A few weeks before AI Engineer Singapore, Singapore's Foreign Minister Dr Vivian Balakrishnan published a detailed technical write-up of his personal AI second brain on GitHub. He had built a system that answers his questions, researches topics, drafts speeches, and gives him daily briefings.
A retired eye surgeon turned diplomat, he had every reason to leave that kind of build to someone else. Instead, he assembled it himself, the way an engineer would.
Sitting in the front row watching him speak, that's what struck me. Not his credentials, but the specific choice he had made: to understand AI by building with it, rather than being briefed on it.
A group photo with Minister Vivian Balakrishnan after his keynote
I was one of 20 students sponsored by 65Labs to attend the conference's first Asia edition, held over three days at SMU and the Capitol Theatre. Most of the people around me had shipped products, managed engineering teams, or were running their own companies.
I expected a lot of talk about model capability: benchmark scores, parameter counts, which lab is ahead. There was some of that. But the talks that landed hardest kept returning to a different question: given that you have access to a capable model, what do you actually do with it?
The system is the product
One answer came from Tejas Kumar of IBM, and it pushed the question further than I expected. His argument was direct: with the right harness around a model - the right prompts, tools, and structure - you can get strong results even from a weaker model. He demoed this live on stage using GPT-3.5-Turbo, a legacy model most builders have moved past, building a browser agent to upvote a Hacker News post.
Tejas Kumar presenting on harnesses
What made it stick was watching the same model behave completely differently depending on how the harness was built. With vague tool definitions and no real error handling, the agent hallucinated and reported the task done when nothing had happened. As Tejas tightened the tool definitions and introduced explicit error responses that surfaced real failure states, the agent became honest about what it couldn't do, and eventually completed the task. The model never changed. The harness did.
It's a small demo, but it points at something bigger. For most of the last few years, the dominant conversation in AI has been about model capability, on the assumption that access to the best model put you most of the way there. What the builders at AI Engineer Singapore kept saying, in different ways, is that this is no longer the interesting part of the problem. The model is becoming infrastructure. The system built on top of it is the product.
The human in the loop has to actually be in the loop
If the system is where the work has moved, the next question is who's responsible for it. Gavriel Cohen, creator of NanoClaw - the tool Dr Vivian used to build his second brain - put it plainly in his keynote: agents do the work, but whoever presses approve is accountable. Capability doesn't dissolve responsibility, it concentrates it.
A candid photo with Gavriel Cohen after his keynote
This echoes Dr Vivian's message from earlier. You can outsource computation and memory, but you cannot outsource understanding. And if you're in a position of authority, you can delegate work, but you cannot delegate accountability. A system where the operator doesn't understand what they're approving is a poorly designed system, regardless of how capable the model underneath it is. As Dr Vivian put it: "You cannot govern a technology that you have only been briefed on. You had better get your hands dirty, and then you understand both the potential and the limits, and the problems."
Where Singapore fits
That same shift - from the model itself to the system and judgment around it - applies to countries too.
Singapore is not going to win the race to build the most powerful foundation model. That race is being run in San Francisco and Beijing, with compute budgets and research teams no city-state can realistically match. Singapore's Deputy Prime Minister Gan Kim Yong said as much in the Economic Strategy Review Committee: Singapore is not likely to be at the frontier of model development, but it can be at the frontier of deployment at scale.
The conversations at AI Engineer kept pointing at what that actually means. The frontier isn't in the weights. It's in the layer above the model - the evals, the tooling, the system design, the integration into real workflows. And it's in the layer around the model - the judgment, governance, and understanding - that turns a capable AI system into a trustworthy one.
For builders here, that's a more specific opportunity than it might sound. The companies that will matter in Singapore's AI economy are not the ones training models. They're the ones figuring out how to make those models work reliably in healthcare, in finance, in government services, in the operational complexity of Southeast Asian markets. That requires deep domain knowledge, rigorous evaluation, and the kind of trust that takes time to build. These are things Singapore can compete on.
The bottleneck in AI is no longer the model. It's the system built on top of it, and the people who genuinely understand what they're running. That's where the work is. And it's where Singapore has a real shot.