# Building Asia’s AI Hub: Open Model Infrastructure, Open Models, and AI Deployment at Scale

## Abstract

Key takeaways Understand how technology infrastructure, models, and applications work together to create a successful end-to-end AI ecosystem in Asia. Learn to apply open innovation principles to accelerate responsible AI adoption across industries. Discover how Singapore can strengthen its role as a regional AI hub and bridge between global AI ecosystems. Understand how to balance openness with trust , including governance, safety, and security considerations for AI systems. Learn to scale AI from pilot projects to production while developing local talent and building robust AI innovation ecosystems.

## AI Summary

_No AI summary available._

## Transcript

**[00:00:00 – 00:00:04]** Thank you, my esteemed panelists, for being with us here today.

**[00:00:04 – 00:00:06]** What an exciting day.

**[00:00:06 – 00:00:09]** AI is having a moment, but I think Singapore

**[00:00:09 – 00:00:13]** is having a bigger moment.

**[00:00:13 – 00:00:19]** Back home in Silicon Valley, there is an unofficial motto, which

**[00:00:19 – 00:00:24]** goes like, move fast, break things.

**[00:00:24 – 00:00:29]** But from looking from outside at Singapore, it seems to

**[00:00:29 – 00:00:34]** be more like move fast, make sure things work, and make sure

**[00:00:34 – 00:00:37]** things work at nation scale, right?

**[00:00:37 – 00:00:43]** So we definitely want to learn about how is Singapore doing this.

**[00:00:43 – 00:00:49]** And AI definitely is, you know, game changer, it's not

**[00:00:49 – 00:00:53]** about, the question is no longer what can AI do, but the question

**[00:00:53 – 00:00:59]** is about how are we transforming and translating the AI capabilities

**[00:00:59 – 00:01:01]** into meaningful outcomes.

**[00:01:01 – 00:01:06]** And what better nation to answer that question than Singapore,

**[00:01:06 – 00:01:08]** and especially with you all?

**[00:01:08 – 00:01:15]** Already we kind of heard from HTX, we heard some really

**[00:01:15 – 00:01:18]** great stuff from industry as well as from government.

**[00:01:18 – 00:01:23]** And so let's get right into it, because we want to see how

**[00:01:23 – 00:01:30]** this balance between innovation, but also trust and execution,

**[00:01:30 – 00:01:33]** comes about in practice, right?

**[00:01:33 – 00:01:36]** So together, let's explore this.

**[00:01:36 – 00:01:43]** And how has Singapore's journey been into becoming an AI hub?

**[00:01:43 – 00:01:45]** So let's begin with this.

**[00:01:45 – 00:01:50]** So I'd like to ask each of our panelists that, as a bridge

**[00:01:50 – 00:01:56]** between the East and the West that you have come to become, what

**[00:01:56 – 00:02:02]** are the most important capabilities that Singapore must build or

**[00:02:02 – 00:02:08]** must strengthen, and to stay Asia's AI hub, and where can it create

**[00:02:08 – 00:02:12]** the greatest competitive advantage?

**[00:02:12 – 00:02:15]** So I think Singapore is not just the bridge between the

**[00:02:15 – 00:02:16]** East and the West.

**[00:02:16 – 00:02:19]** It happens to be located as a hub of this very

**[00:02:20 – 00:02:23]** multicultural, very diverse region called Southeast Asia.

**[00:02:24 – 00:02:27]** So living with diverse views, multi-cultures,

**[00:02:27 – 00:02:29]** quite natural for us.

**[00:02:29 – 00:02:32]** Probably not so true for many other people.

**[00:02:32 – 00:02:34]** So that's on one hand.

**[00:02:34 – 00:02:39]** On the other hand, I think Singapore has invested quite

**[00:02:39 – 00:02:43]** early in the technology readiness, you know, the technology ecosystem.

**[00:02:43 – 00:02:46]** I think our smart nation is more than a decade now.

**[00:02:47 – 00:02:51]** AI Singapore has been around since 2017, way before ChatGPT, right,

**[00:02:51 – 00:02:55]** so I think in terms of readiness, we are quite competitive there.

**[00:02:55 – 00:02:58]** And the last bit is what you mentioned, trust, right?

**[00:02:58 – 00:03:03]** And I think we are a trusted hub in some areas, finance, and

**[00:03:03 – 00:03:05]** more and more AI safety too, right?

**[00:03:05 – 00:03:08]** Thanks to the hard work of some people.

**[00:03:08 – 00:03:13]** Now, most of the time, these three elements, they live

**[00:03:13 – 00:03:15]** in their own lanes.

**[00:03:15 – 00:03:20]** Now, in a world where AI is becoming more polarized, East

**[00:03:20 – 00:03:24]** versus West, and people don't trust AI, I think Singapore

**[00:03:24 – 00:03:26]** has something to offer if we become a thought leader

**[00:03:26 – 00:03:28]** in the intersection of these three.

**[00:03:28 – 00:03:33]** And AI is not just capable, but also inclusive and trustworthy.

**[00:03:33 – 00:03:38]** So that's something that's coming from somebody who built

**[00:03:38 – 00:03:43]** the Sea Lion LLMs and also working from time to time with

**[00:03:43 – 00:03:45]** some other partners on AI safety.

**[00:03:45 – 00:03:46]** So we have done something.

**[00:03:46 – 00:03:48]** Yeah, those are some really great points.

**[00:03:48 – 00:03:51]** And so I'd like to ask you the same question about, you

**[00:03:51 – 00:03:54]** know, from your lens in GovTech.

**[00:03:54 – 00:03:56]** Thanks for inviting me here.

**[00:03:56 – 00:04:01]** So I think this is probably the third or fourth time somebody

**[00:04:01 – 00:04:04]** asked me this question in a short one week.

**[00:04:04 – 00:04:08]** I don't know whether it's something indicative or something that

**[00:04:08 – 00:04:11]** we need to build a lot more in.

**[00:04:11 – 00:04:16]** But the answer is, for me, there are two parts of this.

**[00:04:16 – 00:04:20]** First is about governance and evals.

**[00:04:20 – 00:04:24]** I say these two things is because governance, a lot of you have

**[00:04:24 – 00:04:28]** pointed out to me that Singapore is generally trusted and a neutral

**[00:04:28 – 00:04:32]** party in the world, and since we cannot compete on the number of

**[00:04:33 – 00:04:38]** chips we have or number of people that we have, we should be able to

**[00:04:38 – 00:04:42]** compete in the matter of governance and neutrality and so on.

**[00:04:42 – 00:04:47]** So I thought that is quite interesting perspective.

**[00:04:47 – 00:04:51]** And I think there is some grain of truth in that, but

**[00:04:51 – 00:04:55]** more importantly, I think because AI has come up to be

**[00:04:55 – 00:04:59]** something that can be unpredictable and can, might or might not

**[00:04:59 – 00:05:04]** work as opposed to what used to be how we deal with software.

**[00:05:04 – 00:05:09]** Software in the past used to be very deterministic, you know

**[00:05:09 – 00:05:12]** what you're going to get, at least if there are no bugs, you know

**[00:05:12 – 00:05:13]** exactly what you're going to get.

**[00:05:13 – 00:05:16]** But now it's... Maybe, right?

**[00:05:16 – 00:05:21]** So I do think that EVALS will probably be a very good move

**[00:05:21 – 00:05:26]** forward for us to determine whether the AI actually works

**[00:05:26 – 00:05:27]** as intended or not.

**[00:05:27 – 00:05:29]** So I always thought that these two things will

**[00:05:29 – 00:05:31]** probably go hand in hand.

**[00:05:31 – 00:05:35]** But beyond that, which is the second part I want to

**[00:05:35 – 00:05:39]** talk about, is for Singapore, I believe we do have some

**[00:05:40 – 00:05:47]** things that we are quite... good at, for example, we have

**[00:05:47 – 00:05:51]** great airlines, we have Singapore Airlines, we have a great airport.

**[00:05:51 – 00:05:56]** So if we combine these two things together and we put AI

**[00:05:56 – 00:05:58]** in it, which I'm very sure Changi Airport and Singapore Airlines

**[00:05:58 – 00:06:03]** are doing right now, I think we can enhance ourselves even greater.

**[00:06:03 – 00:06:07]** And that could be a real focal point for Singapore because these

**[00:06:07 – 00:06:08]** are the things that we're good at.

**[00:06:08 – 00:06:13]** Other things that we're good at, like your financial hub, you know,

**[00:06:13 – 00:06:17]** things like medical, healthcare and so on, those are possibly things

**[00:06:17 – 00:06:20]** that we can also try to be good at.

**[00:06:20 – 00:06:22]** I don't think we need to be good in everything because

**[00:06:22 – 00:06:23]** it's going to be impossible.

**[00:06:24 – 00:06:26]** But those things that we are really good at, if we can

**[00:06:26 – 00:06:29]** enhance it with AI, I think that will give us an advantage.

**[00:06:29 – 00:06:33]** Those are some great points on strengthening your capabilities

**[00:06:33 – 00:06:35]** that you've already built.

**[00:06:35 – 00:06:37]** And Denise, how about your views on this?

**[00:06:37 – 00:06:41]** Yeah, Shilpa, I'll add on to the perspective shared by Xiao Xiong.

**[00:06:41 – 00:06:44]** So I have lived in the US in Silicon Valley for the past

**[00:06:44 – 00:06:49]** 15 years, and one of the things I look forward to every year is

**[00:06:49 – 00:06:51]** coming home through Changi Airport.

**[00:06:51 – 00:06:54]** And you might think it's a very odd thing to say, right?

**[00:06:54 – 00:06:57]** Because who actually looks forward to going through immigration?

**[00:06:57 – 00:07:02]** But I have suffered at US airports, and it's such a

**[00:07:02 – 00:07:06]** joy to go through Singapore airport where everything just works.

**[00:07:06 – 00:07:09]** And these days, I don't even need to take out my passports anymore.

**[00:07:09 – 00:07:17]** And so to me, good AI is something that feels invisible to the user,

**[00:07:17 – 00:07:21]** and it just solves real problems for the user, and it just works.

**[00:07:21 – 00:07:28]** So the advantage that Singapore has as a hub, being able to scale

**[00:07:28 – 00:07:33]** AI safely and reliably and be able to bring together the technology,

**[00:07:33 – 00:07:39]** the talent and solve real problems that just make AI work.

**[00:07:39 – 00:07:42]** I absolutely agree with the efficiency of the airport.

**[00:07:42 – 00:07:45]** It took me literally 10 minutes from exiting the plane to

**[00:07:45 – 00:07:51]** my cab, and time is money, so thank you, Singapore.

**[00:07:52 – 00:07:57]** So, my view is, first of all, a consideration that Singapore

**[00:07:57 – 00:08:01]** has already shown in many other challenges that it's possible.

**[00:08:01 – 00:08:04]** You don't have actually to break things to make things

**[00:08:04 – 00:08:05]** work all the time.

**[00:08:05 – 00:08:10]** You can actually develop in a controlled environment where

**[00:08:10 – 00:08:15]** you break things down and you prioritize the safety,

**[00:08:15 – 00:08:19]** the governance, as has been said, so that you can, let's

**[00:08:19 – 00:08:23]** say, break smaller things before you break big things.

**[00:08:23 – 00:08:27]** And when it comes to the West and East coming together, I do think

**[00:08:27 – 00:08:34]** that Singapore is one of the ideal contexts, contextual environment

**[00:08:34 – 00:08:39]** to build those harnesses because it can bring the quality and

**[00:08:39 – 00:08:43]** the bleeding edge frontier of the West with the scale, the

**[00:08:43 – 00:08:48]** availability of talent, the drive, the understanding, the capabilities

**[00:08:48 – 00:08:54]** and the speed coming from the East, but in a smaller scale,

**[00:08:54 – 00:08:58]** which allows you to break small things before you break big things

**[00:08:58 – 00:09:01]** and put things into operations.

**[00:09:01 – 00:09:08]** The key aspect that will come more evident as things move

**[00:09:08 – 00:09:12]** from a transition phase to a really production phase, particularly

**[00:09:12 – 00:09:17]** when you talk about physical AI, will be how do we put together that

**[00:09:17 – 00:09:22]** operation fabric that allows things to be solid and scale, and I think

**[00:09:22 – 00:09:26]** that is ideal space for Singapore to show the differentiating

**[00:09:26 – 00:09:28]** capabilities that we have here.

**[00:09:28 – 00:09:31]** That's an important point now that we are transitioning

**[00:09:31 – 00:09:36]** from generative to agentic and now to physical AI and robotics.

**[00:09:36 – 00:09:42]** So that brings us to the topic of scaling AI, right, as countries

**[00:09:42 – 00:09:47]** and companies are moving from experimentation into execution.

**[00:09:47 – 00:09:52]** So what does it take to scale AI from pilots into actual

**[00:09:52 – 00:09:57]** production scale systems, and how do we measure these outcomes?

**[00:09:58 – 00:10:01]** So maybe we start with you, Denise.

**[00:10:01 – 00:10:02]** Thanks, Shilpa.

**[00:10:02 – 00:10:07]** So I see two reasons why pilots don't scale.

**[00:10:07 – 00:10:12]** So first is many pilots are treated as a tech project.

**[00:10:12 – 00:10:16]** So somebody sees an interesting capability, and then they build it.

**[00:10:16 – 00:10:20]** And the pilots work, but it becomes a solution looking for a problem.

**[00:10:20 – 00:10:23]** So it's more like a vitamin and a painkiller.

**[00:10:23 – 00:10:25]** You don't really need it. It's just a good-to-have.

**[00:10:25 – 00:10:28]** And the second reason is ownership.

**[00:10:28 – 00:10:33]** So, you know, like the tech team may build the solution

**[00:10:33 – 00:10:38]** and the business team may get consulted, but after the pilots

**[00:10:38 – 00:10:40]** are done, like, there's no owner.

**[00:10:40 – 00:10:44]** Nobody really sees it as something important to take it further.

**[00:10:44 – 00:10:49]** And so in HTX, we actually address both sides of this problem.

**[00:10:49 – 00:10:53]** So we start with a real problem that is aligned to our mission,

**[00:10:53 – 00:10:56]** and then we have product squads who then work with our home

**[00:10:56 – 00:11:01]** team departments to actually turn these problems into solutions.

**[00:11:01 – 00:11:05]** So the user is involved from day one, and then when the

**[00:11:05 – 00:11:08]** solution is built, they actually get to take it forward.

**[00:11:08 – 00:11:11]** And I will share an example where we have seen success,

**[00:11:11 – 00:11:15]** and that is the R-COP, or the Report Lodging Co-Pilot,

**[00:11:15 – 00:11:19]** where we have these kiosks in police stations, and the members of

**[00:11:19 – 00:11:25]** the public can lodge police reports using AI through these kiosks.

**[00:11:25 – 00:11:29]** And in fact, some time back, I was interviewing a candidate

**[00:11:29 – 00:11:33]** for a job, and the candidate started by saying he really

**[00:11:33 – 00:11:37]** loves R Corp, and the reason was, like, his father got scammed.

**[00:11:37 – 00:11:41]** Actually, is very very unfortunate, but that R Corp actually made

**[00:11:41 – 00:11:46]** his experience a little bit less painful and much smoother.

**[00:11:46 – 00:11:49]** So, I mean, can you imagine getting a positive customer

**[00:11:49 – 00:11:51]** satisfaction feedback in the middle of an interview?

**[00:11:51 – 00:11:53]** I thought that was pretty amazing.

**[00:11:53 – 00:11:58]** But that's what I felt really the test of good AI in production,

**[00:11:58 – 00:12:04]** that is something that you solve real problems that actually

**[00:12:04 – 00:12:08]** have real owners and make the lives of people better.

**[00:12:09 – 00:12:11]** That's excellent.

**[00:12:11 – 00:12:14]** I'll ask you, Sashang, from a GovTech standpoint,

**[00:12:14 – 00:12:17]** what should governments do, and even organizations

**[00:12:17 – 00:12:21]** and enterprises, to prove that AI deployments are successful?

**[00:12:21 – 00:12:22]** How would you measure it?

**[00:12:22 – 00:12:29]** I think measuring successes for many software systems

**[00:12:29 – 00:12:32]** is one of the hardest things to do.

**[00:12:32 – 00:12:37]** At a very high level, the natural thing to say is, as long as you

**[00:12:37 – 00:12:40]** deliver value, it will work, right?

**[00:12:40 – 00:12:44]** But that's a very motherhood statement because what does

**[00:12:44 – 00:12:46]** it mean by delivering value?

**[00:12:46 – 00:12:51]** So coming up with metrics to actually measure how How

**[00:12:51 – 00:12:56]** the value is being calculated, and not only being calculated

**[00:12:56 – 00:13:00]** or being measured, but must be agreed by all the stakeholders,

**[00:13:00 – 00:13:02]** I think that's very critical.

**[00:13:02 – 00:13:06]** That is something that in GovTech we actually have to deal

**[00:13:06 – 00:13:12]** with very frequently because if we develop a system and we deploy

**[00:13:13 – 00:13:17]** it to one user, one government agency, the value that they derive

**[00:13:17 – 00:13:21]** there might not be equivalent to what is being derived in another.

**[00:13:21 – 00:13:24]** And the cost of delivering one could be quite different

**[00:13:24 – 00:13:26]** from the other as well.

**[00:13:26 – 00:13:32]** So I think the way to go about doing this is really to establish

**[00:13:32 – 00:13:36]** the clear metrics of how you want to measure the value, and what

**[00:13:36 – 00:13:38]** it means by delivering the value.

**[00:13:38 – 00:13:42]** And the value sometimes doesn't necessarily mean just optimising

**[00:13:42 – 00:13:45]** the cost or reducing the cost.

**[00:13:45 – 00:13:48]** And as government agencies, our aim is not to generate revenue.

**[00:13:48 – 00:13:51]** It's really how much value we provide for the citizen.

**[00:13:51 – 00:13:55]** That, generally, is pretty hard to measure.

**[00:13:55 – 00:13:57]** The standard way of doing it is probably

**[00:13:57 – 00:14:00]** like customer satisfaction.

**[00:14:00 – 00:14:05]** But then again, there's this of... there's this issue about

**[00:14:05 – 00:14:09]** customer satisfaction surveys are often very subjective,

**[00:14:10 – 00:14:12]** and it depends on the user base, and it depends on the population

**[00:14:12 – 00:14:14]** that you survey as well, right?

**[00:14:14 – 00:14:16]** If you survey people who love your product, then you get

**[00:14:16 – 00:14:17]** very high scores.

**[00:14:17 – 00:14:21]** And generally, product teams end up doing that because,

**[00:14:21 – 00:14:25]** you know... so we do need to balance that out a lot as well.

**[00:14:25 – 00:14:29]** We need to be able to pick the correct user bases to really

**[00:14:29 – 00:14:33]** tap on the... whether the products that we roll out are actually

**[00:14:33 – 00:14:37]** producing value and whether the users are really using it, not just

**[00:14:37 – 00:14:42]** by surveys, but also by measurement of adoption, measurement of

**[00:14:42 – 00:14:46]** usage, whether hard facts that we can actually measure, and

**[00:14:46 – 00:14:51]** then compare that with the actual customer satisfaction surveys,

**[00:14:51 – 00:14:52]** and then do that kind of matching.

**[00:14:53 – 00:14:56]** So it is a bit of a science and a bit of art as well.

**[00:14:56 – 00:14:58]** It's one of the tougher jobs that we have to do, despite,

**[00:14:58 – 00:15:03]** you know, from a very logical point of view, it seems like quite easily

**[00:15:03 – 00:15:05]** derived, but it's actually not.

**[00:15:05 – 00:15:06]** Good points.

**[00:15:06 – 00:15:13]** Maybe I can echo the challenge in trying to go and measure everything

**[00:15:13 – 00:15:18]** when I actually need to align to the business line as well, right?

**[00:15:18 – 00:15:20]** It's like going to the gym.

**[00:15:20 – 00:15:24]** Things change every day, but you don't see the change every day.

**[00:15:24 – 00:15:26]** You have to wait for months and months and months.

**[00:15:26 – 00:15:30]** And when it comes to AI, it is not a single workflow, it is

**[00:15:30 – 00:15:35]** not a single agent that you deploy, it is not a single application

**[00:15:35 – 00:15:38]** that you have developed.

**[00:15:38 – 00:15:42]** For this type of solution, we want to... not just replace

**[00:15:42 – 00:15:46]** what humans are doing with some agentic workflow.

**[00:15:46 – 00:15:49]** You have to inject a little bit of the transformation.

**[00:15:49 – 00:15:53]** You want to look for perhaps a different business model.

**[00:15:53 – 00:15:58]** There's a lot of conversation around the successful companies

**[00:15:58 – 00:16:01]** versus the non-successful company will be those that

**[00:16:01 – 00:16:05]** are able to embrace AI and transform themselves.

**[00:16:05 – 00:16:08]** Now, we don't know exactly how, so that's why it's hard to measure.

**[00:16:08 – 00:16:11]** But you will see when you when it happens because you

**[00:16:11 – 00:16:17]** will see all the metrics kind of reflecting the change in

**[00:16:17 – 00:16:19]** the transformation of the company.

**[00:16:19 – 00:16:21]** I love that analogy, the gym analogy.

**[00:16:21 – 00:16:24]** It's a continuous process.

**[00:16:24 – 00:16:28]** And the results start showing with consistency and improvements.

**[00:16:28 – 00:16:33]** AI Singapore, Dr. William, how, you know, as AI adoption

**[00:16:33 – 00:16:39]** scales and expands, where do you see the most urgent talent gap?

**[00:16:39 – 00:16:43]** Because, I mean, you as the ASG entity are also looking

**[00:16:43 – 00:16:45]** at research, but also talent.

**[00:16:45 – 00:16:50]** And so what, according to you, is the most urgent talent gap in

**[00:16:50 – 00:16:52]** Singapore, but also in the region?

**[00:16:52 – 00:16:55]** Thanks for the question. So for the past few months, I

**[00:16:55 – 00:17:00]** have been involved in a healthcare AI project, and I've been

**[00:17:00 – 00:17:05]** reflecting, if we have more AI engineers per se in the

**[00:17:05 – 00:17:10]** team, will we make more progress, you know, the progress faster?

**[00:17:10 – 00:17:12]** The answer is no, right?

**[00:17:12 – 00:17:16]** But if we have AI engineers who have deep understanding

**[00:17:16 – 00:17:19]** of the healthcare domain, that will make a lot of difference.

**[00:17:19 – 00:17:23]** So I'm a big subscriber for this notion of a bilingual

**[00:17:23 – 00:17:27]** AI talent, but I want to add one deeper layer to this bilingual

**[00:17:28 – 00:17:31]** talent from my observation.

**[00:17:31 – 00:17:36]** I think it's important for these engineers who know the

**[00:17:36 – 00:17:39]** domain very well to know where to unlock the values.

**[00:17:39 – 00:17:42]** I think that's very, very important, using AI.

**[00:17:42 – 00:17:45]** But I observe that there's these other skills, I don't know

**[00:17:45 – 00:17:50]** what's the name of it, maybe domain safety or domain governance, right?

**[00:17:50 – 00:17:55]** Essentially, this instinct to know where to put this

**[00:17:55 – 00:18:01]** human intervention in some parts of the agentic workflow, right, or

**[00:18:01 – 00:18:07]** a fallback plan in case the agent fail, their agent fails, right.

**[00:18:07 – 00:18:09]** I think it requires some skills.

**[00:18:09 – 00:18:11]** You need to have the technical expertise and the deep domain

**[00:18:12 – 00:18:17]** understanding to kind of balance the productivity value capture

**[00:18:18 – 00:18:19]** versus the safety, right?

**[00:18:19 – 00:18:23]** So I think without these skills, you know, I think it's difficult

**[00:18:23 – 00:18:25]** to move from POCs to deployment.

**[00:18:25 – 00:18:29]** Yeah, so domain expertise is very important, and having

**[00:18:29 – 00:18:34]** the right blend of domain experts as well as the deployment experts.

**[00:18:34 – 00:18:35]** That's a great point.

**[00:18:35 – 00:18:40]** And that brings us kind of to the point about openness

**[00:18:40 – 00:18:45]** and trust and innovation and how innovation and speed and

**[00:18:45 – 00:18:51]** agility, as well as safe and trusted AI, can work hand in hand.

**[00:18:51 – 00:18:55]** So obviously, open models broaden the access and they

**[00:18:55 – 00:18:57]** accelerate innovation, you get to.

**[00:18:57 – 00:19:01]** We have the best of all the worlds, best AI is collaborative,

**[00:19:01 – 00:19:05]** right, and inclusive AI is good AI.

**[00:19:05 – 00:19:09]** But what does it take to build that safe and responsible deployment?

**[00:19:09 – 00:19:13]** So, question for you, Sau Sheong, how does openness

**[00:19:13 – 00:19:14]** and trust work together?

**[00:19:14 – 00:19:18]** And what measures are you taking to make sure that open

**[00:19:18 – 00:19:25]** models are trustworthy and suitable for deployments across nations?

**[00:19:25 – 00:19:30]** Thanks. Again, very often asked questions, especially

**[00:19:30 – 00:19:33]** in panels to government people.

**[00:19:33 – 00:19:36]** In general, but for me, I look at it from a little bit

**[00:19:36 – 00:19:40]** more of a technical lens, and I will look at it from

**[00:19:40 – 00:19:46]** the point of view of guardrails and evalves, right, so guardrails

**[00:19:46 – 00:19:52]** really talked about how we can make sure that whatever input that we

**[00:19:52 – 00:19:56]** go in and the input that comes out are properly filtered and properly

**[00:19:56 – 00:20:02]** corrected such that not the wrong, or rather the undesirable

**[00:20:02 – 00:20:06]** data, comes in or goes out, and I think that is a pretty important

**[00:20:06 – 00:20:09]** part of what we want to do.

**[00:20:09 – 00:20:13]** The second part of it is really to assure whether whatever is being

**[00:20:13 – 00:20:17]** generated actually fits the purpose of what it's supposed to do.

**[00:20:17 – 00:20:20]** And these two things, I think, work hand in hand to try to

**[00:20:20 – 00:20:25]** make sure that whatever that's been produced actually is,

**[00:20:25 – 00:20:29]** number one, not harmful, but at the same time is accurate

**[00:20:29 – 00:20:32]** enough because at the end of the day, whatever data that's been

**[00:20:32 – 00:20:37]** generated needs to actually fulfill a purpose and produce value.

**[00:20:37 – 00:20:38]** That's right.

**[00:20:38 – 00:20:42]** Yeah, and the same thing goes with even the, especially

**[00:20:42 – 00:20:43]** in the physical AI era.

**[00:20:43 – 00:20:48]** So I'd like to hear from you, Dr Giuseppe, like how are you

**[00:20:48 – 00:20:52]** looking at, you know, especially now in physical AI, how are

**[00:20:52 – 00:20:57]** you balancing that open innovation and yet innovating faster

**[00:20:57 – 00:21:00]** by putting in all the guardrails?

**[00:21:00 – 00:21:06]** So we we have been doing innovation for now a number of years,

**[00:21:06 – 00:21:10]** and when it stays, let's say, in the digital realm, we have

**[00:21:10 – 00:21:15]** developed a understanding or a process that allows us to work with

**[00:21:15 – 00:21:19]** our customers and, you know, define those pocs with the, you know, the

**[00:21:19 – 00:21:24]** value and try to have a direction and and then iterate, iterate.

**[00:21:24 – 00:21:28]** It can be a use case that they have in mind or perhaps testing

**[00:21:28 – 00:21:32]** a technology, a lot of it in the last couple of years, we just want

**[00:21:32 – 00:21:37]** to test the technology or maybe change a workflow or an operations.

**[00:21:37 – 00:21:42]** When we bring it into the physical AI realm, the level

**[00:21:42 – 00:21:48]** of complexity, this raises a lot, particularly if we go back to

**[00:21:48 – 00:21:56]** the Singapore example and the areas that we are more focusing on NCS,

**[00:21:56 – 00:22:03]** which is more the areas related to use cases that involve the public

**[00:22:03 – 00:22:07]** or the employees or maybe public spaces or physical operations.

**[00:22:08 – 00:22:09]** Those situations are more complex.

**[00:22:10 – 00:22:13]** And when we have higher complexity, your errors tend to multiply

**[00:22:13 – 00:22:17]** and the predictability already goes out the window.

**[00:22:17 – 00:22:23]** So we try to look at bringing together the best of breed

**[00:22:23 – 00:22:26]** in terms of capability across the technology.

**[00:22:26 – 00:22:31]** Our technology is hardly a bottleneck because right now

**[00:22:31 – 00:22:33]** even our technology is quite advanced compared to what we

**[00:22:33 – 00:22:37]** can do and what we can do safely.

**[00:22:37 – 00:22:41]** And then we try to understand the operations around it,

**[00:22:41 – 00:22:47]** break it down and solve small problems, try to iterate.

**[00:22:47 – 00:22:52]** For example, we try to work on use cases on physical AI

**[00:22:52 – 00:22:56]** where there is public engagement, where there is a conversation

**[00:22:56 – 00:23:01]** that may be the robots has to have with the public and how do

**[00:23:02 – 00:23:06]** I manage that conversation safely or maybe some actions, navigation

**[00:23:06 – 00:23:13]** or how do we build the, again, operating fabric to ensure that

**[00:23:13 – 00:23:18]** the mission can be completely safe, completed safely and then I can

**[00:23:18 – 00:23:25]** scale this to 10, 50, 100 robots in operations in a way that I can

**[00:23:25 – 00:23:31]** gain value and the functionality provides the right level of

**[00:23:31 – 00:23:38]** support to the clients where maybe the workforce needs to be extended

**[00:23:38 – 00:23:43]** because there's a requirement to have more and more resources

**[00:23:43 – 00:23:45]** to deliver these type of use cases.

**[00:23:45 – 00:23:49]** So that's the way that we look at the innovation part, particularly

**[00:23:49 – 00:23:51]** when it comes to physical AI.

**[00:23:51 – 00:23:53]** Yeah, and Denise, you are looking to deploy AI in these

**[00:23:53 – 00:23:55]** mission-critical environments.

**[00:23:55 – 00:23:59]** So, from your view, I mean, how are you looking at innovating

**[00:23:59 – 00:24:03]** faster while still keeping those safeguards?

**[00:24:03 – 00:24:06]** Yeah, so, Shobha, I'll add on to the points on governance

**[00:24:06 – 00:24:09]** and safety that Giuseppe and Sao Hsiao has brought up.

**[00:24:09 – 00:24:12]** And I'll share a personal example.

**[00:24:12 – 00:24:16]** So, I've lived in the US for the past 15 years.

**[00:24:16 – 00:24:22]** And I have a 12-year-old daughter, and in the US, I can't let her

**[00:24:22 – 00:24:26]** go anywhere by herself, so I have to pick her up and drop her off.

**[00:24:26 – 00:24:29]** It is not even safe to leave her alone in the house.

**[00:24:29 – 00:24:35]** So, when I returned to Singapore in January, my daughter loves it

**[00:24:35 – 00:24:38]** here because it feels so safe and there's good public transport, and

**[00:24:38 – 00:24:40]** she can go everywhere by herself.

**[00:24:40 – 00:24:43]** And these days, I don't even know where she is.

**[00:24:43 – 00:24:46]** I have to call her and tell her to come home, please.

**[00:24:46 – 00:24:50]** And why this is possible is because there's such a high

**[00:24:50 – 00:24:57]** level of trust in public safety in Singapore and at HTX, in the home

**[00:24:57 – 00:25:02]** team, that is super important for us to uphold this level of trust

**[00:25:02 – 00:25:04]** and reinforced it now with AI.

**[00:25:04 – 00:25:08]** And with AI agents coming into the picture now, it becomes

**[00:25:08 – 00:25:11]** even more urgent and important.

**[00:25:11 – 00:25:15]** And I've seen how openness can scale, because during

**[00:25:15 – 00:25:20]** my time in the US, I was working on the Android open ecosystem.

**[00:25:20 – 00:25:23]** It's an open ecosystem, but we're able to scale across

**[00:25:23 – 00:25:26]** 3 billion devices because we prioritise security.

**[00:25:27 – 00:25:32]** And at HTX, we have a governance playbook and we treat and

**[00:25:32 – 00:25:36]** check for risk before every AI solution is deployed.

**[00:25:36 – 00:25:41]** And so I think that governance is very important; it actually

**[00:25:41 – 00:25:44]** accelerates innovation; it doesn't stop it because when

**[00:25:44 – 00:25:47]** product teams know that there's guardrails in place, they

**[00:25:48 – 00:25:51]** actually have the confidence to deploy solutions faster.

**[00:25:51 – 00:25:54]** Yeah, I have to say congratulations to Singapore for achieving

**[00:25:54 – 00:25:58]** that level of safe spaces.

**[00:25:58 – 00:26:02]** As for me, I have to monitor my kids on Live 360 every 10

**[00:26:02 – 00:26:04]** minutes when they're out and about.

**[00:26:04 – 00:26:06]** But that's a different story.

**[00:26:06 – 00:26:11]** So, William, from an AISG perspective, and you have

**[00:26:11 – 00:26:15]** the Sea Lion model and other open source assets, how do

**[00:26:15 – 00:26:19]** you make these open source assets into commercially viable

**[00:26:19 – 00:26:21]** production level deployments?

**[00:26:21 – 00:26:24]** Are there any best practice examples that you can share?

**[00:26:24 – 00:26:29]** So, speaking from the experience of building

**[00:26:29 – 00:26:31]** the sea lion, an open model.

**[00:26:31 – 00:26:34]** A lot of people always, you know, many people think

**[00:26:34 – 00:26:38]** that the main outcomes of the sea lion development

**[00:26:38 – 00:26:40]** is the models themselves.

**[00:26:40 – 00:26:43]** But in reality, to get to the models, we have to build

**[00:26:43 – 00:26:47]** a lot of, like what you said, open assets, you know, to

**[00:26:47 – 00:26:49]** support the development of it.

**[00:26:49 – 00:26:54]** For example, evaluation, you know, I think because we want to make

**[00:26:54 – 00:26:57]** sure that the training of the model is going in the right direction,

**[00:26:57 – 00:27:00]** so very early we had to build.

**[00:27:00 – 00:27:03]** Because we are dealing with Southeast Asian languages,

**[00:27:03 – 00:27:08]** it's quite challenging, it's a mix of, you know, having community

**[00:27:08 – 00:27:12]** that basically provides the human evaluation and human judgment,

**[00:27:12 – 00:27:16]** but at the same time balancing it with building LLM as a judge.

**[00:27:16 – 00:27:20]** You know, in the right place to ensure that it's still scalable.

**[00:27:20 – 00:27:22]** Now this asset is important, right?

**[00:27:22 – 00:27:25]** And then, you know, we also have to build to some degree

**[00:27:25 – 00:27:29]** automated post-training pipeline because we have to keep up

**[00:27:29 – 00:27:35]** with open model development pacing, how to build AI that's

**[00:27:35 – 00:27:38]** aligned despite data scarcity.

**[00:27:38 – 00:27:42]** All these things are important for the sea lion, but they

**[00:27:42 – 00:27:44]** are also valuable.

**[00:27:44 – 00:27:48]** For other sectoral AI projects that we are currently involved

**[00:27:48 – 00:27:53]** in, in healthcare, and in finance, and so on.

**[00:27:53 – 00:27:56]** And all these sectoral AI transformation projects,

**[00:27:56 – 00:28:01]** they target production, so I think that's where

**[00:28:01 – 00:28:02]** we have been contributing.

**[00:28:02 – 00:28:06]** We have not, in the CLion team, we have not thought

**[00:28:06 – 00:28:10]** too much about commercialization, but if we ever get there, we will

**[00:28:10 – 00:28:12]** apply the same kind of thinking.

**[00:28:12 – 00:28:16]** Allow me to share a little bit about open.

**[00:28:16 – 00:28:22]** I think the fact that the CLion is an open model, it helps us a lot.

**[00:28:22 – 00:28:26]** It accelerates the capability building and also pushes us

**[00:28:26 – 00:28:29]** to a certain level of quality.

**[00:28:29 – 00:28:32]** What does open source bring to us?

**[00:28:32 – 00:28:36]** It brings us community of users who give us feedback.

**[00:28:36 – 00:28:41]** Importantly, it also gives us communities of collaborators

**[00:28:41 – 00:28:46]** who take whatever we build and then build something else on top

**[00:28:46 – 00:28:48]** of it and then share back with us.

**[00:28:48 – 00:28:51]** I think we've been focusing a lot on communities locally

**[00:28:51 – 00:28:56]** with GovTech, NCS, and also in the region.

**[00:28:56 – 00:29:01]** But I'm happy to share that recently I learned actually our

**[00:29:01 – 00:29:04]** influence goes beyond our region.

**[00:29:04 – 00:29:07]** I had the privilege to be in Lagos recently in Nigeria

**[00:29:07 – 00:29:09]** to attend Deep Learning Indaba.

**[00:29:09 – 00:29:13]** It's one of the most prestigious AI conferences in Pan-Africa.

**[00:29:14 – 00:29:19]** And in one of the follow-up meetings with the Uganda AI Lab,

**[00:29:19 – 00:29:23]** they shared with me that they took the Seahelm Evaluation Pipeline,

**[00:29:23 – 00:29:28]** they took the linguistic diagnostic that we built, And apply it and

**[00:29:28 – 00:29:32]** adapt it to Luganda, to Swahili, to other African languages.

**[00:29:32 – 00:29:38]** Now because African languages are a little bit more voice-oriented

**[00:29:38 – 00:29:42]** than text, they actually take the next step of adapting that

**[00:29:42 – 00:29:46]** linguistic diagnostic evaluation.

**[00:29:46 – 00:29:56]** I think this kind of thing can only happen if it is open source.

**[00:29:56 – 00:29:59]** And one last thing about open source that we have these

**[00:29:59 – 00:30:04]** regional communities, our collaboration with big tech like

**[00:30:04 – 00:30:08]** NVIDIA, Google, Alibaba, all these things become more meaningful.

**[00:30:08 – 00:30:11]** And I think we learn so much from this kind of collaboration,

**[00:30:11 – 00:30:14]** which we diffuse back to all the sector of AI transformation

**[00:30:14 – 00:30:15]** that's happening.

**[00:30:15 – 00:30:17]** I think the international collaborations that are enabled

**[00:30:17 – 00:30:25]** by open models is how AI becomes even more inclusive.

**[00:30:25 – 00:30:27]** And so we are going to actually switch to a

**[00:30:27 – 00:30:29]** rapid-fire round, right?

**[00:30:29 – 00:30:31]** So I'm going to ask you all a question.

**[00:30:31 – 00:30:33]** Anybody can choose to answer it.

**[00:30:33 – 00:30:37]** And it has to be a quick answer, super fast.

**[00:30:37 – 00:30:40]** It's like the TV shows have the rapid-fire rounds, right?

**[00:30:40 – 00:30:42]** And you'll get a prize, which is...

**[00:30:42 – 00:30:43]** So, no thinking tokens, you know?

**[00:30:43 – 00:30:45]** No thinking tokens, yeah.

**[00:30:45 – 00:30:47]** Don't consume too many tokens.

**[00:30:47 – 00:30:53]** And there'll be a prize, which will be the Kuelipa, the...

**[00:30:53 – 00:30:56]** Five-layer cake, the real cake outside at lunchtime.

**[00:30:56 – 00:31:00]** Yeah, yeah. Lapis, kwe lapis, kwe lapis, sorry.

**[00:31:00 – 00:31:04]** I'm trying to improve my Singlish.

**[00:31:04 – 00:31:08]** All right, so what is the one thing children born today

**[00:31:08 – 00:31:12]** may never have to do because of AI?

**[00:31:12 – 00:31:14]** Start from zero.

**[00:31:14 – 00:31:17]** Okay. Start from zero, from a blank page for anything.

**[00:31:17 – 00:31:20]** That's a very good point, all right.

**[00:31:20 – 00:31:22]** If you had unlimited compute for one year,

**[00:31:22 – 00:31:24]** what problem would you solve?

**[00:31:24 – 00:31:28]** I'll answer that. So if I've unlimited compute, I'll rent

**[00:31:28 – 00:31:31]** it out to everyone who needs GPU and I'll put a subscription

**[00:31:31 – 00:31:32]** plan with credit limits.

**[00:31:32 – 00:31:34]** But that's just joking.

**[00:31:34 – 00:31:38]** So I want to build like AI agents for every role in home

**[00:31:38 – 00:31:43]** team and the employees' lives become better and they can better

**[00:31:43 – 00:31:44]** achieve the mission for Singapore.

**[00:31:44 – 00:31:46]** That's beautiful.

**[00:31:46 – 00:31:51]** Alright, if Singapore created an AI ambassador, what would

**[00:31:51 – 00:31:54]** its first assignment be?

**[00:31:54 – 00:31:56]** I guess that's me.

**[00:31:56 – 00:32:01]** I'm hungry now, so I guess I will want the AI ambassador

**[00:32:01 – 00:32:05]** to help me to figure out where's the next best place to eat.

**[00:32:05 – 00:32:09]** Okay, great. Maybe it's outside.

**[00:32:09 – 00:32:14]** All right, and bigger models or smaller domain-specific

**[00:32:14 – 00:32:14]** and local models?

**[00:32:14 – 00:32:19]** Which matters more for Southeast Asia?

**[00:32:19 – 00:32:23]** Based on Hugging Face download of Sea Lion, smaller models broadly,

**[00:32:23 – 00:32:26]** but we learned from Nemotron that actually to make capable

**[00:32:26 – 00:32:32]** smaller models we can benefit from a big model like Nemotron that

**[00:32:32 – 00:32:35]** allows some kind of distillation, but needs to understand Southeast

**[00:32:35 – 00:32:38]** Asia, so hin hin, maybe we should work together on that.