The Long-AND, Not the Short-OR

Capability without deployment is a slide. Deployment without enforceability is a liability.

The Deployment Decade; What Singapore, Beijing and MIT Are All Describing

I. Three dispatches, one signal

Three pieces landed on my desk this week, and taken together they are not three separate stories. They are three cross-sections of the same shift.

The South China Morning Post reports that Singapore has become the place where frontier models get turned into working tools rather than trained in the first place. The World Economic Forum reports that China is running the world’s first large-scale live experiment in what happens to a workforce once that deployment actually lands inside a company. And MIT FutureTech, in the most rigorous piece of AI risk research I have seen this year, has just told 272 experts to stop treating “AI risk” as one undifferentiated worry and start naming which five risks will actually produce catastrophic-scale harm in the next five years.

Read separately, these are a labour market story, a workforce story, and a governance story. Read together, they describe a single object: the deployment layer. That is the unglamorous plumbing that sits between a frontier model and an organisation that can actually run on it, and it is where the real work, the real risk, and the real value of this decade will be created. Not in the model. In the layer underneath it.

II. Singapore and the Long-AND, not Short-OR

Jordan Seow joined OpenAI in June 2025 as its first locally hired forward-deployed engineer in Asia-Pacific, when the Singapore office had fewer than twenty people. Since then the pattern has scaled sharply: OpenAI committed more than S$300 million under a joint agreement with the Singapore government, Google Cloud opened an engineering centre with Grab and DBS, and Microsoft is investing US$5.5 billion in local cloud and AI infrastructure through 2029. Anthropic joined the list on 16 September, opening a Singapore office across applied AI, finance, marketing and business roles.

What’s notable is the shape of the roles being created. Singapore is not building a major centre for training frontier models but is instead adapting those models, integrating them into existing systems, and turning them into practical tools that banks, hospitals and ministries can actually run in production. That is not a consolation prize. It is the harder problem. Anyone can call an API. Getting a hospital or a bank to trust an agent inside a regulated production workflow is an entirely different discipline, and it is the discipline Singapore has decided to own.

This is the clearest live proof I have seen of Long-AND, not Short-OR. The industry spent 2023 to 2025 arguing capability versus deployment as if a country, or a firm, had to pick one. Singapore’s answer is that you need both, running concurrently: capability access AND deployment capability AND the regulatory scaffolding to make deployment trustworthy. Andrew Ng has been making a version of this argument at the enterprise level for two years now, and his 2026 message has sharpened rather than softened: as model access commoditises, the differentiation businesses fight over will not be which frontier model they can call, but how well they can build the agentic workflow around it. Singapore has essentially industrialised that insight at the level of a nation state.

One qualifier belongs here, because I do not want this section to read as unclouded enthusiasm. Singapore’s entire model depends on continued access to frontier systems built somewhere else, and that access is not guaranteed. Washington’s decision this year to briefly restrict non-US access to Anthropic’s most capable model was a small but real demonstration of how quickly the ground can move under a deployment hub that owns no frontier lab of its own. A regional strategy built on being the best place to deploy other people’s models is, by construction, a strategy with a single point of geopolitical failure sitting outside its own borders.

III. China’s three-stage transition, and the Frozen Workforce risk

The WEF’s dispatch on China is the most concrete workforce transition data I have read this year, and it deserves to be read as a leading indicator, not a China story.

The Adecco Group’s Ian Lee frames the transition in three stages: disruption, as AI compresses existing tasks and roles; augmentation, as people and AI work together within redesigned processes; and creation, as new roles and forms of economic activity emerge. The first stage is already visible: coding and customer service tasks are compressing fast. But Lee’s more interesting observation is what is happening in parallel: experienced workers are being brought back into organisations to guide AI-enabled work, bringing context and business judgement that task-level automation cannot provide. His line for management is one I intend to steal outright: you cannot have AI in a bad process. Sequence matters. Understand how work actually moves through the business before you automate, augment or reassign any part of it.

This maps directly onto Erik Brynjolfsson’s J-curve, which he has been refining publicly all year. Brynjolfsson’s framing is that productivity dips before it lifts, precisely because organisations are spending the disruption stage building the “intangible assets” (redesigned processes, retrained people) that the augmentation stage will actually run on. His sharper and more useful reframe for 2026 is task-level, not job-level: almost no job disappears wholesale, but almost every job contains tasks that AI changes, and the businesses that win are the ones augmenting rather than merely automating those tasks. That is the difference between his “Turing trap,” where AI is used only to mimic and replace human labour, and the augmentation path that actually compounds.

But the WEF piece surfaces a risk that the productivity literature tends to skip past, and it is the one I want to flag hardest. Junior, routine work has always been the training ground where professional judgement is built. Exploratory research cited in the piece names the specific danger: organisations may automate some of the tasks through which junior employees developed judgement, while leaving behind work with limited learning value, and strong AI-assisted output can conceal whether the employee behind it could actually explain or challenge the result. This is the Frozen Workforce risk in its most literal form: not mass unemployment, but a leadership pipeline quietly starved of the reps it needs, years before anyone notices the gap. China’s own response, embedded in its Employment Priority Strategy, is instructive: a stated push towards new forms of human-AI collaborative work, a large-scale youth employment skills initiative aiming to help 1 million young people, and around 1 million internship positions. Whatever one makes of state-directed labour policy, that is a government treating workforce infrastructure as AI infrastructure, funded and sequenced, rather than as an afterthought to be handled by HR once the technology has already landed.

Renmin University’s Wenxia Zhou has a phrase for what this does to individual careers that I think will outlast the cycle: the “mapless career era,” where the old ladder (education, entry, gradual promotion) gives way to something closer to cultivating a forest. That is HX = CX + EX in one image. The employee experience side of this transition is not a soft add-on to the technical rollout. It is the load-bearing wall.

IV. MIT’s five risks, and the enforceability test

If Singapore and China describe how deployment is scaling, the MIT FutureTech study, with 188 co-authors and a lead team including Peter Slattery, Alexander Saeri, Jess Graham, Michael Noetel and Neil Thompson, tells you where it can break.

The study used the Delphi method to ask 272 international experts to score 24 AI risk domains for likelihood and severity out to 2030. The headline finding is sobering on its own terms: under business-as-usual, experts judged that 18 of the 24 AI risk domains had at least a 10% probability of catastrophic outcomes, where catastrophic was defined as harm involving more than a million deaths, more than $100 billion in losses, or comparable civilisational damage. Even under “pragmatic mitigation,” where organisations and governments make cost-effective efforts to reduce harm, five domains still cleared that 10 percent bar: AI systems possessing dangerous capabilities, AI-enabled weapons and cyberattacks, environmental harm, inequality and unemployment, and power centralisation.

Two of Slattery’s explanations deserve to be read in full rather than summarised away. On why weapons and cyberattacks rank so high: AI is disproportionately good at exactly the skills that matter for offensive cyber work, coding and performing pattern recognition and information synthesis, against infrastructure that is already software-dependent and therefore offers a large attack surface. On why competitive dynamics belong on the list at all, given it is not itself a harmful use of AI: it is an instrumental risk that creates other risks, the condition under which firms and states, believing AI confers decisive advantage, are incentivised to move fast, resist constraints, and underinvest in safety.

This is precisely the dynamic Dario Amodei has spent the past year trying to name from inside the industry that is doing the racing. His “race to the top” thesis holds that a safety-focused lab has to stay at the frontier, on the theory that ceding it to less careful competitors produces a worse outcome for everyone. His starker June 2026 formulation, that the most powerful systems risk resembling something closer to weaponisable material than ordinary technology, reads less like doom-mongering and more like a plain restatement of what Slattery’s competitive-dynamics finding says in academic language: the race itself is the risk multiplier, independent of what any single model can do.

The finding I keep returning to, though, is the structural one buried near the end of the MIT piece, because it is the one with the sharpest implications for governance design. AI developers and governance actors have the primary responsibility for addressing AI risks, but AI system users and the stakeholders such systems affect are most vulnerable to those risks. Slattery’s own word for this is apt: misaligned incentives. The people best positioned to fix the problem are structurally distant from the people who bear its cost.

This is exactly the failure mode TrustOS was built to close, and it is the enforceability test in its purest form. A responsible-AI policy that sits in a slide deck while the people exposed to its failure have no standing to invoke it is not governance, it is decoration. The WEF’s own governance research this year found that fewer than one in a hundred organisations has implemented full responsible AI practices. Put the MIT misalignment finding next to that number and you get a fairly precise diagnosis of why: governance obligations have been written onto the actors with the least exposure, and the actors with the most exposure have no enforceable lever to pull. Any AI governance framework that cannot answer “what can the exposed party actually do about it, and to whom” is not going to survive contact with the risks MIT just ranked.

V. What this actually asks of a leader

Put the three dispatches on one table and the sequencing writes itself.

Deployment engineering is not a downstream, lesser task behind model capability. Singapore has made a national bet that it is the higher-value, harder-to-copy layer, and Andrew Ng’s enterprise-level argument says the same thing at company scale: differentiation lives in the workflow around the model, not in which model you can call.
Workforce redesign is not HR’s problem to solve after the technology lands. It is AI infrastructure, on the same capital-planning footing as compute and data. China’s Employment Priority Strategy treats it that way. Most Western enterprises still do not.
Do not wait for a complete risk taxonomy before acting, but do not accept a governance framework that cannot pass the enforceability test either. Slattery’s own advice is the right one: fold AI risk into the governance conversations you are already having about cybersecurity, privacy and business continuity, continuously, not as a one-time compliance exercise.

The Fork here is not really Star Trek versus Mad Max at the level of civilisation. It is smaller and nearer than that: it is whether an organisation’s AI governance actually enforces anything for the people most exposed to it, or whether it just reads well in the annual report.

Singapore is answering the deployment question. China is running the workforce experiment in real time. MIT has just handed everyone else a ranked list of what to worry about while they do both. The organisations that treat these as three separate initiatives, owned by three separate teams, are the ones that will discover, in about eighteen months, that they built the deployment layer on a workforce that was never redesigned, governed by a framework nobody exposed to its failure could actually invoke.

Sources: South China Morning Post, “Why AI giants are choosing Singapore to turn models into real-world tools” (20 Sep 2026); World Economic Forum, “Lessons from Chinese business leaders about integrating AI in the workplace” (27 Aug 2026); MIT Sloan, “These are the most urgent AI risks, according to 272 experts” (20 Jul 2026), reporting on Saeri, Graham, Noetel, Slattery and Thompson, “Prioritization of Risks From Artificial Intelligence,” MIT FutureTech and University of Queensland; Fortune, “Both U.S. and Chinese AI firms are setting up shop in Singapore” (19 Jun 2026); public statements from Dario Amodei, Andrew Ng and Erik Brynjolfsson.