From Labs to Nation-Scale AI Products

Building AI that actually works in the real world isn't just about having the most cutting-edge technology. It requires resource management, building trust, and understanding human context. Dr. Leslie Teo, Senior Director of AI Products at AI Singapore, joined Dr. Tutanon Sinthuprasith of SCBX to share critical insights on driving AI adoption across Southeast Asia. Here are the key takeaways:

1. Think Like an Investor: 3 Principles for Survival

Dr. Leslie applies principles from the finance world to AI development to ensure scalability:

• Make a clear bet: Focus on areas where you have a durable, long-term advantage.
• Diversify: Don't put all your eggs in one basket: be ready for the uncertainty of fast-changing tech.
• Build Trust: Trust is essential to weather shocks and is the primary driver for getting people to actually use the technology.

2. The Journey of "SEA-LION": An ASEAN LLM

The SEA-LION project is a prime example of developing a nation-scale model. Its evolution is divided into three distinct phases:

• Phase 1: The initial bet was on "Multilinguality" using regional data. They trained the model from scratch to prove that local data truly makes a difference.
• Phase 2: Shifted to "Continued Pretraining" on top of the best existing open models. This method cuts costs and reduces "catastrophic forgetting" while successfully retaining regional language capabilities.
• Current Phase: The focus is now on smaller, efficient, and safer models. The goal is to run these models on standard laptops—not just expensive H100 GPUs—making AI truly accessible to everyone.

3. Speed vs. Ethics: You Need Both

Many believe that worrying about "Ethics" (safety, privacy, bias) slows down development. Dr. Leslie argues the opposite: "Trust makes you faster."

If users are confident that the model understands their culture and isn't designed to replace them, they will open up. They become willing to adopt the technology and share their data. Therefore, prioritizing local context and safety isn't just a "nice-to-have"—it is a necessary strategy for speed.

4. The Enterprise Blocker: Not Tech, but the "3 Ps"

While foundation models act as infrastructure (like roads), implementing them in a company is unique to each organization. The real blockers are rarely the models themselves, but the 3 Ps: People, Process, and Policy.

• Don't start with hardware: Don't rush to buy expensive servers or GPUs. Start by renting (Cloud) to test and prove the Return on Investment (ROI) first.
• Embrace the "boring" work: Success isn't about using the absolute latest model. It’s about the backend systems—clean data pipelines, consistent monitoring, and governance. These "boring" tasks are the only things that make your AI reliable and usable in the long run.

5. Advice for Business & The Future

• SMEs, don't rely on AI as a magic pill: AI is a tool, not a business model. Successful companies solve real customer problems first, then apply AI to help—not as the sole reason for their existence.
• From Models to Workflows: In the future, we will talk less about "models" and more about "Workflows" and "Agentic Patterns" (AI systems that can execute multi-step tasks).
• Partner globally, but control the core: You can use the best tools from around the world, but you must always keep your "strategic capabilities" and "critical data" under your own control.

Conclusion

The key to success isn't building one single, perfect model. It is about building adaptable systems. Start simple, prove the business value (ROI) quickly, and scale through partnerships—while never losing sight of the foundation: Safety and the Data that is the heart of your organization.

Watch the full content at https://www.youtube.com/watch?v=bJBQpjib8Sg