Brainstorm AI Singapore 2024: Public - Private Partnerships In Ai Governance
Jacqueline POH, Managing Director, Singapore Economic Development Board Su Shan TAN, Managing Director and Group Head, Institutional Banking, DBS Group Moderator: Ayesha KHANNA, Co-founder and CEO, Addo AI; Co-chair, Fortune Brainstorm AI Singapore
Category
🤖
TechTranscript
00:00Hello. Good morning, everyone.
00:03So, Jacqueline, Singapore wants to be the leading AI-powered economy in the world.
00:10That's a tall order for a small country.
00:14Tell us how you plan to do that.
00:16Thanks a lot, Ayesha.
00:18I'm not an expert in AI.
00:20A long time ago when I was running GovTech,
00:23I did a little bit of practical AI for citizens.
00:26So I've got a little experience on that
00:28and the kind of impact it has on productivity and jobs.
00:31But here in my role in the Economic Development Board,
00:34we work very closely with organizations
00:36like the Ministry of Digital Development and Information
00:39as well as IMDA, the Infocom Media Authority,
00:44and the National AI Office in the furtherance of our AI objectives.
00:49The reason I say that Singapore is more interested
00:51in being the most AI-powered economy in the world
00:54as opposed to the economy that has the most data centers
00:58with GPUs in them or the economy that has the most LLMs
01:03or foundational models or sovereign foundational models
01:07is because this is a choice that we've taken.
01:10We believe that AI has a huge transformative potential for the economy
01:15and we think that we've seen this before
01:17because when we started Smart Nation and we decided on a digital economy,
01:22we wanted to reach into every facet of the economy
01:25to be able to use digital to improve products
01:28and to improve productivity.
01:30That journey's taken more than 10, 15 years now
01:33and we're actually really seeing the fruits of that happen.
01:35The digital economy in Singapore is close to about 17, 20% of GDP now
01:42if you include not just the verticals and the companies in the ICT verticals
01:46but also the ones in the rest of the economy.
01:49That's really possible to do also with AI
01:52but it's going to take a lot of effort.
01:54It's going to take a lot of talent.
01:56It's going to take compute and most of all,
01:59it's going to take companies willing to change their business models,
02:03their operating processes, the way they engineer data
02:07and the way they provide experiences to their customers
02:11in order to actually make that happen.
02:14Like I say, we've done it before with digitalization.
02:17We think we can do it with AI but not easily.
02:2184% of Singapore companies say they can't really implement AI
02:29as effectively as they would like to
02:31simply because their data is found in too many silos.
02:35Many of them say they can't get the talent they need
02:38in order to be able to implement simply the kind of data engineering
02:43and the integration of solutions
02:45from what is actually still a very evolving AI stack
02:51and it's evolving all over the world.
02:53These are the kind of questions that we work with various companies
02:57to make happen.
02:59We work with two types of companies.
03:01The ones that are at the edge,
03:03in other words, the ones that are most ambitious
03:05in terms of their implementation of AI
03:08and for that, we're looking for 100 companies
03:11who want to put up AI centers of excellence in Singapore
03:15and these include companies that are born on the internet like Grab.
03:20They're in companies like have huge amounts of data
03:23and operations processes like the Changi Airport
03:26or the Port of Singapore
03:28and they're ones in the sectors that we think
03:31are the most promising sectors for AI to flourish
03:35because they have the data
03:37and because they have the use cases and problem statements
03:40and these include sectors such as financial services.
03:44We work with companies like DBS and other companies
03:48like Prudential, for example,
03:50to actually implement AI in their various processes
03:55including fraud detection and customer service.
03:58There are companies in healthcare.
04:00We think healthcare has a huge potential
04:02and there's such a huge range of implementation
04:06and use cases in healthcare.
04:08We're doing AI for drug discovery on the one hand
04:11but we're also looking at AI in operations processes
04:14in pharmaceutical manufacturing plants
04:17and then there's AI in the healthcare delivery system,
04:21literally the hospitals, the clinics,
04:24and really how AI is delivered to each patient
04:27and there are other areas as well.
04:29For example, advanced manufacturing.
04:32Not many people realize
04:34but Singapore is not just a services hub.
04:37Singapore is not just a business hub.
04:39Singapore has 20% of its GDP in manufacturing.
04:44We make stuff and that means that there's a lot of potential
04:48for what we call AI in industrial,
04:52sort of industry 4.0 processes,
04:55industrial AI to actually make manufacturing more productive
05:00and then finally there's a whole range of professional services
05:04which is done in Singapore, legal accounting and so on,
05:07for which I think AI makes a huge amount of sense
05:10to be implemented in their work streams.
05:14So this together with supply chain logistics
05:18presents a huge opportunity for Singapore
05:21to be the world's most AI-powered nation.
05:25I love it.
05:26So Jacqueline just laid out our vision and the use case
05:31and she said, Sushant, can you give me some real examples
05:36of successes and failures on the ground?
05:39Is this possible, what Jacqueline is saying?
05:42So, by the way, if these two ladies tell you
05:45they're not AI experts, don't believe them.
05:47But like everyone in the room
05:49and frankly anyone operating in business,
05:52we're all learning as we go.
05:55And so my perspective is whether it's the customer
05:58or the employee, the key is to encourage experimentation,
06:05allow failure, get the feedback,
06:09and then keep improving and learning and changing
06:12and saying, oh my God, that worked, that didn't work,
06:14and being brutally honest with ourselves
06:16and allowing leaders to say, oops, that was a fail, right?
06:21And then us doing the post-mortem and diagnosing
06:24why one model worked and one didn't work.
06:27So I think that the beauty of AI
06:29is the ability to scale experimentation
06:32and it then behoves all of us businesses to Jack's point
06:36to make sure our data is organized.
06:39And the beauty of Gen AI is, guess what,
06:41it helps with unstructured and structured data
06:43so you can do both.
06:45And for us at DBS,
06:50I'll share both the internal and the external perspective.
06:53So the internal perspective was how do you move
06:5530,000 to 40,000 employees into this new way of working?
06:59Not easy.
07:01Well, you have to behave like a startup,
07:03but you're really a very mature company.
07:05So how do you do that?
07:07So one of the things we did which really worked
07:09was we used gamification.
07:11We used video games, right?
07:13So in 2020, we did this thing called DeepRacer with AWS,
07:17and that was just a,
07:19you can teach your Formula One virtual car
07:23how to drive with machine learning.
07:26So you don't need to code.
07:28You just need to learn how to tell it to go right,
07:30to go left, see a wall, crash,
07:32learn from that wall that's there,
07:34turn right when you see that wall in future.
07:36And so I was one of, like, I don't know,
07:395,000 employees that learned, you know,
07:42to build my own fast car using machine learning.
07:45And that then made me understand,
07:47ah, you know, you can cut and paste code
07:50to make your model better and faster,
07:52and you can do this really quickly.
07:54So that was for us an ah-ha moment.
07:57Then, you know, being able to bring all the customer data,
08:00we got 15,000 customer points for certain use cases.
08:03We've rolled it out to 3.5 million customers
08:06in the consumer banking side.
08:08And then refining the model because, you know,
08:11if you've been doing the same job for a long time,
08:13you think you know it all,
08:15but the data might tell you something else, right?
08:17So one of the use cases I like to cite
08:20is, you know, you all know that the FX rates
08:24and interest rates have been very volatile.
08:26And time and time again, your bankers will come and say,
08:29we've got to cut, you know, you've got to discount this
08:32or cut this because some startups come up
08:34with a lower price or a better price or whatever.
08:38And, okay, but let's experiment.
08:41Let's try some use cases where the prices are lower.
08:44Let's try some use cases where the message is different.
08:47Sometimes it's all about the how, not the who and the what.
08:51The who and the what, you know, but the how.
08:53How do you message the offer?
08:57The offer could be, hey, we've got a great click here.
09:00The offer could be, oh, I've missed you,
09:03haven't seen you in a while.
09:05Come back and do this trade for us
09:07and we'll give you something, right?
09:09Hey, guess which one worked better.
09:12The I miss you message actually worked better.
09:16So, you know, so we all have our embedded biases,
09:20unconscious bias.
09:22But being able to have these AI machine learning models
09:25then tell you otherwise can be quite a humbling experience
09:29for many of us.
09:31But then we learn, we go, okay, which is why now
09:33we manage by dashboards.
09:35Every meeting we have internally now, we have a live dashboard.
09:39And so if the manager says, no, we've got a discount,
09:42okay, so let's have a look at what the feedback loop tells us.
09:46So the discipline that we now have is you must always have
09:50a feedback loop, right?
09:52So then if you're a product owner and we now organize
09:55by horizontals, right, horizontal organization.
09:58So in every customer, every customer journey,
10:02the horizontal organization comes together and everyone
10:05is incentivized by the same outcomes.
10:07We want great customer experience.
10:09We want them to give us a five-star click.
10:11We want, you know, successes.
10:13We want new clients to come on, new to product,
10:15new to bank, et cetera.
10:17And if everyone is incentivized for the common outcomes
10:19and everyone owns the failures together, right,
10:22so that's a real cultural mindset shift.
10:24And then when you manage by dashboards, it's pretty instant
10:27and the data doesn't lie.
10:29So then you can't say, okay, look, let's look at the data,
10:31what the customer is telling us, what the feedback loop tells us,
10:34and keep refining and refining and refining and learning
10:37as you get along.
10:39So it's a journey, and it is a mindset,
10:43cultural mindset shift.
10:45But I think you can, the human brain is so agile, right?
10:49So you can change an old hag like me to become, you know,
10:53more of a machine-learning, AI-driven individual.
10:56You can change anyone.
10:58And what's really inspired me is we have got people
11:01who've worked in the service center, call center for 30 years,
11:04people who used to be bank tellers for 30 years,
11:06and they are the ones helping us with the model-making.
11:09And that, to me, is what it's all about.
11:12It's getting both your stakeholders,
11:15your internal stakeholders, and your external stakeholders,
11:17because you need them to tell you, hey, i like this,
11:19or i don't like this.
11:21I often say it's a fine line between being relevant
11:23and being creepy, right?
11:25And so being, you know, bringing these stakeholders together
11:28and what they see works for them so that you can
11:32hyper-personalize to a point of n equals 1.
11:35I think, Sushant, listening to you, it's so reassuring
11:38that real results and real implementation.
11:41Because if I compare this to the media and the news,
11:44there's so much hype, Jacqueline.
11:46Like, on the one hand, you have DBS really using it
11:49in all these different ways and admitting to experimentation
11:52and failures and successes.
11:54But then there's multi-billion-dollar valuations.
11:57What do you think, Jacqueline, from Singapore's perspective,
12:00is this real? Is this hype?
12:03When are we going to see this big bang of the singularity?
12:07I'm not sure that we are definitely moving along
12:10the direction of the singularity.
12:12But I'll tell you, it's a singular moment
12:15of hyper-inflated expectations for AI.
12:19There was an article in today's Financial Times,
12:22which I found rather interesting, which stated
12:25that the real winners of the AI boom to date
12:29have been the management consultants.
12:32That's true.
12:35This is not...
12:37That was a very salient point for me.
12:41So, of course, I took that article
12:44and I forwarded it to all the management consultants I knew
12:48who were heavily involved in AI consulting in Singapore.
12:52And they said, nay, that is not true.
12:55It is the infrastructure guys.
12:58It's actually NVIDIA.
13:01Our ROI is just a fraction of what they're making out of this.
13:06And I think that both of these things are true.
13:10I think we are in the middle of a hype cycle.
13:14Most of you who understand the Gartner hype cycle matrix
13:18will probably feel it and know that we're in the middle of a hype cycle,
13:22which doesn't mean that it isn't real.
13:25It simply means that some expectations are inflated
13:29and you have to manage to keep the core of the value
13:34throughout the whole process
13:37until you get through the trough of dissolution,
13:39until you get to a plateau of productivity
13:42and still maintain those capabilities.
13:45It's important because we've seen this before,
13:48not just in digital, but in AI itself.
13:51It's gone through several of these cycles.
13:54But every single time the technology has improved,
13:57it's come back at a higher level.
13:59And the economies that have preserved those capabilities have done well.
14:03I was recently in Canada, Toronto, Montreal, and so on,
14:08and meeting some of the AI experts in Canada.
14:11Canada has really maintained a lot of the institutes
14:14and deep tech capabilities in AI.
14:17And they've also produced some really great companies like Coher,
14:20who've actually emerged out of this,
14:23and disciples, friends of Yoshua Bengio,
14:26or like Yann LeCun, or Geoff Hinton.
14:29And I wondered why they were very successful
14:33in managing through these periods of time.
14:36And I realized they are called AI winters.
14:39And some places are better at working through winter.
14:44Places with a lot of winter.
14:47I think it's the ability to weather a winter.
14:52And it's the ability to have a hibernation mode.
14:57And to keep your capabilities warm during a period of hibernation.
15:02I'm not saying that a winter is going to be upon us.
15:05I'm saying that for a lot of deep tech subjects,
15:08it is important to know when to go out and get funding,
15:12when to go into hibernation,
15:14and when to preserve fundamental capabilities.
15:17For the current cycle, we have, apparently,
15:211,100 AI startups in Singapore.
15:24Some of them are very advanced.
15:27Some of them I'm not sure are AI startups.
15:30So discernment is going to be of great value
15:34in the whole part of this process.
15:37Even in terms of where the infrastructure boom
15:41is taking place for AI with AI data centers,
15:44discernment is also needed to figure out
15:47how well-utilized those data centers are.
15:49Who is utilizing those data centers?
15:51How are you managing power and resources and water?
15:55What part is more sustainable?
15:58Which parts are not more sustainable?
16:01Which is why I go back to our fundamental thesis in Singapore
16:05about an AI-driven economy.
16:07Because until the moment that businesses find that they can
16:11Get an ROI out of using AI in order to improve their
16:17Productivity or to create new product, as DBS is doing,
16:22Or SIA, which is another big one that's using this now,
16:26Or, you know, any of the other companies.
16:29Even in our own stable, we worked with GlaxoSmithKline
16:33To find a way to use the Google and the Google Ignite program,
16:38The Google AI trailblazers program, to actually bring
16:41Together these companies to actually find value out of an
16:45AI implementation in the pharma manufacturing.
16:48And we really, like, saved 5,000 man-hours and so on.
16:52So until these things are really scaling in companies, it's hard
16:57To see where, you know, the hype ends and the reality of that
17:02Productivity starts to take place.
17:05It may be some have said that, you know, in some senses we're
17:10At the edge of a revolution where you can only see the deep
17:14Embedding of the technology within the next 5 to 10 years.
17:17I think you can see, you can measure the outcomes, at least
17:21For our industry, right?
17:23And the outcomes that you measure and you quantify will be
17:27Either loading factors, right, per human.
17:30Now, how many clients can you cover?
17:32It can be the customer number of transactions.
17:36It can be the time to market. And how fast a service call is
17:41Resolved, you know, and whether you can, it's one call
17:45Resolution, that's also a good measurement.
17:48So i think there are a lot of quantifiable, certainly different
17:52Industries will have different quantifiable matrix, but you can
17:56Measure them and then refine your models to make them more
17:59Efficient.
18:01And building on that, like the discernment that you need on
18:05The quantity of ai models, like what is your approach and your
18:09Experience at dbs in terms of should you just do a lot of
18:13Different ones and how do you measure that?
18:16So it is, and that's a great question because i think no one
18:21At the moment, i don't know if everyone really knows the real
18:25Cost, jack, to your point, of all this generative ai and it's
18:29A lot of models, right? so there's the real capex
18:32Cost, the cost you pay to the big tech guys and some of the
18:36Start-ups, then there's the adjacencies around energy or
18:39Water and other things, and then the operating costs.
18:43And then the cost of training, you know, new target operating
18:46Models, et cetera. So there's a lot of adjacencies
18:49That we still don't know, but i think what we've been trying to
18:52Do is we've been trying to refine the way we measure the
18:56Successes of each model, as i said, by things like cost
18:59Income. Overall, you just have to look
19:02At your overall business cost income ratio.
19:04That's your top-down, you know, approach.
19:07Then you look at the individual productivity ratios.
19:10Then you look at the customer returns ratios.
19:13So you can look at different lenses, your own costs, customer
19:16Returns, you know, employee productivity, et cetera.
19:19You can dice it up, but in my experience, number one is get
19:23Everyone to drink the kool-aid. That's not always easy, but if
19:28You put the weapons in your colleagues' hands, so to speak,
19:32Let them own the model, let them own the feedback loop, let them
19:36Own the outcomes, and let them own both the failures and the
19:39Successes. You know, that's one important
19:42Thing, is ownership, accountability, and being fully
19:45Aligned through the whole horizontal organization for the
19:48Outcomes that you're generating, number one.
19:50Then get the customer feedback. To me, that's so important,
19:53Right? if the customer is not telling
19:56You what they like and don't like, it's not very useful.
19:59My third learning is, you know, you don't always have to be the
20:03First. Sometimes if you are the first,
20:06I mean, i will share one failure, i won't mention names,
20:09But 2014, we were the first bank, i think, in the world or
20:12In asia to use ai for wealth management.
20:14I was the wealth management head in those days.
20:17We ended up teaching the model a lot of things.
20:19We ended up helping to build the model.
20:21I had to put a lot of resources into it.
20:23For a whole year, that's what we did.
20:25The outcomes were not great because llms were not that
20:28Strong in those days, and my bank was like, this is not
20:31Working, it's a waste of our time, stop it.
20:34So you also have to know when to say, okay, not working, you
20:37Turn. Kill.
20:39Kill switch. Knowing when to turn on the
20:44Kill switch is also important. So as i said, a lot of this
20:49Cultural mindset shift is bringing the whole team along,
20:52Getting the feedback and learning, experimenting, fail,
20:54Fail fast, move on.
20:56I actually do have a perspective on this.
20:58We are looking at a lot of companies for whose
21:01Motivations might not be the right ones for doing ai
21:05Projects or implementing ai. We have quite a few companies
21:10For whom the ceo and the c-suite team are doing it
21:14Because the board said you've got to do ai.
21:19Without giving them a path to either scaling or an alternate
21:24Objective, which is to build capabilities and just learn.
21:29So either you're doing it for scaling or you're doing it for
21:33Learning, but those are two clear objectives, at least you
21:37Would be the aims, not my board said i had to do this or else.
21:43So if you have and then if there aren't those paths, there are a
21:48Lot of pocs that don't end up anywhere because the board
21:53Didn't also authorize scaling expenditure.
21:57But on the other hand, you know, not having a clear win that is
22:02Just having, you know, 50 people in your organization really
22:05Understand this and are ready for the next round, then that's
22:08Also not very helpful. So you can try a learning
22:11Objective or a scaling objective, but not to have
22:14Either and my board just told me to do this is not a great
22:17Recipe for success.
22:19But often it's the people at the cold face who will then tell
22:22You, hey, i think this can solve this problem.
22:25Some of my team used to say, there's so much grunt work in
22:29Our admin. Everyone is employee toil.
22:32What's the grunt work? for example, when we're
22:36Settling documentary trade, there's a lc, export lc, import
22:40Lc, then you have to read everything on the trade, make
22:44Sure the legal documents are fine, and reading legal docs
22:48Many times over and over again, right, we all know how grunty
22:52That is, right? so much grunt work.
22:55So today the gen ai, the great thing is it can read,
22:59Synthesize, summarize, and then tell you, yeah, check, check,
23:02Check the box, this document is perfect.
23:04Take that out of the human equation, the board does it,
23:07Perfect, and the board does it better than the human in some
23:10Instances, great. So the gen ai can help with a
23:13Lot of grunt work. Then where it's getting a
23:16Little bit more scary is when it can generate great pitch
23:19Books, sometimes better than the intern that came in last
23:23Summer or better than your youngest analyst.
23:26Then you have to think, okay, now if the pitch book generation
23:30Was used to be done by a human, is that good?
23:33But you keep the human in the loop to stop the hallucination
23:36Still. We still have to do that.
23:38I haven't met any tech that solves for hallucination yet.
23:41But if you have that, then you think, okay, young analyst, you
23:45Don't have to spend hours now. We used to have investment
23:48Bankers spend hours overnight sleeping in the office churning
23:51Out pitch books. Those days are over.
23:53Brilliant. So what does the analyst do?
23:56The analyst starts talking to customers to get a sense of what
24:00The customer wants, maybe coming up with ideation, right?
24:04And that's far more creative and far more fun than churning
24:07Out pitch books.
24:09So this really means that talent and culture, as you're
24:12Emphasizing, is so important. Jacqueline, when you're going
24:15Out there trying to attract the best talent and the best
24:18Companies, what are we offering to them as singapore in terms
24:21Of compute and an environment for ai start-ups?
24:24I do think that a lot of ai start-ups have come to
24:28Singapore, especially in the last two years.
24:32And they bring a lot of capabilities with them.
24:36What we've done is tried to ensure a couple of things.
24:39One is trained personnel. We've committed to training
24:4215,000 ai experts in singapore for the next few years.
24:45And actually that also provides sort of a base load of talent on
24:49Top of what is already a fairly digital country with quite a bit
24:53Of digital expertise in the nation. So i think we have some of that.
24:58We also offer a tech pass and a one pass, which are visas,
25:02Which are two years or five years long for individuals with
25:05Specific talents. And the tech pass has been
25:08Very attractive, especially to digital tech talents who are in
25:12The ai space. So basically it is actually
25:16Something that is quite attractive to a lot of people
25:19Because they're kind of golden visas for tech talent to come
25:22Into singapore, particularly to do ai.
25:25There's also quite a lot of funding available.
25:28So i mean, we remain the venture capital funding hub of
25:33Southeast asia and some say asia. So the money is looking for
25:38Really good prospects. The money is looking for real
25:43Presentations, real business models, and ai companies that
25:47Can help other companies solve their problems.
25:49So the money is going to be there. The talent is going to be there.
25:55And then finally the compute. A few years ago we had a
26:00Moratorium on data centers in singapore in acknowledgement of
26:03Our climate goals. And the fact that actually data
26:07Centers are very, very carbon intensive, ai data centers are
26:11Particularly carbon intensive and extremely energy dense.
26:16So recently we've decided to do a couple of things, one of which
26:21Is, of course, to lift the moratorium and we issued about
26:2480 Megawatts last year and we made an announcement that we're
26:28Going to be issuing another 300 megawatts in the next few years.
26:32And that's mainly for very strategic workloads such as ai.
26:35So i think the compute will be available.
26:38And, you know, we've also worked with our neighbors to
26:43Ensure that within a low latency sort of environment that the
26:48Compute is available at a place near singapore where data that
26:52Is maybe less sensitive, maybe not, say, financial sector, can
26:56Be stored and then be used to inference models for customers
27:00In singapore. So on all those fronts i think
27:04Singapore is really well positioned to lead in the ai
27:09Revolution of the next ten years.
27:12And i think this panel kind of really represents the singapore
27:16Public-private partnership, this tango between setting good
27:20Policy and leaders at the best institutions implementing and
27:24Executing it with real results. Thank you so much, sushant and
27:27Jacqueline for your time today.