Singapore’s AI Scene Is Already Here

Singapore’s AI Scene Is Already Here | 65labs

“You cannot outsource your personal understanding.” These were the words spoken by Dr. Vivian Balakrishnan, Singapore’s Minister for Foreign Affairs, in a room of AI builders at AI Engineer Singapore, which was held from 15 to 17 May 2026.

Initially, this felt like a personal rule to follow while working with AI. However, at the end of the three days, it became a principle for the wider ecosystem. If we want AI systems to go beyond just the demos, it is not sufficient to simply use them or wait for better models. We have to understand how they work, what can go wrong, and how it can be fixed.

This stayed with me throughout the conference. Across the workshops, talks and demos, I saw a visible shift in focus from what AI can do to what it takes to make AI work with real users, data, evolving requirements, security risks and constraints. This is what made the conference feel unique. It was not simply a showcase of tools or technical deep dives into AI concepts. Rather, it brought together founders, researchers, developers, product managers, designers and students in the same room. It created the kind of density that Singapore’s AI ecosystem needs more of, that is, people moving from passively following updates to actively asking questions, having conversations and building real systems.

However, this density did not appear out of nowhere. There is already a lot happening in Singapore through hackathons, build nights, and smaller community events. AI Engineer Singapore brought this energy together at scale. Personally, this was one of the strongest signals from the conference. Singapore already has people who want to create something. What I believe it needs now is to go beyond one-off meetups by providing platforms to help builders keep going after the first conversation, share work, find collaborators, receive mentorship, and turn their interest into serious projects.

Over the three days at the conference, AI builders were thinking less about models in isolation, and more about how to create the systems around these models. Reliability, evaluations, security, deployment, and human judgment became a recurring theme throughout. One of my biggest takeaways is that the hard problems today are no longer simply about whether a model is capable. They are about whether they can work in environments where we actually need it.

On Day 1, there were about 20 workshops alongside a leadership track. The sessions I joined, by Exa, Convex, Resaro, and Arize, covered different parts of the AI stack: web-scale search, real-time backends, scenario-specific evaluation, and production agent workflows. Together, they reiterated that the model is only one part of the product. Search for AI agents is not the same as search for humans. Broad benchmarks are not enough for evaluations, and a better alternative could be operational design domains, golden datasets, and test cases that reflect the specific conditions that the system should be able to handle. Production agents require planning, context management, observability, and feedback loops. What I took away from these workshops was not just the tools, but rather a different way of thinking: AI engineering is about understanding what needs to be retrieved, tested, monitored, and bounded, so that the system can behave reliably. The talks over the next two days reiterated the same idea. The topics spanned across coding agents, world models, design, developer tools, infrastructure, security, and research problems. The speakers, across OpenAI, DeepMind, Cursor, Figma, and many more, did not only talk about what was possible. They also discussed how we make these possibilities reliable, useful, and safe enough for people to depend on.

This is important because if we simply treat failures as a model problem, the answer would be to always wait for a stronger model. However, if we treat failures as a system problem, it introduces actions that AI builders can take, such as improving workflows, adding feedback loops, and testing more carefully. I learnt that reliable AI does not come from a clever prompt or a better model. It depends greatly on the small engineering decisions that we make consistently.

Beyond these sessions, the most valuable part of the conference for me was the conversations I had. Some of my best learning happened at the booths and after the talks, which started with a single question but gave me valuable insights and tips I can apply as a student. This is also where the sponsored student tickets mattered. The student cohort, comprising twenty students, was not just given access to the events. We were also given the opportunity to ask speakers questions directly, meet the people behind some of the greatest work at the forefront of AI, and understand how they think about real problems. That made a huge difference because it treated students not as future participants in the ecosystem, but as people who should already be part of the conversation.

Speaking with Gavriel Cohen, the creator of NanoClaw, made security feel less like a separate topic, and more like something that has to be designed into agent systems from the start. He spoke about prompt injection, containerisation, agent architecture, and taking the projects end-to-end. His advice is very valuable for young builders, especially for students, that is, start with what interests you, play around with it, and then go deep enough to understand the whole system.

Other conversations and talks also widened this view, especially around research. As someone who is really interested in AI research, I was excited to hear from both Daria Soboleva and Sara Hooker. Daria’s work on mixture-of-experts models connected ideas we had learnt academically to research and system questions that builders are thinking about. Sara Hooker’s focus on model adaptation, and our conversation on interpretability and evaluation, showed that “the future is adaptable”, but we need better ways to understand how they can be evaluated.

In our conversation with him, Geoffrey Huntley offered a different perspective. He spoke about the value of building things from the ground up, and how new agent problems we see today map back to older computer science ideas such as schedulers, memory management, tool execution loops, and formal verification. He highlighted that fundamentals still matter, especially when tools make it easier to build quickly without understanding what is happening underneath. For all students and early-career builders, this is a really important point. While it is tempting to chase every new model or framework, in order to be at the forefront of advancing the field, we must not only keep up with the tools, but also pay attention to the current pain points such as reliability, evaluation, security, deployment, and accountability.

This kind of access that we received as part of the student cohort made a big difference. I learnt a lot from the workshops and talks, and just as much from the conversations with speakers and builders. Being at AI Engineer Singapore was also really motivating. What stayed with me was the feeling of being surrounded by so many people and students who are building, asking questions, and sharing what they know. Being part of the student cohort also gave me a small community within this large conference - friends who were curious about research, security, products, and real-world AI systems, and who were trying to understand where they could contribute. The energy in the room makes all the difference between thinking about starting a project and actually starting one.

Thus, for Singapore, AI Engineer Singapore highlighted, as Dr. Vivian Balakrishnan stated, that while Singapore may not be at the frontier of model development, it can be at the frontier of “deployment at scale.” This is really important because if the next phase of AI progress depends heavily on making systems reliable, secure, and useful, then deployment becomes frontier work.

What made his point powerful was that he was not just giving the talk as a minister. He spoke about building and using his own “second brain,” which showed what it can look like when leaders engage with the tools. As he put it, “you cannot govern a technology you have only been briefed on.” If AI is going to shape policies, then it is important for the leadership to have a hands-on understanding in order to make more grounded decisions.

This also applies to the wider ecosystem. It is clear that Singapore already has strong talent. Even within the student cohort, I met peers working on research, security, products, and building real AI systems. However, such talent also requires platforms such as meetups, hackathons, mentorships, and funding pathways. We can create a strong ecosystem if we not only celebrate adoption, but also reward people for trying and building.

This is what I think would make Singapore’s AI ecosystem stronger. We already have the people. We also have smaller communities and events that are already happening. AI Engineer Singapore showed us what it feels like when this existing energy is brought together at scale. In the closing remarks, the idea that we have “no scene” in Singapore was pushed back on, and every single person in that room was, in fact, the “scene.”

This deeply touched me because it made me realise that we don’t have to wait for permission to share what we are working on. The ecosystem will become stronger when more people are willing to build in public and learn from one another. The next step now should be to keep showing up after the conference: build more, share more, ask more questions, and make it easier for builders and students to take their ideas further.

By the end of these three days, my biggest takeaway is to start and be consistent.

I left AI Engineer Singapore with a lot of thoughts, questions, and most importantly, the motivation to build something. The future of AI no longer feels distant, but more like something we can all be active participants in. If Singapore’s AI scene is going to flourish, it will be because enough of us decide to keep showing up for it.