Written by Vihaan Motwani.
AI Makes Building Cheaper. Understanding Is Now the Bottleneck
5 min read · Part of Sponsoring students at AIE became a test of Singapore's AI ambition
Written by Vihaan Motwani.
5 min read · Part of Sponsoring students at AIE became a test of Singapore's AI ambition
AI is making it easier to build things we do not fully understand.
That was the uncomfortable thread running through AI Engineer Singapore. The conference was full of coding agents, design tools, robotics workflows, agent harnesses, and talks about software abundance. The demos were impressive, but the bigger question was harder: when building becomes cheaper, who is responsible for knowing what is actually being built?
Dr Vivian Balakrishnan gave the clearest version of that question. In a technical talk on AI adoption, personal workflows, and accountability, he said we can now outsource calculation, computation, memory, and knowledge retrieval. But we cannot outsource personal understanding. And if we are in positions of responsibility, we cannot outsource accountability either.
It also captured what made the conference feel different. This was not a room asking whether AI could generate code, interfaces, documents, or agent workflows. That question already felt answered. The more serious conversations were about what has to surround those systems before they can be trusted: evals, sandboxes, checkpoints, traces, deterministic boundaries, human review, and better feedback loops.
This is probably the most important shift for AI builders in Singapore right now. The bottleneck is moving from access to judgment.
For the last few years, much of the AI conversation has been about who gets to use the tools. That still matters. But at AI Engineer Singapore, the more interesting issue was what happens after the tools are in everyone’s hands. Starting is easier now. Shipping a prototype is easier. Generating code is easier. But understanding the system well enough to deploy it responsibly is not getting easier at the same pace.
JJ Geewax from Google DeepMind made this point in a practical way. Models are already useful enough to build with, but production systems still need boundaries. You cannot keep waiting for the perfect model. At some point, builders have to decide where the model belongs, where deterministic logic should take over, and how to make the whole system reliable enough for real users.
That is very different from simply prompting a model well.
Geoff Huntley’s talk pushed the idea further. His argument on software abundance was deliberately provocative: if software development becomes cheaper than minimum wage, then the structure of companies starts to change. Smaller teams can do more. More people become builders. Some organisations may discover that their old processes were designed for a world where software was scarce.
But abundance creates its own problems. When everyone can produce software, producing software is no longer the advantage by itself. The advantage shifts to choosing the right problems, building the right feedback loops, and knowing when not to automate.
The design talks made the same point from another angle. Several speakers pushed against the idea that AI output is automatically good output. Taste, intent, brand fidelity, editability, and context still matter. One framing from Jay Demetillo stood out to me: prompts do not have opinions. People do. AI can generate options, but it cannot decide what a product should feel like for a specific user in a specific moment.
For students and early-career builders, that is probably the most useful lesson. The future does not belong only to people who know the latest model or framework. Those tools will keep changing. The more durable skill is learning how to think with them without surrendering judgment to them.
For founders and operators, the takeaway is more immediate. Do not treat AI adoption as a tool-buying exercise. Build evals before demos become decisions. Treat AI fluency as a leadership skill, not just an engineering skill. Keep humans accountable for systems even when agents do the work. Most importantly, do not confuse faster software creation with better product judgment.
Singapore has an interesting role to play here. We may not be the country training every frontier model, but we can be a place where AI adoption is taken seriously across government, startups, enterprises, and education. The conference brought these worlds unusually close together: a minister talking about his own AI workflow, engineers talking about agent harnesses, designers talking about taste, and robotics builders talking about simulation and real-world evaluation.
That mix matters. Southeast Asia does not need to copy Silicon Valley’s AI story exactly. The region has its own operational realities: multilingual users, resource-constrained SMEs, public-sector complexity, cross-border markets, and uneven access to technical talent. These constraints make responsible deployment harder, but they also make the region a useful testbed for AI systems that have to work outside perfect lab conditions.
A conversation I had with Bhavan Jaipragas from The Straits Times added a wider layer. We spoke about equal access to AI and the future of employment. If AI makes software and services cheaper to produce, who benefits from that abundance? Universal Basic Income is one possible answer. Another more radical idea we discussed was whether compute itself could someday become a public utility, like electricity, provisioned in some basic amount to every household.
That was not the main theme of the conference, but it echoed the same underlying question. When capability becomes abundant, responsibility does not disappear. It moves somewhere else.
That is what I left AI Engineer Singapore thinking about. AI will let more people build, faster than before. That is genuinely exciting. But the real work is not only in making things. It is in understanding them deeply enough to know what they are doing, where they fail, who they affect, and what should happen next.
The tools can do more of the work now.
They still cannot do the understanding for us.