Brainstorm AI Singapore 2024: How AI can bridge Asia’s linguistic and cultural divides
Kisson LIN, Co-founder & Chief Operating Officer, Mindverse AI Sambit SAHU, Senior Vice-President, Silicon Design, Ola Krutrim Leslie TEO, Senior Director, AI Singapore, Lead, SEA-LION Caroline YAP, Managing Director - Global AI Business and Applied Engineering, Google Cloud Moderator: Matt HEIMER, Executive Editor for Features, FORTUNE; Co-chair, Fortune Brainstorm AI Singapore
Category
🤖
TechTranscript
00:00Let's start by talking a little bit
00:01about defining the opportunity.
00:041,000 different languages, I'm not good at math,
00:06but I think that means about 50 billion opportunities
00:09every day for miscommunication, for crashing
00:12into a language barrier.
00:14How much economic potential could we
00:16unlock if we were able to have an AI tool that
00:20broke down these barriers?
00:21And what are some examples, you think,
00:22of the opportunities that are out there?
00:25I mean, Kisum, what's your thoughts?
00:27Yeah, sure, I'll be quick.
00:28I think there are definitely a lot of opportunities.
00:31I came from China.
00:32I know that the language barrier could be a huge issue
00:35until recently.
00:35So a few of them, work collaboration,
00:38when we have different colleagues
00:40from different regions speaking different languages.
00:42It's not only that people find it hard to collaborate,
00:46but also hard to bond with each other.
00:48So that's one of the issues.
00:50And another is in education.
00:53Equal education opportunity comes
00:54from having the access to listen to the courses
00:58from different languages.
01:00Entertainment, Medicare, and a lot of other things.
01:04So just a few data points, right?
01:06I think there are some earlier data points,
01:08like from European Commission, that the different languages
01:14actually cost almost about 1% of the GDP per year.
01:18And for customer service or marketing this specific area,
01:25we also have data from Alibaba that after they start
01:29AI translation for different products,
01:32the revenue actually went up to like 10%.
01:36And so for a lot of our clients as well,
01:39I think AI translation and also taking
01:42into consideration the cultural differences
01:44can actually benefit a lot.
01:46And Sambit, you're based in India.
01:47You're working in a market where, again,
01:49thousands of different dialects, thousands
01:51of different cultures.
01:52What are some of the opportunities
01:53that you can see in removing some of these barriers?
01:57I take what Keshav said.
01:59Lots of wonderful points.
02:01Health care, education is extremely important,
02:05and AI can play a big role.
02:08I think the other thing in India is
02:10we produce a lot of movies, lots of media
02:13in different languages.
02:14So we can translate that, then the whole entertainment
02:18industry, the media industry, the movies,
02:20they can pretty much propagate to the whole world.
02:24Yeah.
02:25I want to flip the question a little bit to the fact
02:28that generally speaking globally,
02:31the LLMs that are most widely used right now
02:34are trained and owned either in the West,
02:38the English-speaking West specifically, or in China.
02:42What problems are caused or what opportunities
02:45are missed for a user in this region
02:48if they're working with an LLM that originates
02:52from a sort of monolingual perspective?
02:55Curious about this from Leslie and Caroline in particular.
02:59So Matt, maybe it's important not
03:02to fixate too much about who built the model or where it's
03:06built, but to remember that these models are
03:09built on data.
03:11And what they learn is from the data.
03:15So the challenge for our region, by the way,
03:17my AI says there are more than 1,000 languages.
03:20I'm not sure about you, your AI.
03:24This region is not well represented
03:27in the internet or digital space.
03:30And therefore, what I believe, very powerful models
03:35are fed data that's missing the region's context.
03:41The region's language, yes, that's easy to see.
03:46But its way of thinking, its food, its history,
03:49its culture.
03:52And this is more of a problem for those languages
03:55that don't have huge populations.
04:00Not so bad for Vietnamese and Thai and Indonesia,
04:03but certainly very lacking for Khmer or Lao.
04:09And in Singapore for Tamil, Singapore Tamil.
04:12So for me, data is one.
04:14Second is, how do you define it's good or not?
04:17Most performance metrics, most benchmarks,
04:22whether we say a model is good or not,
04:24is in English or in Chinese.
04:28And we build the models to do better in these benchmarks.
04:32But they're all from a particular perspective.
04:35So those are my two answers, data, benchmarks.
04:40Caroline, you're working with Google, arguably
04:44a company with access to data on a scale that most of us
04:46can only dream of, but also, perhaps,
04:49a company that needs to think more broadly
04:51about its benchmarks.
04:52And tell us a little bit more about the 1,000 Language
04:55Project at Google and some of the recent accomplishments.
04:59It's a very ambitious goal of where we are.
05:03And recently, we added another 140 languages
05:06to that, including Cantonese.
05:08Now, for a lot of people who know of Cantonese,
05:11they might say, yeah, but it's a very well-established language.
05:15There's a written corpus as well as a spoken corpus.
05:17Why did it take that long?
05:19And a lot of models might have been
05:22used to the way grammar and sentence structure
05:26works for Mandarin.
05:29And not necessarily for Cantonese.
05:32And so then when you actually start to go, OK,
05:34to Leslie's point, the benchmarks
05:37were then created based on Mandarin
05:40and not necessarily on Cantonese.
05:42And so there was a step that was needed to actually change
05:46the way the benchmarks were even created so that we can measure
05:50the efficacy and, from a responsible AI perspective,
05:54how relevant even the outputs would be in Cantonese
05:58so that it was culturally correct.
06:00The other thing I would also say is, like Hokkien, for example.
06:04Hokkien is a language that's spoken by 46 million people.
06:08But there are so many differences
06:10in Hokkien as a whole.
06:13A lot of you are familiar with the Ang Mo Kio neighborhood
06:15in Singapore.
06:16They speak a version of Hokkien.
06:19I am from Malaysia.
06:20My family is actually from Penang.
06:22Penang Hokkien has a lot of Malay words in it.
06:25And so Teochew Hokkien, and Teochew is its own language,
06:30but it has a lot of Hokkien words as well
06:32and similarities.
06:34But that also then varies and changes.
06:37And so how do you inject a lot of these cultural nuances
06:40into the language itself so that it doesn't just
06:43look super binary and black and white
06:46but takes into the regional aspects how it's used,
06:50what it's meant to convey, formal, informal, casual,
06:55and make it where when someone asks the question,
06:58perhaps the model's response should actually
07:01be to collect a little bit more context
07:03before giving you the output.
07:05And the cultural nuances should drive
07:08what the next question should be back to someone like us
07:12to say, are you looking for it because you're greeting someone
07:14elder or younger, and then give you
07:17the output versus just assuming a direct translation is
07:20the only way to go about it.
07:22Those are great points.
07:24And there's also the issue of text versus tonal languages,
07:27which as a Westerner, I'm sadly ignorant about on many levels.
07:30But the notion that a given syllable or a given character,
07:35depending on the vocal inflection,
07:36can change meaning entirely.
07:40Kisan, you work in an area that you describe
07:42as AI personalization, which is focused a little bit less
07:47on the building of the models and more
07:48about building something that each individual can carry
07:51around and maybe reflect their individual characteristics.
07:54Can you talk a little bit about that
07:55and how language issues inflect the work that you're doing?
08:00Yeah, that's a great question.
08:01So we talk about different values in different cultures.
08:04And also, even in one country, people have different values.
08:08But let's put it further.
08:09Even for different people, they think about things differently.
08:12Male and female think about things differently.
08:14And you and I can probably approach the same questions
08:17differently.
08:18And we have different preferences.
08:20So just for example, I probably think
08:23whenever I have a fight with my boyfriend,
08:25he should apologize first.
08:27But he probably doesn't think so.
08:28So how do we reflect that kind of differences in AI model?
08:32We cannot rely on chat GPT to give us an answer.
08:36You're right, and he's wrong, obviously.
08:39Yeah, that's my preference.
08:42So here is the thing.
08:44We think that eventually, everyone's differences
08:46in preference, their value system,
08:48their chain of thoughts should be reflected by AI.
08:52But it's really impractical to build LLM for each individual.
08:56And it's impractical.
08:57It's expensive.
08:58It's not necessary.
08:59LLMs, for us, is more like a calculator.
09:02It's like our frontal cortex.
09:04But every one of us will have our own memory system,
09:07remembering who you are.
09:09Matt works at Fortune.
09:11And some of our own friends, their preferences
09:15and our own chain value system, et cetera.
09:18And so think about that LLM as your brain, a huge brain.
09:25But then this memory system that you built
09:28is more like a dictionary that you carry on with you.
09:31And that dictionary can update itself every day.
09:35So whenever I have another conversation with Matt,
09:38my dictionary updates about what I know about Matt.
09:41And that solves the ultimate issue
09:44of making AI personalized, but also scalable.
09:47So we eventually will have personalized AI model
09:50that serves everyone.
09:51And that can be plugged in with ChatGPT, with CLion,
09:55with a lot of other AI models.
09:57It doesn't matter what LLM model we use.
09:59But this dictionary that we bring with us
10:02is personal, is on the edge, and is with everyone.
10:05And so that, together, can be used
10:08in many different applications, like personal AI assistance
10:12or your agents, et cetera.
10:15And so that's eventually what we build.
10:17So do you think that we should create
10:19our own personalized AI agents so
10:22that the time for inference and things for the questions
10:25we're asking, because it's already so personalized to us,
10:28it's a lot quicker?
10:29Yeah, I think a lot of times, people
10:32have different preferences.
10:33So I can say I'm open-minded.
10:36I don't mind using different models.
10:38But some other people might be very specific
10:40on their preferences.
10:42And our philosophy is we think that we should
10:45give people back the choice.
10:48Let them choose how much influence they want.
10:50The opt-in option, which I think is getting bigger,
10:53especially now with data privacy and security and things
10:55like that, right?
10:56Absolutely, yeah.
10:58Leslie, just today, you've rolled out
11:00a second instance of the CLion model, which
11:04is trying to wrestle, of course, with all of these issues
11:06that we've just been talking about.
11:08Can you tell us and tell the audience a little bit more
11:10about what's involved in that and what
11:12some of the challenges of addressing
11:14these different cultural nuances and personal nuances
11:16have been?
11:17Sure.
11:18So we, fully by this discussion, there's
11:21a need for regionalization and localization.
11:24But a couple of other nuances we will add to it.
11:26In order to make this technology accessible,
11:31you do, yes, need to regionalize and localize.
11:34But you need to keep it cost-effective.
11:37You need to keep it quite fast, right?
11:41And by the way, the knowledge about how you should do this
11:45should be in the open.
11:47So CLion, like good public infrastructure,
11:52because that's how I model it, is
11:55meant to enable a lot of commercial activity
11:58and a lot of public use cases.
12:01But we do this in an open source way,
12:03doing things that commercial companies may not
12:06have time or money to do so, unless you're big like Google.
12:10So for example, most important, actually,
12:13is the data, cleaning the data, and open sourcing the data,
12:17building the benchmarks, and open sourcing the benchmarks.
12:20And lastly, having the models.
12:23We had a pre-trained from scratch model last year.
12:27This year, we are pivoting a little bit
12:29to take advantage of the work that's
12:31been done by companies like Google and Meta.
12:34So CLion version 2 is built on top of Meta for now.
12:39We are looking forward to building it
12:41with Google, a version on top of Google.
12:44But the idea is, actually, the models
12:47are there because some companies and actors
12:51can't train the model.
12:54But that's not the whole story.
12:55The idea is the benchmarks, the data, and the expertise
13:00is in the open for anyone to copy and follow.
13:04And we hope this creates an ecosystem
13:08to help Southeast Asians use this technology.
13:12In practice, in the open source model,
13:15have you received a lot of information
13:17that has fine-tuned and improved it?
13:19Have there been a lot of contributors saying, oh no,
13:21tweak that or adjust that if you want
13:23to be truly accurate of our culture, truly representative?
13:27Oh, I'm sorry.
13:28Have you been getting a lot of useful information inbound
13:30to your team from the open source model?
13:34So open source is not just about the model.
13:37It's a coalition.
13:39Google is actually part of that coalition.
13:41But we have AWS, Google's competitor.
13:45We have, importantly, key partners
13:49in every part of Southeast Asia building and contributing
13:53data.
13:54So in true open source fashion, we're
13:57saying if we have better data benchmarks,
14:01all boats are lifted.
14:03This technology is used.
14:05Our citizens benefit much better.
14:09Doesn't mean you can't make money,
14:10because you will take that 80% model now.
14:13It was previously 50% there.
14:15And you will have a better product.
14:18And I think the story resonates, because we do have 600 million
14:23citizens in Southeast Asia.
14:25And according to my AI, 1,300 languages.
14:30So we do have a big problem.
14:33And if we all contribute, I think the payoff will be huge.
14:38And actually, Matt, at Google, we've
14:43been able to innovate at the forefront
14:45to solve for some of these human challenges.
14:49But it's not enough for us to do it alone.
14:51We need really, really good public-private partnerships.
14:54And that's why it's been so good to work with Sea Lion.
14:56And also, traditionally, our competitors
15:01in different parts of our business,
15:03because the goal is to actually further where humanity can go
15:06and how AI can actually accelerate that,
15:08versus just making it all about lock-in and proprietary
15:12and things like that.
15:13And so that's why it's been such a big deal for us
15:15with the work we've done with also Minister Teo
15:18and the Singapore Area Trailblazers Program
15:20to understand how we can help Singaporeans develop
15:24solutions to solve for local problems,
15:26but with a global reach.
15:27And this whole thing with language and culture
15:30is really near and dear to me, because there are so many
15:33things that we do here that's very different.
15:36Four languages in a sentence to order a meal, right?
15:39And so how do you capture that context
15:41as you switch from one word to another
15:43to be able to capture that?
15:45Those things are really, really key.
15:47And I mean, India has a lot of this.
15:49It very much does.
15:50And Sambit raised a really fascinating question,
15:52which I'd love to come back to, which
15:54is, as you're trying to bring together all these languages
15:58and find a program that communicates across all
16:01of them, do you risk losing some of the individual cultural
16:05nuances?
16:06Do you risk saying, well, everyone, we all
16:08need a lingua franca in the center,
16:10so we have to give up some of our quirks
16:12so that we can communicate effectively?
16:15And I'd love to know, how do you,
16:17thinking about working in India, where
16:19you have so many languages and so many different cultures,
16:21how do you make sure that the coming together in the middle
16:25doesn't require everyone to leave their distinctiveness
16:27at the door?
16:28Yeah, great question.
16:29As we all know, India has a lot of languages
16:33and a lot of different cultures.
16:35There are 22 scheduled languages as per the Constitution,
16:38but there are 20,000 languages and dialects spoken
16:42all over the country.
16:43So I think the model is as good as really the data
16:48that we feed it.
16:49So we have to make sure that we get data
16:51from the different languages, different states,
16:54different cultures that we train the model with.
16:57So that's really very important.
17:02And then once you train the models,
17:06you have to make sure that there is no under-representation.
17:09You also certify the models with experts
17:12from different languages, different culture,
17:14different states.
17:15You have to make sure that you certify them and test them out
17:20so that there are no biases.
17:21There's no under-representation.
17:24The third piece which we are attempting
17:26is really how do you develop the ecosystem?
17:29I mean, model is just one piece of it,
17:31but really how do you enable the model
17:33to reach out to the broader Indians and the developers,
17:38applications?
17:39So as you grow more and more models, more data,
17:45more applications, I think you'll
17:47see a fair representation across broader India.
17:51Hey, Matt, I have a question for my fellow panelists.
17:53How much focus are you putting in
17:55on what we call the human in the loop
17:58and the reinforcement learning through the human feedback?
18:00Because Leslie, you said that there's
18:02all of these different collaborations
18:04with all the different open source communities, et cetera.
18:06How are you building the right frameworks
18:08to capture a lot of that?
18:09Because we have our own framework,
18:11and we've published our AI principles,
18:14and we have a fairly robust way to test a lot of our frameworks
18:19and stuff.
18:20I'm just curious how all of this is also
18:22being done and what you're all finding as best practices,
18:25if that's OK.
18:26Sure, absolutely.
18:26Yeah, yeah.
18:27So one thing that we are doing, we
18:29are reaching out to the universities, developers
18:33to try out the models, develop applications,
18:37and how to use our models effectively,
18:41how to find our problems.
18:42So I think reaching out to a broad set of developers
18:46and reaching out to universities, college students,
18:51if it enterprises, as much as you can reach, it's good.
18:55So it's early days, I would say.
18:59But I would say that over time, we'll evolve.
19:02And we'll have a lot of developers
19:04and a lot of users using it.
19:06That's a great question, though, to think of.
19:08Forgive me for interrupting.
19:09But to think of, is the human in the loop someone
19:11from one of these smaller communities,
19:13maybe a language that only has a couple of hundred thousand
19:16speakers rather than a couple of million,
19:19but you still need a human from that?
19:20Do you still need a human from that constituency in the loop?
19:23I hope the answer is yes, but are we able to find one?
19:27Maybe I'll add to this.
19:29I think that's part of the solution.
19:32Three points to make.
19:32One is, at least for me, the golden sauce
19:35is still a human, no matter how good our benchmarks
19:39and measurements are.
19:41Second, just like we are an exam-focused country,
19:48we are very focused on metrics.
19:51But all of us who have hired for HR
19:54know that exams are only a indicator, maybe not even
19:59the best indicator.
20:02So I have that view of many LLMs.
20:04They work really well on benchmarks.
20:07They don't in a use case.
20:09The third point is, and we try to do this
20:12with partners like Google and with help from your team,
20:18actually, once we build the cultural representation
20:20for Southeast Asia, it has to be done by Southeast Asians.
20:26So people in Indonesia, even in Indonesia,
20:30it can't just be done by people in Jakarta.
20:34We can't just have students.
20:36Actually, students are the easiest group for us
20:38to reach out to.
20:39But you can imagine, students will give you a certain view.
20:42And that's not representative of society or even your users.
20:47So across the region, how do we do this
20:51in a way that is also correct from a technical point of view,
20:56but correct from a representation point of view,
20:59and also correct that we don't exploit any of these data
21:04providers, if you like.
21:06I think I just want to add one more point is,
21:09for broader outreach, I think we have
21:11to make chips less expensive.
21:13Today, training chips are extremely expensive.
21:16And the second piece is, we have to make sure
21:19that these large language models,
21:21they're accessible with our personal devices
21:24like phones and wearables.
21:25Absolutely.
21:26I want to jump in very quickly here.
21:27I think on top of that, we also want ultimate solution
21:31is to make the AI trainable by itself.
21:33So whenever they interact with every user,
21:36they will take the input from the user and adjust itself.
21:40I think that's the ultimate solution.
21:43And that point about access is so important,
21:45especially for someone who's working just off a phone.
21:50The farmer in the field needs the smaller device
21:54with the fewer parameters.
21:56Well, it's exciting to see all the work that you're doing.
21:59And I'm so grateful that you've been able to share it with us.
22:01We're coming close to the end of our time.
22:03I just want to encourage, I hope you'll all
22:04be sticking around the conference for a while.
22:06And I hope that everyone here in the audience
22:08will feel free to hunt us down afterwards
22:10and ask more questions, because there's a lot of knowledge
22:13on the stage.
22:13So Caroline, Leslie, Sambit, and Kisan,
22:16thank you very, very much for joining us.