Building an AI Engineer Nation

Make AI useful to businesses and workers

A small business may see promising AI applications yet struggle to access the technology. An employee may have little reason to learn new tools until their employer starts using them. For Josephine Teo, these are connected problems: enterprise adoption gives workers a reason to acquire AI skills, while smaller firms need help participating. The national strategy she describes starts with better jobs, higher wages and businesses able to offer greater value.

In this Latent Space conversation, published in October 2024, Teo explains how generative AI prompted Singapore to refresh its national strategy. Consultation included practitioners and Singaporeans working overseas, especially in the Bay Area. The gathering transcribed as RAISE grew out of that outreach; Teo credits its participants with sharpening the refreshed National AI Strategy and encouraging greater ambition before its December 2023 publication. Applications developed and scaled in Singapore could, she argues, deliver public benefits elsewhere too.

The choice of digital development in the ministry’s name emphasizes the impact on communities. Teo describes four responsibilities under Singapore’s Smart Nation efforts: - Digital economy: support vibrant economic activity. - Digital society: preserve cohesion amid technologies that can deepen polarization. - Digital government: advance government’s own use of technology alongside wider adoption. - Digital security: address cybersecurity and people’s safety when using digital services, devices and social media.

Her analogy is city development: conceptualize, plan, implement, monitor and correct problems, while choosing partners and considering long-term impact. For AI, the enabling work includes making compute available to researchers and companies, updating data-protection laws to clarify legitimate business uses, and organizing data for sectors such as finance and logistics.

Talent includes the people who use the bots

The talent strategy distinguishes three groups: - AI creators: scarce specialists whom Singapore wants to engage through local work and international projects. Collaboration need not require relocation. - AI practitioners: data scientists, machine-learning specialists and engineers who support development and application. - AI users: people who must become comfortable applying the tools available to them.

A company can have a team building bots and identifying use cases while colleagues remain uncomfortable using them. Engineering capacity alone does not complete adoption. Singapore’s compact size and existing training infrastructure are presented as advantages for addressing these different needs. Providers whose courses lead to good employment outcomes are described as likely to receive support for subsidized training.

Build practical safety tools—and learn through use

The Singapore AI Safety Institute is described as working with counterparts in the US, UK and elsewhere on responsible development and deployment. [AI Verify](https://aiverifyfoundation.sg/) was developed for traditional AI; generative AI requires different testing capabilities, including red teaming and benchmarking. The tools available at the time are presented as imperfect starting points.

The conversation briefly moves from testing to mathematical proof. One participant argues that tests provide limited reassurance without a proof of safety; another raises the possibility that safety may be fundamentally unprovable. That possibility is acknowledged. Proof remains a research ambition, with no method or result offered.

Workplace experimentation provides a more immediate route to familiarity. A host recounts meeting Grab CPO Philipp Kandal and hearing that Grab paused company work for a week of generative-AI training, requiring employees to undertake an activity using the technology. This secondhand account illustrates employer commitment; the discussion anticipates future returns rather than reporting measured productivity gains.

Singapore Airlines is also described as having a dedicated team exploring AI use cases. Airline operations combine optimization opportunities, compliance obligations and prescribed processes, making them a promising application area, though specific implementations are not disclosed. Within government, officers had received generative-AI guidelines and tools for experimentation, and some were building bots for colleagues. The policy rationale is that hands-on use helps government design better guardrails.

A host challenges the emphasis on safety, arguing that some AI circles associate it with censorship, over-regulation and reduced capabilities. The host sees a possible conflict with frontier exploration but acknowledges uncertainty about the right balance. The response distinguishes largely voluntary governance guidance from recently introduced, not-yet-passed legislation specifically addressing synthetic election content.

The guidance asks developers to explain where training data comes from and what uses a model is intended to serve. Transparency should also help affected people ask better questions of developers and deployers. An electricity analogy supports rules tied to particular uses and desired outcomes: allow broader experimentation while addressing specific harms. The respondent explicitly excludes harmful conduct from that permissive approach and argues that premature regulation can be counterproductive.

Give sectors ownership and keep responsibilities connected

Singapore’s single tier of government and cooperation across ministries are presented as advantages for coordinating adoption. But each sector needs to own its applications. Transport and healthcare officials must advance work in their domains; a relatively lean digital ministry cannot own every deployment.

Rail maintenance makes the operational choices concrete: - Corrective maintenance: repair equipment after it fails. - Preventive maintenance: intervene before failures, at the cost of maintaining a broad set of equipment. - Predictive maintenance: use machine learning to prioritize interventions. The proposed benefit is a more reliable railway without substantially higher cost—a potential application, rather than a measured deployment result.

Healthcare requires a different division of responsibility. AI could help with diagnosis and treatment planning by considering a patient’s full set of health indicators, while physicians retain judgment over those inputs. The respondent does not expect physicians to be replaced in the foreseeable future.

Digital services also cross organizational boundaries: citizen-service delivery depends on security and resilience, and government standards must account for industry and cloud-provider architectures so systems can interoperate. Flexible responsibilities matter as much as formal reporting lines. The working rule is to retain ownership of an unresolved problem until someone else takes it on.

For founders seeking an entry point, the Economic Development Board is recommended as a first contact for connections to relevant people in Singapore. An unfamiliar opportunity without an institutional champion should prompt a search for the right group. A host connects that openness to plans for an AI engineering conference alongside the machine-learning conference ICLR in Singapore the following year, hoping visiting researchers will also encounter the local engineering community.

Turn training into jobs and coherent career paths

AI engineering means turning foundation models into useful products. Domain knowledge and practical skills can provide an entry point without a PhD or a computer-science degree. Entrants may change careers or train for this work from the outset. Against the strategy’s target of growing AI practitioners from 5,000 to 15,000, a host argues that recognizable job titles matter: “AI engineer” gives employers a hiring category and workers a concrete learning goal.

The host worries that courses might produce certificates without helping people succeed in industry, while acknowledging limited familiarity with Singapore’s programs beyond SkillsFuture. The response offers two routes: - Company-led training: government funding helps employers cover training resources. Employer commitment signals which skills matter. - Individual career planning: workers whose employers do not offer training can use sector-specific guidance to identify changes in their tasks and their own learning needs.

Jobs transformation maps describe how technologies could change particular roles over three to five years. The horizon is deliberately short enough to support task-level guidance: robotics may change logistics work, while AI and machine learning may change financial-services tasks. More than a dozen maps are reported as available across sectors.

Accounting illustrates the intended transition. As routine tasks move to AI tools, workers would need to learn those tools and develop other skills. Audit experience could provide a foundation for digital forensics, where investigating how fraud occurred still requires human understanding of the problem.

The maps prompt an application idea: make them accessible through retrieval-augmented generation (RAG), using retrieved documents to support a model’s answers. But career guidance also requires planning: the proposed steps must make sense in sequence and lead toward a destination. A host connects this distinction to interest in planning and reasoning around OpenAI’s `o1` model. The career assistant remains a proposal, with no system demonstration or measured result.

A general transition map and a database of individual career histories would serve different purposes. The respondent recalls exploring career-history databases with US recruitment firms: observed paths might reveal the intermediate moves needed to reach a new role. The maps provide a framework but do not contain those personalized histories. Separately, a host proposes collecting auditors’ explicit reasoning as training material for language models—another idea, rather than an implemented method.

The discussion then broadens to a host’s proposal for coding as a decades-long national project, with Python as an additional common language. The response distinguishes broad AI literacy from universal coding education. Earlier computer-literacy programs, including adult learning, make widespread AI literacy plausible. Whether everyone should learn programming remains an open educational question; wider introduction would depend on a clearer conclusion about its necessity.

Place compute where the workload requires it

The infrastructure discussion separates where computation runs from how its uses are governed. The government’s earlier cloud-first strategy is a hybrid approach: suitable workloads move to the cloud, allowing capacity to expand without permanently dedicating resources to every workload. Unsuitable workloads remain on premises with separate supporting infrastructure.

That split also shapes security spending. Moving suitable systems to cloud providers narrows the set of on-premises systems requiring direct protection, potentially concentrating safeguarding resources on that smaller set while cloud workloads rely on provider security architectures. The respondent expects government AI workloads to follow a similar hybrid approach.

At national scale, the argument is that Singapore cannot host all the AI activity it hopes to enable. Manufacturing supply chains provide the analogy: activities can move to locations suited to them, with regional specialization benefiting the wider system. Latency requirements help determine placement—some workloads may need local hosting, while others may tolerate running abroad. The respondent expects room for redistribution, without offering a completed allocation plan.

Domestic data-center expansion remains part of the approach, pursued through both lower energy consumption and greener energy. A host then proposes government-provided AI services for small businesses, including AI accountants. The response addresses compute access: a budget had been set aside for startups and researchers, though no amount or capacity figure is given.

Private interest extends beyond large cloud providers to data-center operators offering compute as a service. The stated government role is to monitor unmet demand and complement private provision. Adult education follows a related principle: choose providers according to their ability to meet industry needs. Private training intermediaries may understand employer requirements closely; polytechnics and universities also offer useful programs, without needing to be the sole providers.

AIBots: find past decisions and understand their reasoning

AIBots brings the discussion back to government employees. The internal initiative encourages staff to experiment with ministry and agency memos that preserve past decisions and institutional thinking. In a tax-appeal example, a newer Inland Revenue colleague could use records of previous cases through RAG to find how similar issues were decided more quickly.

Finding a decision still leaves the question of why it was made. Bringing that reasoning into the interaction is described as a potential next step. Earlier assistants reportedly returned standard answers and repeated them when users asked follow-up questions. The newer bots are described as substantially better at handling questions, though deployment remains early and no evaluation or measured time saving is provided.

For one host, transferring institutional knowledge and improving government services for an aging population are especially motivating applications of AI engineering.

The election boundary: depictions of actual words and actions

The conversation closes by returning to synthetic election content. The respondent distinguishes disagreement over political opinions from realistic fabricated depictions of someone saying or doing something they never said or did. Such depictions can confuse voters and unfairly help or harm the person portrayed. The stated objective is political disagreement grounded in actual words and actions.

The election-specific requirements are intended to preserve factual accuracy and public trust by limiting depictions to what people actually said or did. At the time of the conversation, the legislation had been introduced but had not yet passed. This is the concrete use-specific boundary accompanying the broader invitation to experiment with AI.