Brainstorm AI Singapore 2024: Identifying High-Impact AI Initiatives For Your Business
Saemin AHN, Partner, 500 Global,
Chris CHELLIAH, Senior Vice President, Technology and Customer Strategy, Japan and Asia Pacific, Oracle,
Miao SONG, CIO, GLP,
Gina WONG, Head of Global Alliances, Kyndryl,
Moderator: Jeremy KAHN, FORTUNE
Chris CHELLIAH, Senior Vice President, Technology and Customer Strategy, Japan and Asia Pacific, Oracle,
Miao SONG, CIO, GLP,
Gina WONG, Head of Global Alliances, Kyndryl,
Moderator: Jeremy KAHN, FORTUNE
Category
🤖
TechTranscript
00:00Welcome, I'm Jeremy Kahn, I'm Fortune's AI editor.
00:04We're here to discuss the AI rollout,
00:05identifying high impact AI initiatives for your business,
00:09and we've got a great group of speakers.
00:11Now for a little bit of housekeeping
00:12before I introduce them,
00:13a reminder that this session will be on the record.
00:16Additionally, when we turn to audience questions,
00:19and we will at some point come out
00:20to audience for questions,
00:21please note you'll have to raise your hand,
00:23someone will come by with a microphone,
00:25then if you can stand up and sort of state your name
00:27and also your affiliation, that would be great.
00:31With that, let's start with some introductions of the panel,
00:34but do remember this is meant to be a group discussion,
00:36and we look forward to incorporating
00:38as many voices from the room as we can.
00:41Immediately to my left is Semin Ahn,
00:44partner at 500 Global.
00:46Then to his left is Mia Song, CIO at GLP.
00:51Then we have Chris Chilae,
00:54Senior Vice President Technology and Customer Strategy,
00:57and Asia Pacific at Oracle,
00:59and Gina Wong, Head of Global Alliances at Kindrel.
01:03I want to start out,
01:04because I want to get a show of hands in the room.
01:08How many people in the room have experimented
01:10with using generative AI,
01:11at least at sort of a proof of concept level
01:13in their business?
01:15How many people are experimenting?
01:16Good number.
01:17And then I wonder how many of you
01:18have actually deployed something
01:20at sort of full-scale deployment
01:22of a generative AI product in your business?
01:25Actually, it's a decent number of hands.
01:28I think that's interesting,
01:30because globally, statistics are,
01:32well, there's a lot of experimentation,
01:33there's very little full deployment,
01:35so it's very interesting we have
01:35a number of people in the room
01:37who have deployed generative AI at scale,
01:39so maybe you'll have some lessons for us,
01:40and maybe for the panel on that.
01:43I want to start out by asking all of our panelists,
01:45I think a lot of people,
01:46what they struggle with is how to identify
01:48the best value use cases for AI,
01:52because I think people just don't know
01:53the best use cases to target,
01:55or sometimes the things that seem highest value
01:56also seem the highest risk,
01:58and so people are sort of afraid to take that on.
02:00Sameen, let's start with you.
02:01When you're looking at portfolio companies, potentially,
02:03where do you look for to see,
02:05oh, that's a really good high-value use case,
02:07and how do you kind of assess that?
02:10Gotcha.
02:12Yeah, is Sameen's mic on?
02:13There we go, all right, there we go.
02:15I think generally, when you look at the use cases,
02:19for me, because I look at the growth stage of investments,
02:22I look at where the macro has impacted
02:25and hasn't been impacted by the LLMs and AIs of the world.
02:29So if you look at things that have happened,
02:30especially on the freelancer level,
02:32things like video editing,
02:33things such as copywriting,
02:35things such as bidirectional general communication by email,
02:38it's become immensely impacted to the point
02:40where you increase the proficiency and the effectiveness
02:45of these HRs, about three to four X,
02:48which also decreases the need for those human resources
02:51by a factor of two.
02:52So things on the more generalized interaction side
02:56of customer service and customer interaction
02:59have already proved massively, massively interesting.
03:02If you look at the larger top end of things,
03:04Microsoft just announced their Q2s, right?
03:08They're on a $5 billion run rate for Azure AI.
03:11That's not accounting for OpenAI's $5 billion
03:14that they spend on.
03:15So that means that already,
03:17their Azure generalized platform
03:18accounts for 5% of their top line, right?
03:21And if you look at what Bank of America has said,
03:24they said only 5% of enterprises
03:27has actually inputted and accepted
03:30some sort of general AI market.
03:32That means if you put those two figures together,
03:34there's a lot of runway to go on the potential
03:37and the actual upside of AI in general.
03:40Great.
03:41Miao, what's your sort of view
03:43on how to identify the best use cases for AI?
03:45Yeah, I think there are a few things I want to call out.
03:48So AI has been there for decades.
03:51So it's not just generative AI.
03:53So it's machine learning, predictive AI, and gen AI.
03:57So when I look at the value,
03:59I start to look at the AI vision.
04:02I think importantly, you need to have AI strategy
04:05to define where you want to travel.
04:07So at the moment, a lot of AI use cases
04:11are purely focused on enhancing of productivity.
04:15So all the copilots you're running with Microsoft,
04:18it's basically saving some hours of people, right?
04:21So that's it today.
04:23But I think broadly, I am looking at
04:26much more value-added AI use cases
04:29that relevant to grow the company's top line.
04:33So rather than focus on bottom line,
04:35which is productivity gain,
04:37you have to think about the top line.
04:39For example, how you use AI to drive customer engagement.
04:45How you use AI to identify future opportunities.
04:49How you use combined AI solution
04:52to predict what's gonna happen
04:54in the future of your business.
04:56I just show you some of the live use cases.
04:59So that's one of the pillar of AI.
05:01But also, you have to look at also the adoption.
05:05So in a small scale adoption,
05:07which save few hours per week,
05:09it's not gonna support the need of the business.
05:13So one of key thing is scalability.
05:16How you actually scale up the AI solution
05:19in a much larger scale to make it much sustainable.
05:23So for me, it's not about creating a chatbot.
05:27So that chatbot talking about HR policy,
05:29travel policy, saving,
05:31it's not about just building Excel
05:33and convert it to PowerPoint automatically.
05:36I think most importantly,
05:37all the adoption is around how you actually
05:40create a much bigger business case.
05:43One of things a lot of people don't talk about
05:45is building a digital product.
05:47Today, the chatbot, the startup,
05:49the AI solution, co-pilot,
05:52it just direct conversation with generative AI.
05:55In order to make it more useful or scalable,
05:59and I think it has to be focused on digital product.
06:02So everything you build has to be a digital product.
06:06I mean, then you scale it up
06:08into a much sustainable solution.
06:11It could be industry focused.
06:13It could be solving the financial services problems
06:17that you help your organization
06:20to actually have a deep understanding
06:22of your financial performance.
06:24Identify the root cause,
06:26why you fail to meet the target of your investors.
06:29It could be that develop a holistic marketing campaign
06:35using generative AI,
06:37but also identify the future customers.
06:41A live example,
06:42if one of the potential customers search on your website,
06:45you can potentially automate your marketing campaign,
06:48but also enhance the reach of your customer,
06:51and eventually drive the conversion of your customer.
06:55And that's a live example of actually
06:58contribute to a revenue growth.
07:00So that's how I look at a use case.
07:03The other I want to mention is,
07:04I call this AI framework,
07:06which is AI architecture.
07:09Yes, generative AI, chat GPT is dominant,
07:14but you have to think about a different layer, right?
07:16There are other ROI models,
07:19and different model exists because there's a need.
07:22So for me, I focus on creating this,
07:25I call it a five-layer architecture of AI.
07:28Bottom layer is a cloud infrastructure.
07:32On top of that, you can integrate
07:33with all the large language model,
07:37other AI models or APIs, that's a second layer.
07:40The third layer is where you identify the integration,
07:46and then the top layer will be
07:48how you attribute a use case.
07:50Through this multi-layer architecture,
07:53I was able to build capability,
07:55but also identify people,
07:58talents needed to build this solution,
08:00because at each layer,
08:02the need for different skill sets
08:04and talents are very different.
08:07In order to do that,
08:09I'm thinking about setting up AI COE
08:11to support my AI strategy.
08:13So that's how I look at AI holistically.
08:15Interesting.
08:16That's just a lot to think about,
08:17and I want to circle back to some of those points,
08:19because I think some of that,
08:20it sounds great,
08:21but companies, I think,
08:22struggle to figure out exactly how to do it,
08:23and they worry about that talent piece in particular.
08:25Do they have the right people to actually deliver on that?
08:28I think sometimes they're finding the answer is no,
08:31and they're not really sure how to bring in the right talent.
08:32But Chris, when you're dealing with people at Oracle,
08:37and they're trying to build out AI solutions,
08:40how do you identify for them,
08:41and help them identify what some of those
08:43highest value added areas are going to be?
08:46And then, yeah, maybe there's this issue around,
08:48okay, that's a high value area,
08:50but can I actually deliver that?
08:52And there's sort of a thought that I should be able to,
08:55but I don't know how.
08:56Yeah, look, and I'll pick a couple of threads
08:58from my previous speakers here as well.
09:00I think there's two horizons of delivery.
09:02There's the productivity immediate horizon,
09:05and then there's that horizon of innovation,
09:07so that's sort of the bottom line efficiencies
09:09and the top line innovation pieces.
09:11And I think the productivity,
09:12and I'd love to get the summary from the show of hands here
09:15that says you've sort of deployed full scale.
09:17I'd almost say that most of them, typically,
09:18in my seeing, customers have done
09:20a lot of that productivity side of things,
09:21so bringing in chatbots,
09:23and bringing in sort of customer service agents
09:26that'll help their agents relate better to their customers.
09:30But I think when we, and those are good.
09:32Those are absolutely good.
09:33They're great starters into the cycle.
09:37I think when we look at the high yield fruit,
09:39the high yield fruit is when I think you can bring AI
09:42into your organization, and what does that mean?
09:45So when you take, you know, one litmus test for me
09:47is if you're deploying just a large language model,
09:50well, that's general knowledge.
09:52That's like, you know, everybody's got the same thing.
09:54It's when you can bring that model into your organization
09:58and infuse your organization's knowledge into it,
10:01and to me, that's when you're gonna drive that innovation,
10:04and so I think, you know, where I think the opportunity is
10:07for us to go and help customers identify,
10:09hey, which pieces of data scattered around your organization
10:12can we use to augment the use case?
10:15And then you're building differentiation
10:17versus just productivity, and I think that's what we try
10:20to help our customers do.
10:22We have a bit of an unfair advantage
10:23having industry solutions, so you know,
10:26whether it's in the hospitality, or in retail,
10:28or in construction, or utilities, you know,
10:31we have a software stack that helps customers
10:33go and take that to market as well.
10:35Right, excellent.
10:36Gina, maybe you can talk about Kindle's experience with this,
10:40and I mean, do you agree with what Chris has said here
10:42about how to sort of identify these value-added cases?
10:47Yeah, you know, it's interesting,
10:49as I listen to the first three speakers, right,
10:51it sort of all comes in together.
10:54I mean, the top line, the bottom line,
10:56the productivity, the innovation,
10:58and Kindle, for those of you who don't know Kindle,
11:01we are two and a half years old,
11:02we are a spin-off from IBM,
11:04and I just want to add on how we look at use cases
11:07internally as customer zero.
11:09So Kindle now, after the spin-off,
11:12we are a technology services company,
11:15but we had something, we had an intellectual property
11:18in an area that's pretty unique, right?
11:20So coming from IBM, 30 years of managing MNCs,
11:23global companies, banks, insurance, airlines,
11:26managing the IT estate, we have accumulated
11:29a lot of IT insights and patents.
11:32We've developed 6,000 automation playbooks.
11:35So how do we then use AI to monetize
11:38on an advantage that we've got?
11:40So we built something called Kindle Bridge.
11:43So Kindle Bridge, think of it
11:44as an open integration platform,
11:46enterprise-grade platform as a service,
11:49interconnected with the customer's environment,
11:51monitoring system, tools, and so on and so forth.
11:53So it's an entire AI ops platform,
11:57and in there, we've integrated AI at scale
12:01within Kindle Bridge, okay?
12:03So we can do things like outage prediction,
12:07anomaly detection, change risk analysis,
12:10and so on and so forth.
12:12But if you think beyond how we monetize Kindle Bridge,
12:15which, by the way, we've deployed
12:16in 1,200 enterprise customers today,
12:19so it has really helped them to gain insights
12:21that they never got before.
12:22They were able to reduce outages, tickets,
12:27in advance, all the preventative stuff, right?
12:30But if you think the next step
12:31beyond this AI ops monetization, right?
12:35Because it's an integrated platform,
12:37we integrate with many different AI technologies
12:40and ecosystem alliances.
12:42You know, by the way, we are partners, too.
12:44We are partners.
12:45And so if you think about AI,
12:48AI, first of all, is not going to,
12:50you can't realize the true value
12:52with a few technologies or a few companies.
12:55It takes the whole village.
12:56It takes the whole ecosystem, right,
12:58to get the best out of it.
13:00So I think as we advance through this journey
13:03on Kindle Bridge with our customers and partners,
13:06I think there's a lot more that we can monetize
13:08and bring to upsell, cross-selling to the customers.
13:12Interesting.
13:13Why don't we go out to the audience,
13:14if anyone has sort of questions or comments at this point.
13:17If not, I've got lots of more questions
13:19I can ask our panelists.
13:21But I want to sort of integrate this into the room.
13:23So if you have a question, please raise your hand
13:27and we'll get a mic to you.
13:28If not, I will carry on.
13:29I'm not seeing any right away,
13:30so I'm going to carry on with questions.
13:32Oh, there's one there.
13:33Actually, there's one here.
13:34Hang on, let's get a mic to you.
13:36You can state your name and where you're with.
13:39Awesome.
13:40Hi, I'm Serena.
13:41I'm the founder of Fuzzy Sequence.
13:43I'd love to know from your experience
13:45what are the fastest applications of AI right now
13:49and which use cases have been having the most friction.
13:53So I'd assume the ones with more data points
13:56requiring more precision, taking longer to adopt.
14:00But what have been the fastest?
14:01I think I'm more interested in that one.
14:02Yeah, so what's been the sort of fastest?
14:05Fastest rate of adoption for use cases in AI in enterprises.
14:09Yeah, so a few things in enterprises.
14:11The first one is chatbot.
14:13So it's very easy to build chatbot.
14:15I met some companies that chatbots are not built by IT
14:19or third-party vendors.
14:21In the organization, the users themselves
14:24are building chatbots through a copilot studio.
14:29So live example of chatbot.
14:31That's the easiest.
14:32The second one, I'm talking about Gen AI at the moment.
14:36So the second one, which is a very typical use case,
14:39is cognitive search.
14:41Just imagine you have very structured documents
14:44in your organization, a knowledge base,
14:47which is very common to every industry, every company.
14:50And you have business problem to solve,
14:54which is you might find the right information
14:57in the right document.
14:58Typically, it take hours, days, months
15:01for people to do this.
15:02Through just using general AI capability,
15:07you can get that information in a few seconds.
15:10Those are typical use case.
15:11I think the third use case I would say
15:14is around at enterprise level
15:17is basically automation of some of the summarization.
15:22I'll give you example.
15:24I think one of the local companies
15:25actually recently announced their result.
15:28And they said that result was written by Gen AI.
15:32So summarization is another, again,
15:35is another typical use case,
15:37which doesn't cost money at all.
15:39You can train your own organization,
15:41your own people start to do that very easily.
15:44Right, Chris, do you agree with those
15:45as sort of the fastest implementation ones?
15:48I think those certainly got a place.
15:50And I'll add a couple to that.
15:51I'll add a couple where you can bring in.
15:54I think size and performance matters with AI, right?
15:57So what does that mean?
15:58So it's gotta be ubiquitous.
15:59So large language models,
16:00we've heard a lot about large language models,
16:02but what about small language models?
16:04Because small language models are the ones
16:06that can come all the way to the very edge.
16:08So I think we're starting to see use cases
16:10where people are taking large language models
16:12to get the vocabulary,
16:15but then quantizing them and summarizing them
16:18and bringing them down to a small form factor.
16:21I'll give you an example of a startup in Australia
16:22that's working with us at the moment
16:24where you cough into the phone and you get a diagnosis
16:28of whether or not you may have an infection.
16:30Now, that data is now private, it's resident.
16:32We've taken away some of the adoption inhibitors
16:35of moving my data to the cloud
16:36because the data's private and result.
16:38So I think if you can bring AI to the data,
16:43to the edge, to the end, and to do that,
16:45you need small language models, not large language models,
16:48and you need to have that AI pervasively available
16:51wherever you need it to be.
16:52Now, those use cases, we're sort of flipping the pyramid,
16:56because there's a whole risk of adoption.
16:59If the risk is taken away,
17:02then you're gonna get rapid adoption.
17:03So we're seeing examples like that
17:06of taking smaller models to the edge and turning innovation.
17:10But you have to make sure the capability
17:11is matched to the use, right?
17:14Because you still have this issue about,
17:16that sort of takes care of the data privacy issue,
17:18which might be a bottleneck.
17:19Doesn't necessarily take care of the reliability issue.
17:22Correct, correct, correct.
17:24There are different issues to solve for, absolutely.
17:28Does anyone else want to get in on this one about,
17:29Samin, you look like you have something to say
17:31on fast adoption, or maybe the opposite.
17:34She was asking where the friction points are,
17:36what are you seeing there?
17:37So if you look at what's happening on a macro level,
17:41the people who are adopting violently generative AI
17:45are the verticals that do not look at cost
17:48as the main pain point of adoption, right?
17:51So you look at two specific things.
17:52Things that are happening in the pharmaceutical,
17:54computational biology sector,
17:56and secondarily, a lot of defense companies
17:58that are really, really adopting generative AI.
18:02So a lot of interesting things that are happening,
18:04especially with companies like Palantir
18:06and how they're utilizing LLMs
18:08and different forms of transformers to further productize,
18:11whether it's strategy development
18:13or general collateral development,
18:15to computational biology,
18:16where you see guys like Evolutionary Scale.
18:19They built their own protein folding,
18:22a systemic computational biology system
18:24to build really, really interesting molecule formats
18:28so that they can be produced and they can be beta tested.
18:31So we've seen massive amounts of adoption there,
18:33specifically because the potential outputs
18:36far outweigh the cost of whether or not
18:39it's one token for one cent or one token for $10.
18:43Interesting.
18:44And Gina, are you seeing the same sorts of trends?
18:46I would say, with our customers,
18:48I think one of the early pilot POC use cases
18:51that they will choose relates to the contact center area,
18:55customer service contact center area, right?
18:57And they will start with typically emails
18:59before we do the phone calls.
19:00So on the email, again, it's about how you use AI
19:04to analyze the email content that's coming through,
19:07provide the CAN template,
19:09the agent just sort of scan it through,
19:11and if it's almost right,
19:12they can send the AI response, you know,
19:14and so on and so forth.
19:16Eventually, we think about the rest of it
19:18that comes through, right?
19:18Sentiment analysis, you know,
19:21and helping the core agent search for the information
19:26and have the suggested response
19:29in advance of the agent responding.
19:31All of this eventually will lead to a reduction
19:35in the call center agents
19:36because you don't need as many agents
19:38to process the same volume of emails and phone calls.
19:41So I guess that's one of those that people see the impact
19:45in terms of the ROI, potential ROI for AI.
19:48Now, in terms of friction,
19:49you mentioned a really interesting one.
19:51You know, just think about Co-Pilot, for example.
19:53You know, today, a lot of us are using Co-Pilot, right?
19:56We are probably one of the largest GSI
19:58around customer zero using Co-Pilot.
20:02But when I talk to customers,
20:03customers are still nervous about the data
20:06that are being analyzed, right?
20:08Because it's within your SharePoint,
20:09within your repository.
20:10Some of us may store things
20:12and may not have encrypted those data.
20:14We may not have classified those data
20:16as confidential data within the company's definition
20:18and so on and so forth.
20:19So there's still a lot of anxiety
20:23around to what extent, where is AI crawling the data
20:27and where is it sharing and disseminating?
20:32So I think anything that relates
20:33to the HR, private salaries, for example.
20:36What if you are talking about people's salary
20:37in the companies?
20:38So there's still a lot of anxiety there.
20:41And that boils down to one of the very important things
20:45that me and Mel and a few of us,
20:46when we did the prep run, we were discussing, right?
20:48What are some of the preparatory work
20:50that you need to be ready for AI,
20:51which is your data, right?
20:54The underlying data infrastructure
20:56from strategy to governance to classification,
21:00management, processing, and so on and so forth.
21:02So I think once we solve more and more of those pieces
21:06based on whatever guidelines
21:07that the world has been coming up with
21:09on GDPR and data privacy, PDPAs,
21:12I think people will become more and more comfortable
21:15adopting AI.
21:16It's just like cloud, if you think about cloud
21:18in two decades ago, right?
21:20People are not necessarily rushing to the cloud
21:22if they're not an early adopter.
21:24There's a lot of nervousness around security compliance.
21:27It's the same now with the AI era.
21:30It's just education comes a long way.
21:34Regulatory compliance comes a long way
21:36to help pave the way.
21:38Right, I want to get to other questions.
21:39Who else has questions in the audience?
21:42Please raise your hand and we'll get a mic to you.
21:44If not, I'll carry on asking some questions.
21:47I know, Samin, we were talking in a call
21:50we had preparing for this session
21:52a bit about this issue of data
21:53and where are people getting enough data.
21:55And there's been this big move
21:56towards people adopting synthetic data
21:58as a solution to maybe not having the right data
22:02or data privacy being concerned.
22:04Well, if we create synthetic data,
22:06then we don't have to worry about data privacy.
22:08And you were saying, actually,
22:09we have to be pretty careful about this.
22:11There's been a lot of talk, I know, this past week
22:13in sort of industry circles about this Nature paper
22:16that came out about if you use too much synthetic data
22:19over several generations of iterations,
22:22you get model collapse.
22:23And how do we avoid that?
22:25Is that a concern?
22:27Maybe you could talk a bit more about that.
22:28And then how are you seeing some of the companies
22:30you're working with maybe try to deal with this issue
22:34of can we use synthetic data?
22:36If so, how much and how do we use it?
22:38Yeah, I mean, the paper on Nature is really interesting
22:41because it's telling you if you give suboptimal data
22:44that's synthetic to train an LLM,
22:47what is the eventuality of that LLM or that transformer?
22:52And they're saying that it starts spitting up garbage,
22:55hallucination ratios go up to 80 to 90% of the responses.
22:58It's a very extreme case that Nature actually showed.
23:02What's interesting here is that it is,
23:04it actually emphasizes the importance
23:06of why human-in-the-loop model is that much more important
23:10because we're gonna deal and we're gonna go across
23:12a situation in the next 10 years
23:15where data becomes much more rarefied
23:17and it becomes a commodity, right?
23:19They talk about how there's like C4 buckets,
23:21C3 buckets of data, ultra high quality data.
23:25About 25 to 35% of that data already right now
23:28is copyrighted in the world.
23:30This is why Japan, for example,
23:31is playing a specific geopolitical game
23:35where they're allowing a large breadth of freedom
23:37for companies to train on IP and data in Japan
23:41without being persecuted for IP or any penalties, right?
23:44If you look at it on a longer duration scale,
23:47what we need to understand is the actual efficiencies
23:51that are gained with just a general model
23:53that does not need more and more data and context.
23:56But in the end, we're gonna have to come up
23:57with a better business model that derives the actuality,
24:01but derives the efficiencies from maybe taking
24:04three or four pieces instead of 10 pieces of data
24:07and still coming up with the right output.
24:09That's interesting.
24:10And Chris, I know when we were speaking just earlier,
24:11you had an interesting point about the potential
24:14to create sort of a higher value use case around AI
24:17with using generative AI to create synthetic data
24:19that you then feed to a predictive AI system,
24:22potentially be able to make some predictions
24:24that you couldn't make otherwise.
24:26Can you talk a little bit about that?
24:27Yeah, you think about some of the use cases, right?
24:29So preventative maintenance,
24:32it's preventative, it's very predictable.
24:33But how would you use generative AI
24:35in a very predictable preventative maintenance scenario?
24:38Well, preventative maintenance has been typically geared
24:41based on historical data of the failure of components
24:44in that chain.
24:46But what about components that don't fail
24:48and you don't have data on that?
24:49So how do you do predictive maintenance on that?
24:53So by spinning up generative AI models
24:56and an element of synthetic data
24:59and modeling effectively a digital twin,
25:01you're able to then check real-time run status
25:04against what the generative model is saying.
25:06And then, I think it's very important
25:07what Sangmin said around the feedback loop.
25:09I think it's really important for us to make sure
25:12that it's not just all synthetic
25:14and that there is a feedback loop between that.
25:16And so we're seeing cases like that
25:18in preventative maintenance, for example,
25:21where we're using, it's been around for a long time,
25:24anomaly detection's been around for a long time,
25:26but using generative AI to make that even more efficient.
25:30And I think we're seeing a lot more
25:32of these sort of hybrid systems
25:33where you use maybe a large language model
25:36to classify or parse something on the front end
25:39and then feed it to a predictive system
25:41or even a hard-coded system.
25:44But I don't know, are you seeing more of that?
25:47Mao, do you see that?
25:48Yeah, definitely.
25:49I use a live example, let's say the consumer business,
25:53which I'm very familiar with.
25:54And one of the challenges of a lot of large consumer company
25:58and retail companies is that inventory.
26:01To maintain the service level,
26:02they keep a very large inventory in all their warehouses.
26:05This includes Walmart, Tesco, everyone.
26:08So now AI has been used to predict
26:12the inventory level at each store.
26:15So the data collected are the sales data.
26:19The data collected are the predictive forecasting
26:23around the demand of the consumers of the e-commerce.
26:28And data has been used,
26:29this is a typical AI use case, right?
26:31So through AI adoption, the companies are able
26:35to predict the inventory at the store level
26:39and eventually lead to a forecast
26:41of their warehouse inventory.
26:43That tremendously reduced the inventory level.
26:46I saw a company reduce 30 million US dollar inventory
26:50in the US by doing this alone.
26:52But now with generative AI,
26:54they actually add narratives on top of that
26:57to have a summary to tell the whole leaders,
27:00tell the leaders of the whole supply chain
27:03what they need to do in order to fulfill the demand planning.
27:06How are they going to do,
27:08they use generative AI to summarize
27:10the new process of demand forecasting.
27:13So this is live already.
27:14I'm not talking about something fresh new.
27:16This is definitely live in a lot of companies.
27:18So it's taking that predictive data,
27:20then creating a kind of narrative story around it
27:22that can be used to convince managers
27:23that they should actually do this.
27:24It becomes a tool for internal communication
27:26more effectively of the data
27:28that the predictive system's generating.
27:30That's interesting.
27:31If I may pick on something you said as well earlier around,
27:33there's no one model that fits all.
27:35And I think we're seeing more and more
27:37a chain of models effectively,
27:39sort of picking things up.
27:41So you're gonna have a generative,
27:42a large language feeding into generative.
27:44And we see this, for example,
27:45in use cases in construction sites.
27:47We have industry solutions for construction sites
27:50where you're taking safety procedure manuals,
27:52which is very structured,
27:53but you're also taking audio cues,
27:55the forklift that's moving around
27:57and people's got ear muffs on
27:59and they can't hear the forklift,
28:01and taking video cues,
28:02and taking this multi-models and feeding them
28:05and providing real-time safety information
28:07to say, hey, duck out of the way,
28:08there's a forklift coming your way.
28:09Wow.
28:10Right?
28:11So multi-model, chains of models,
28:13generative, predictive working together.
28:15That's where we see it going.
28:17But this is a bit what Mia was saying
28:18about creating digital products around each solution,
28:22but then you get into this issue,
28:23do you have the talent to actually do that?
28:24It becomes more,
28:26when you start talking about chaining models together
28:28and then maybe also having different UXs on that
28:31and different code
28:32that actually would couple those systems together,
28:34it becomes more of engineering lift.
28:36Samin, I know you said that there's an issue about talent,
28:38really, to do this within companies.
28:40I don't know if you can talk a little bit
28:41about kind of the war for talent
28:42and how you're seeing that shake out.
28:44I mean, there's maybe three pillars
28:45that we could talk about that's really interesting
28:47and maybe thought-provoking.
28:49In the past eight to nine months,
28:51the mega cap tech companies have fired
28:53about 110,000 people.
28:55Now, about 30 to as much as 60% of that
28:58is attributable to what can you do
29:00with less people with generative AI.
29:03At the second time, when you think about engineering
29:05and we think about implementing new solutions
29:08so that the larger organization can use it,
29:11this is definitely a talent issue
29:13of does the actual engineering manager
29:15or the VP understand how to actually produce this
29:18with the data privacies in mind, so on and so forth,
29:21actually do this so that there's less hallucinations?
29:24But also, there's a different aspect of this
29:26where it's about culture, right?
29:29Does your 50 to 60-year-old engineering VP,
29:32are they comfortable downsizing
29:33their engineering team by half
29:35because they know that the copilot will enable them
29:38to actually utilize the engineering team
29:41and to be extra productivity?
29:43So there's a war on talent
29:46and there's a war on generational culture
29:47that needs to be actually very much underlined
29:51and really, everyone needs to underwrite that process
29:54in a very predictive and successful way.
29:57Interesting.
29:58Other comments or questions from the audience?
30:00There's one there, if we can get a mic to that gentleman.
30:03If you can state your name and where you're from.
30:04Hi, I'm Andrea from Zühlke Engineering.
30:07Thanks a lot for your insights, it's very interesting.
30:10I'm interested to get your take on build versus buy.
30:13So at the moment, we see many new products
30:15on Gen AI or AI coming up.
30:18Many companies have similar problems to solve
30:21so this could also be a signal
30:22that there's more products to come.
30:25So how should company go about it today?
30:27How should companies go about it today
30:29in deciding whether to build something for themselves
30:32or actually buy or maybe even wait for a product?
30:35Right.
30:36Build versus buy, yeah.
30:37We start with Gina, actually,
30:38because she's nodding at the end and then we can come down.
30:40It's a very good question and so pragmatic, right?
30:44I think as part of, if you think about the best practices
30:46in ROI for Gen AI, right, one of the things
30:50is to make sure that you start
30:52with an initial cost assessment
30:53of how much you're gonna spend,
30:54your hardware, software, your infrastructure,
30:57and so on and so forth.
30:58But at some point, you will need to test
31:01and as part of the implementation strategy,
31:03you start small with pilots
31:04and to test the different technologies
31:06out there in the market, right?
31:07And then you can figure out later on
31:09how whether you're gonna build it
31:11or whether you're gonna continue to leverage
31:12the open-source or the off-the-shelf tools, right?
31:18When it comes to eventually the LLM,
31:20I think at this point, I think it's still very costly
31:23for a lot of companies to try
31:24and build their own LLM models.
31:26So they're still relying a lot
31:27on whatever that's out there.
31:30Of course, if you wanna be cost-savings,
31:33you start looking at some of the open-source tools
31:35and technologies, but you've gotta equally make sure
31:37that you've got the talent and the people behind
31:40who understands how to use, implement, and operate.
31:45But you raise a very interesting point,
31:46which is collaboration, right?
31:48Because if you think about the like-minded industry,
31:53network companies, can you come together
31:55to collaborate on solving similar problems
31:59where you can then share the cost of the whole setup
32:02and set up like a mini-COE equivalent, right?
32:05Partnerships is another one, right?
32:07Partnering with the likes of Microsoft, Google, Amazon,
32:11and other technology partners where they are out there
32:13doing a lot of trailblazers on AI today, right?
32:16So you can be there where they will naturally give you
32:21the right expertise, they are architects,
32:22they are data scientists, they will give you
32:25maybe even a sandbox environment
32:27to test your early pilots, right?
32:29And give you free training potentially to your people.
32:31Chris, build or buy?
32:33Yeah, build or buy.
32:34I'm gonna say both, that's a cheating answer.
32:37That's a cheating answer, but let me get,
32:38there's a couple of litmus tests I use
32:40for both of these things.
32:42I think you should let what's available,
32:44what's going to be sort of off the shelf
32:47and horizontal efficiencies, I would buy those,
32:50I wouldn't build those, right?
32:51And if you look at it, if you're doing PDF summarization,
32:55every PDF reader now is gonna build in summarization for you
32:59so then you look at things like in horizontal applications,
33:02we have SAS suites for human capital management.
33:04So things like job applications, processing them,
33:07resume summarization, job descriptions,
33:10performance appraisals, well all of those things
33:12are horizontal, company A can do it, company B can do it.
33:15So don't build it, just buy it, just buy it.
33:18I think where the value's gonna come
33:21is when you bring your data into it.
33:22And I think that's kind of where we're focused on.
33:25We're really focused to help customers
33:27bring AI to the data.
33:30And so tool sets become important,
33:33skills become important.
33:34And look, listen, your spoken language
33:36is now your programming language, right?
33:38Because you can speak to your assistant.
33:40And now imagine if that assistant and that AI
33:43already respects the security and privacy
33:46that you've got built into your existing data sources.
33:49So you're not picking your data and going elsewhere,
33:51you're leaving it here and you're putting an agent
33:53on top of that that respects that.
33:55That's what we're focused on
33:56and that's where we think you can build
33:58and you can sort of cross the chasm
33:59a little bit quicker on the skills gap
34:02because you're talking to it but you're respecting
34:04what you have already in place.
34:05That's where I would build.
34:07Interesting, and I'll come back to the other panels
34:09but I know there's some other questions out here.
34:10I saw at least one, maybe another one there.
34:12Let's go to Alex here, I saw him next
34:14and then maybe the one there.
34:17So this is gonna be a tricky question
34:21and it's actually two questions.
34:24So many of us here live through the dot-com bubble
34:28and we remember 2001 very clearly.
34:31I remember it like yesterday.
34:34And even before 9-11 when everything collapsed, right?
34:39And we remember the huge hopes
34:42that the internet was promising, right?
34:44And people were deploying infrastructure like crazy
34:48and much of that infrastructure actually
34:49was deployed by startups and provided by startups.
34:53Now, generative AI provides us with even a bigger promise.
34:59My first paper on generative AI was 2016,
35:03the many, many papers, 2017.
35:05So now what we see is hyper-investment in GPU.
35:10So NVIDIA is $3 trillion, right?
35:13Close to $3 trillion.
35:15It's actually more than the size
35:16of the pharmaceutical industry, pretty much.
35:19If you think about it, just think about that, right?
35:22And now many of those companies, the big seven,
35:26they are part of the S&P, right?
35:28So it means that all of your savings
35:30are also going there as well, or some of your savings.
35:36Where are we in the hype cycle?
35:40Is $3 trillion NVIDIA market cap justified?
35:47If we are kind of closer to the decline
35:50when everybody has deployed the equipment
35:52and now they don't have the money
35:54to buy the next generation,
35:57and how is it going to affect your companies,
36:00especially like Oracle, Kindrel?
36:02All right, Kindrel, Oracle, you guys are on the hook here.
36:05Are we at the peak of the hype cycle?
36:07Are we still on our way up the climb?
36:09Oh, Gina's not going to take the bait on that one.
36:12All right, Chris, what do you think?
36:13Look, I'll go for it, why not?
36:16So I think there's a lot of focus, like I said,
36:21on the large language modules and big spending on that.
36:24But if you then look at it,
36:26we're slightly different with the internet.
36:27The internet was when you built a business model,
36:30that business model was hosted somewhere
36:32and everybody in the world connected to it, right?
36:35And it was in a particular market then.
36:37Well, if AI is going to be pervasive,
36:40which is, that's what we're all talking about here,
36:42it's going to be at the edge, it's going to be everywhere.
36:44And I think where the market's going to flip
36:46is going to flip from training to inferencing.
36:49And inferencing is going to drive a big portion of it.
36:52So every single one of us using, engaging,
36:54interacting with AI is going to be the consumer.
36:57And it's not going to go to central service somewhere,
36:59it's going to be at the edge in our market somewhere here.
37:02So I don't think the infrastructure
37:04that we're deploying in the market here
37:05is going to go to waste,
37:06because the inferencing engine's going to pick up
37:09if we get this right.
37:10The second thing I'd say is, you know, let's look at us.
37:13We've got two-thirds of the world's population,
37:15and I cover Japan and Asia Pacific.
37:17I've got two-thirds of the world's population here.
37:19If two-thirds of the world's population
37:20is going to start inferencing and using AI,
37:22I think, you know, I don't think we're at the cusp
37:26of going down.
37:27I think there's a huge uptake that's going to use.
37:31We're still going up.
37:32All right, Miao, what do you think?
37:34And I'll get to Samin, and then we're going to have to wrap up.
37:36I think it's more sophisticated and complex
37:37than just saying hype and AI, AI,
37:40just read the flags of AI.
37:42I think one of the challenge, yes, AI is a start.
37:45By the way, I think with Gen AI,
37:48this is a journey, starting journey of adoption.
37:51However, there are a few challenges.
37:53If you look at what happened in the U.S. stock market,
37:55S&P yesterday, all the Magneto 7 jumped.
37:59And one of the things that despite Microsoft
38:01is doing very well in Q4, their Q4,
38:05and still, they did not meet the expectation of the investor
38:08because investors has much higher expectation on AI
38:13than the reality.
38:14We have to meet the fact, right?
38:16So that's point one.
38:18Point two, I think what is missing,
38:20despite all the startups, the big seven,
38:23the Microsoft, Google, Oracle world,
38:26there's a foundational modeling of IOM model.
38:30I think still application side,
38:33adoption side of AI is lacking of a few things.
38:36First, scalable, industrialized solution.
38:41Nothing is, so you just imagine,
38:43there has to be something like SAP AI,
38:47code and code SAP AI, code code SAP Oracle has to happen.
38:51Don't mention SAP, they're not on the list.
38:53I don't mind, by the way, I think neutral, no wonder.
38:56Just use that word, right?
38:57There's nothing big enough to generate the next demand.
39:02So I think the next way doing things of AI
39:06is to industrialized AI solution,
39:10end-to-end, scalable, directly relevant to the business
39:15will take off.
39:17Only that day comes and AI will come.
39:19Samin, where are we on the hype cycle?
39:21Yeah, so I think it's better
39:23to approach this macroeconomically.
39:25If you look at what's happening right now,
39:28the S&P 500 actually dropping has more to do
39:31with the yen deleveraging and the yen carry trade
39:34deleveraging in the past three or four days
39:37than that of retail actually selling out or degrossing.
39:40Look at your phones right now,
39:42the BOJ just raised the interest rates of the BOJ
39:45by 25 basis points.
39:46Ah, one person doing it, thank you very much.
39:49What's gonna happen right now
39:50is for the next three or four more days,
39:52everyone's gonna cycle out of growth stocks,
39:54which is basically the mega cap stocks.
39:57Hype and bear case is defined
39:59by what the stock market is doing right now.
40:0179% of the world's stock markets, S&P 500 mega caps,
40:06have reported earnings
40:07that have exceeded analyst expectations.
40:10This means that as long as the money is not being spent,
40:14that does not go lower
40:15than the expectation of the analysts.
40:18There will be less of a bear case
40:19and more of a bull case for these markets.
40:22You have to understand the larger picture
40:24of what's going on,
40:25because if that large picture defines
40:28the actual trade drives and traction.
40:31Interesting.
40:32On that point, I'm afraid I'm gonna have to wrap up.
40:34I know there's still questions in the room.
40:36We have some more time for you guys to circulate,
40:38network, before we have to be back at the main stage.
40:42But I wanna thank my panelists
40:43and thank you all for listening.
40:44And please continue to enjoy some dessert.
40:47You can come up and talk to the panelists individually.
40:49Hopefully you can get your questions answered.
40:50I'm sorry we weren't able to get to all of them
40:52in the time we had, but thank you very much.
40:54Thank you all for coming.