them the kind of thing the policy is meant to frame.

The significance is the widening gap between the pace of AI deployment and the pace of AI governance. Bangladesh is now regulating AI in pieces — a central-bank directive here, a cyber amendment there — while the overarching policy that is supposed to make those pieces coherent remains a draft. Piecemeal governance is better than none, but it risks inconsistency: rules made ministry by ministry, without the common framework the national policy would provide.

Who benefits, and who pays? In the near term, the absence of a binding framework gives deployers flexibility. The ones who pay are the citizens whose protections remain aspirational, and the coherence of the country’s whole AI-governance effort, which fragments the longer the unifying policy stays unfinished.

The Dhaka view is the one I keep returning to: finalize it. Every week that passes brings another AI deployment or directive — several in just the past two weeks — that the policy is meant to govern. The draft is good. The gap between a good draft and an enforceable, coherent framework is the single most consequential unforced delay in Bangladesh’s AI story, and it is entirely within the government’s power to close.

Wrapping up

Ten stories in seven days. Globally: Anthropic launched Claude Science and pointed its own models at rare-disease drug discovery, the constructive face of a month of alarming AI news; the U.S. lifted the export controls that had disabled Claude Fable 5 and Mythos 5, closing a three-week arc that proved frontier-model access is now a variable under discretionary state control; the EU Council finalised its AI Act revision, delaying the toughest high-risk rules but banning AI-generated non-consensual intimate imagery outright; Together AI raised $800 million, a market vote of confidence in the open-model infrastructure layer; and Meta launched a cloud business to sell its spare AI compute, intensifying competition beneath the models.

At home: Bangladesh Bank restricted staff use of AI tools over fears of confidential data leaving the country — a model of institutional governance other bodies should copy; Parliament passed the Cyber Security (Amendment) Bill, 2026, arming the state to remove harmful AI-generated content while raising real questions about how those powers will be used; a report warned that 40 per cent of the workforce is exposed to automation, the downside story that Bangladesh’s AI conversation keeps under-weighting; a KOICA-backed national AI hub moved forward; and the National AI Policy stayed unfinished as the deployments it is meant to govern kept multiplying.

If there is one thread this week, it is institutions drawing lines. Anthropic drew one around what its models are for. Washington redrew the one around who can use them. Brussels drew a hard one around AI-generated abuse. And Dhaka’s central bank drew a sharp, admirable line around its own confidential data. The frontier’s raw capability is no longer the only story; increasingly, the story is what institutions decide to permit, prohibit, and protect. Bangladesh, this week, drew some good lines — the central bank’s data rule, the intent behind the cyber amendment — and left the most important one, its national policy, still undrawn.

If something here 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.