Select a choice

Episode 3:
The Future AI Generation and Gen AI 2.0

This episode introduces the reality of using AI solutions in Singapore and how dependent or aware are citizens on their usage and how can one leverage the solutions to attain and master 21st century competencies. We chat with Darius Liu from AI Singapore as he shares more about his experiences.

AVAILABLE AT YOUR FAVOURITE PODCAST DESTINATIONS

AVAILABLE AT YOUR FAVOURITE PODCAST DESTINATIONS

Transcript

Part 1: The Beginning of AI	0:22	Hi Darius
	0:33	Hi there, Shahidah
	0:35	Thank you so much for coming and joining us today for The Teaching Point.
	0:45	Of course, my name is Darius.
	0:46	I’m Darius Liu, I'm the Head of Strategy and Partnerships at AI Singapore's product team.
	0:52	This is the team that builds products that we open source for the communities use, almost like a digital public good or common infrastructure, so to speak, right? Right.
	1:01	If you take a step back, AI Singapore as an organization has been around since 2017.

	1:05	It is the government's national AI program.
	1:08	And we exist to bring the benefits of AI and AI solutions from academia into the industry, right?
	1:16	So we do this via research.
	1:18	We do this via organizing the grand challenges, for example.
	1:23	And we do this also by putting out open source tools and products, which is what my team does.
	1:29	I think just to kick things off,
	1:34	Could you maybe explain a little bit more or share a little bit more about your first experience delving into the world of AI?
	1:42	Sure.
	1:44	I guess when you mean AI in this context, it's probably generative AI as we understand it, right?
	1:50	And like most of us, the first introduction or the brush with AI was when ChatGPT came right out of the scene somewhere around end of 2022.
	2:01	And that's when, you know, the Singapore government and Singapore set up and say that, look, you know, this is probably a step change in how the AI landscape is going to evolve.
	2:10	And that led to rethink about how we want to position the product scheme.
	2:15	And that's where this initiative about large language models for Southeast Asia came out.
	2:19	I see. So having that, you know, surge of interest in ChatGPT or in generative AI, in your opinion, what would you say is the level of awareness that people have of the amount of AI or generative AI at the very least that they are using on a daily basis?
	2:42	I think it's increasing.
	2:45	Maybe if you ask this question to me, maybe a year or year 1/2 ago, I think people may not fully understand how much of generative AI, even in that level of development, is embedded into the tools and the solutions that we use everyday, right?
	2:59	Like if you talk to a voice bot, it could very well go into some kind of a large language model which generates a solution.
	3:07	Whether the solution is good or bad, that's up to you to decide.
	3:10	But the solution does embed some form of AI.
	3:14	If it's not a large language model, there could be natural language processing engine behind it or for some other applications, perhaps some computer vision, right, which is also another branch of AI.
	3:25	So I think Fast forward to today, I think there's a lot more appreciation about how much AI is being embedded to the extent that people are starting to second guess whether somebody on the other line into talking to them is actually a bot or it's actually a human.
	3:39	So the awareness is increasing, but I would say that there's still ways to go.
Part 2: The Key Uses of AI	3:44	Could you maybe share some examples, maybe in Singapore as well as the region, on how large language models have been put into place and how they were effective?
	3:53	Yeah, for sure.
	3:54	So I think this is an example that's of course has been made public, so I'm free to share this.

	4:00	So if you happen to go to Indonesia and you speak into the Gojek app, like to call a car to make a payment using the Go Pay application, you tend to speak into it.
	4:12	If you're an Indonesian, of course, Bahasa Indonesia or if you happen to speak one of the major dialects like Balinese or Sudanese and everything, you are able to do that, right?
	4:20	And the reason why you're able to do that is because the Gojek group, the maker of Gojek and Gopay has integrated a variant of a large language model that is able to be conversant in Southeast Asian languages, specifically Indonesian languages.
	4:38	We are proud to say that we worked with Gojek to create the model, which is called Sahabat AI and it's serving millions of Indonesians as we speak.
	4:50	So this is 1 good example of localization and how language is important.
	4:56	So if you, you know, rewind the clock one year ago and you tried the same thing using one of the Western, you know, American models or the Chinese models and you try to speak to it in Bahasa Indonesia, it probably doesn't sort of understand you.
	5:12	You try to even speak to it in English, but with our own Singaporean accent.
	5:16	Yeah, yeah.
	5:17	It, it does a pretty not so good job, right?
	5:21	You have to do the fake American accent and here.
	5:24	So.
	5:24	So it is a problem, right?
	5:25	Yeah.
	5:25	It directly impacts the usability of AI.
	5:29	Yeah, true.
	5:31	Maybe we could take you back a little bit and you know, we talk about how it has been effective using like the Gojek and the Sahabat thing.
	5:39	But maybe for those of, you know, our listeners or even for some of us who are not as knowledgeable about how it works, maybe could you share with us a little bit more about, you know, what are the key steps that you took in terms of getting the data, you know, not specific data, but like just the process.
	5:56	Yeah, correct.
	5:57	So of course the first part of call is whatever is open in the digital realm, right, the Internet and unfortunately there is not much Southeast Asian data on the Internet.
	6:06	But whatever is there, we take it, we sieve out those parts that are not suitable.
	6:14	For example, they are mainly advertising or machine translation or something a little bit more not safe for work kind of content.
	6:23	We filter out the copyrighted material because we are very cognizant that we need to be do this properly and we end up with a decently sized corpus, not very big, but get off to good start.
	6:38	What we do a lot is work with our partners in the region.
	6:42	I see because a lot of the Southeast Asian data is not in digital form.
	6:47	If it's even if it's in digital form, it's not in a form that's consumable by a model training algorithm.
	6:53	So for example, it's in the PDF, or it's in the OCR or it's in a scanned document.
	6:58	So a lot of work goes into processing that data to cleaning it up, to putting it in into a format that is able to be ingested by model training algorithm.
	7:08	And then we add it to newer and newer generations of the model that we are creating and we put it out in open source.
	7:15	I see.
	7:16	If you were able to quantify maybe in a time frame, then how long would this process actually take?
	7:24	The first iteration took about 10 months and then obviously the pace of change accelerates there on, right?

	7:33	So the next iteration took maybe six or seven months.
	7:37	Again, the next iteration took about four months.
	7:39	So the time scale of the model development is decreasing.
	7:43	Yeah, welcome to the AI arms race.
	7:46	So it's actually interesting because I was expecting you to say that it's going to take about maybe a year for the initial because of the lack of data and how much you really have to see about right And how much you have to work with partners and whatever jurisdictions that they have and whatever logistics or basically regulations that they have, it will take slightly longer.
	8:06	So I'm actually surprised 10 months for a considerable amount of change.
	8:09	It's actually quite fast.
	8:11	And the fact that in every run it becomes faster.
	8:15	It just goes to show how much AI can be integrated into our work life, our daily life and as well as the systemic level.
	8:24	So given that we are, so I would say maybe in the past year or so we've got acclimated to AI, right.
	8:32	May I then ask, do you think that there's still a growing stigma if I say that I use AI in my job?
	8:39	I think there would be a stigma if you say that you do not use AI in your job.
	8:44	Again, that whole perception has changed.
	8:47	Two years ago, if you use AI in your job, people thought you were trying to cut corners or you're lazy or whatever.
	8:53	Today, if you do not use your AI, they'll be asking, you know, why aren't you sort of working smart or using the latest tools or, you know, getting with the whole sort of flow of where the technology is going.
	9:05	So I think that that mindset has definitely shifted.
	9:10	I would say, yeah, it is about using what's best out there rather than dogmatically sticking to some, you know, way of doing things, which has worked well in the past.
	9:20	But you know, with the way that technology is moving, it may not be the best way to go for.
	9:26	Do you think then if let's say you were to compare the younger working force with maybe those who are more mature?
	9:35	And do you think that there is still that struggle for the mature ones to come into terms with the more AI-driven way of work?
	9:44	Or do you think that right now it's a lot more of an open environment and even the mature, you know, adults are a lot more receptive,and a lot more interested in trying out?
	9:57	Yeah, I think it's, to me, it's a bit of a myth that the so-called more mature workers in the industry find it more difficult.
	10:09	AI is one of those technologies where if you bring in domain expertise, you bring in experiences from the manual way of doing things into and you utilize AI, you overlay AI solution on top of that.
	10:25	Those act as multipliers to be able to allow you to use AI in a more intelligent manner, right?
	10:31	Versus let's say, for example, somebody coming fresh out of school or, you know, working for one or two years and hasn't had those, you know, experiences in the workforce, hasn't stepped on some of these land mines before.
	10:43	And then you take the the AI output and say, what do I make of this?
	10:47	How do I review this?
	10:48	How do I critically assess what the AI is telling me?
	10:52	The ability to do that is a meta-level skill, is a higher-level skill that comes with experience and domain expertise.
	10:59	And you know, more experienced workers do have a hitch in that.
	11:03	So in order for AI to be harnessed, it has to, you know, be coupled with real-world experiences.
	11:10	It's not just a be all and all.
	11:11	It's not a magic genie that you can.
	11:13	Yeah.
	11:13	So just ask anything and you get the right answer, right.
	11:17	So I would imagine that even creating the quiet environment and process will allow not only like the mature working adults to be more receptive, but will also create a process that's a lot more elevated and can synchronize work at a much faster rate.
	11:34	Yeah, correct.
	11:35	So, I mean, if you one good example would be this idea of a mod, this idea of you have a model and you press a button and the output comes out versus the need to be able to structure a series of models into a certain workflow, right?
	11:50	The ability to synthesize a workflow into various model architectures is a skill that perhaps is not as easy as pressing a button and getting it out, right?
	12:01	You need to understand the business use case.
	12:03	You need to understand where the efficiencies are and where the bottlenecks are.
	12:06	And that is something that, you know, comes with domain expertise, like I said, and experience.
	12:11	Yeah, OK.
	12:13	But I think this section definitely is a lot more a lot insightful and it definitely opens up the thought of not having AI go away anytime soon.
	12:24	So I think people used to think that it's a phase and it will go away.
	12:28	But I think now that it is a lot more integrated in our work life, I think people understand that it's probably a permanent solution here on out.
	12:37	Yeah, definitely a permanent feature, right.
	12:39	If you look, take a look at how technology has evolved, where are the step changes in the last maybe 30 years of computers mobile in our AI, Yeah, we mentioned quite a great deal about, you know, large language models.
	12:56	We also mentioned about, you know, how it has been more how AI has been effective in making sure that it's not just a copy and paste situation, but it's also integrating it into the system, having that knowledge be complementary instead of it being the driving force, right?
	13:14	So it's let's say we were based on like the residents or the people, end users that you've met in your time in AI Singapore and even in your experience with working AI projects.
	13:24	Could you maybe share some of the tips that you would recommend to any user?
	13:27	Because we always expect that you know, one to learn lifelong, right?

Part 3: The Tool to Master 21st Century Centuries	13:32	So even in their adulthood or even in their teenage years and who may think that AI is such a daunting task or such a daunting concept, What will somewhat will be some recommendations that you give them to start things off with AI?
	13:46	Yeah.
	13:46	So the recommendation I would say in my advice to these learners to be is don't start with AI as the first concept.
	13:56	Start with what problems in your life are you trying to solve?
	14:00	Are you could potentially automate, right?
	14:03	Will you automate letter generation or could you automate some kind of a data processing, whatever it is, right, Or e-mail writing, right?
	14:10	Everybody writes emails to some extent, yeah.
	14:13	Could you automate that?
	14:15	And then you have, we all watch YouTube, right?
	14:20	That would be a good place to start.
	14:21	Not all YouTube videos are the best, admittedly, but you know, it's a good place to start.
	14:26	And you can search online for tools.
	14:30	There are now AI tools for just about anything you can possibly imagine.
	14:33	So start with what you potentially think that you can solve using AI or improve using AI and look for something that works in that domain, right?
	14:43	Or works for that task and I can get almost guarantee you that there will be some kind of application with a free or freemium model that will be able to speak to that.
	14:53	The other advice I would give is that there are a lot of very good education courses online for like the Coursera or AI.
	15:01	Singapore also puts out this very fundamental open-source online learning called AI for Everyone.
	15:10	And as its name suggests, it is really a very introductory course on what AI is, what AI isn't, what it can be used for, what it cannot be used for.
	15:21	And that's really a good place to start.
	15:24	So I, I think there are a lot of materials out there and I like the way you mentioned, right?
	15:29	If you look at the problem first, before you look at the solution and problem solving is such a crucial task given how, how complex is our daily lives, right?
	15:39	We in our personal life or in our hobbies and definitely in our work life or at the systemic level.
	15:45	Now, as we talk about learners, right?
	15:46	And we in education, we always talk about the concept of understanding and mastering 21st century competencies, or these are like your critical thinking, self-directed learning, as well as your problem solving.
	16:00	How do you think AI is able to still foster these skills or competencies despite them performing the task for you?
	16:12	Right. So one of the things that AI in its current shape or form is very good at is idea generation.
	16:20	For example, you want to write a poem of, I don't know, Hariraya or something, right?
	16:26	You can tell the AI give me 5 poems, right?
	16:30	And you look at the five and say that, OK, maybe this one is good.
	16:34	Expand on this one.
	16:35	I like this passage.
	16:36	I don't like this passage.
	16:37	Then give me another 5.
	16:38	And the way at which you evaluate the five, that's critical thinking, right?
	16:43	So you don't necessarily have to go down to the writing, the poem itself, but you have to elevate yourself to the, the OK, I can review the output.
	16:53	Just like how in traditional corporate workplace, you start off as a grant worker doing a manual labour and then you have the supervisor who's reviewing the work.
	17:03	Now replace all the interns or the entry level workers with an AI tool, right?
	17:09	And then here you are, you are forced to elevate yourself to a reviewer role.
	17:13	And that's where we see, I see the workforce going, the workforce over time, role development and upscale naturally because the AI is able to take on a lot of these other tasks.
	17:29	I also agree, I think that once you kind of not eliminate, but once you're able to complement any job with that already first line of education, then naturally and person or even a young person or no matter the age, it's required to kind of push themselves forward, right?
	17:46	And look at it from an analytical standpoint.
	17:50	Now you mentioned about, you know how the workforce is shifting it that way.
	17:54	How else do you see AI shaping the future, maybe near future or maybe in 5-6 years to come in terms of shaping the workforce?
	18:06	Like I said it, it sort of will elevate everybody's role in a way to in a way a higher level abstraction of whatever task that you are, you know, employed to do, right?
	18:20	So for code reviewers, like you write your programmers to code reviewers, right, Code reviewers to serve code deployers, something like that.
	18:31	So I think just like how you know, in computer programming 20 years ago, everybody wrote in a certain thing called assembly language.
	18:39	Nobody does that anymore because the programming language does that for you, right?
	18:44	Yeah.
	18:45	And in future maybe you probably may not write code as much as you people do today.
	18:52	It may just be pseudo code or logic that you prompt the engine and the AI works to generate the code for you.
	18:59	But whether or not that code is efficient, whether that code gives the output that you want, or if you have 5 versions of the code, how do you make sure that you are able to understand which version works the best in the context that you provided it?
	19:16	That's the skill, right?
	19:18	It is no longer about generating the code, it's about reviewing and critically looking at the various aspects of the code to see whether or not it's fit for purpose.
	19:27	That's the skill that will naturally come because the the jobs will call for it.
	19:34	Yeah.
	19:34	And human beings, as a race, we are fully adaptable and we've adapted in the past.
	19:39	We will adapt again.
	19:40	So there's also that minor way.
	19:44	If we are going to be a bit over complacent, then you might end up in one of those pitfalls where we don't we either replicate what the AI can do or we try to overcompensate or under compensate and therefore be providing or creating gaps in the system itself.
	20:04	Yeah, I think so.
	20:05	So I think again back to the point about fully understanding what AI at that current juncture is able to do or not the.
	20:16	Underlying models, the solutions, what are the advantages, what are the disadvantages?
	20:21	I think that's really, really important.
	20:22	And I think if there's one thing that will leave the listeners today, it is really about fully understanding the nature of AI, right?
	20:33	The nature of the solution.
	20:36	It is there to perform certain tasks in certain ways and it's not there to perform all tasks that you asked about it, right.
	20:45	And the ability to distinguish one from the other is what will make a good AI user versus somebody who's, you know, maybe a little bit less clued in on the other topic.
	20:59	Well, I think this conversation has been helpful to me and I think it will be helpful to our listeners as well because not only do we first understand what is AI, but the different components of AI like large language models and how within the South Asian, Southeast Asian region, it's so important for us to have this element to ensure that, you know, whatever processes that we're using, it's contextually accurate and it's also acclimatized to our specific needs.
	21:28	So thank you so much, Darius.
	21:29	Thank you, Shahidah.
	21:30	I definitely learned a lot and I think our listeners have as well.
	21:34	So thank you.
	21:34	Thank you so much.

More about our guest below

Darius Liu

Darius currently holds the position of Head of Strategy, Partnerships & Growth at AI Singapore, a national program dedicated to amplifying the societal and economic impact of AI research and technology. His responsibilities include leading the development of strategic partnerships with government and industry partners, and facilitating the commercialization of AI Singapore's product and technology solutions.

Prior to AI Singapore, Darius co-founded ADDX and created the world's first fully-regulated platform for digital securities which integrates issuance, custody, and secondary trading of private market assets in a digital platform powered by blockchain and smart contract technology. (www.ADDX.co)

As the Vice President of Total Portfolio Strategy (TPS) at GIC, he specialized in asset allocation and portfolio construction for GIC, Singapore's sovereign wealth fund. He initially joined the Human Resource and Organization Department in 2013 before transitioning to the TPS team in 2015.

In his earlier roles, Darius served as a Singapore Government Administrative Service Officer and the Head of Strategic and Operational Planning at the Singapore Ministry of Defence. His significant achievements include establishing the strategic planning and capability development framework for the organization, and operationalizing the enterprise budget and resource allocation mechanisms. He also served in the Singapore Ministry of Finance, managing strategic issues related to Singapore's financial reserves and overseeing matters pertaining to the national reserves held and managed by GIC, Temasek Holdings and MAS.

Academically, Darius is a recipient of the Singapore Public Service Commission Overseas Merit Scholarship. He graduated Summa cum Laude with a BSc in Mechanical Engineering (2003) and an MSc in Mechanical Engineering (2004) from Cornell University. He later earned a Master of Science in Financial Engineering from the National University of Singapore. Professional certifications include the CFA and CAIA Charters. Additional qualifications include a Performer’s Certificate in Violin (Trinity School of Music, UK) and a Private Pilot License (IFR land and sea) from the US FAA.