Intel Corporation (INTC) Earnings Call Transcript & Summary

Hi, everyone. Thanks for coming to our GenAI monthly webinar series. My name is Sancha Norris with Intel AI Marketing. Real quick, first a few housekeeping items because we get these questions. The presentation will be about 40 minutes, and then we'll take Q&A afterwards. The PowerPoint slides will be made available. So be sure to look at [Audio Gap] in your follow-up e-mails and go ahead and ask questions throughout the presentation, we'll answer as many of them as time allows, but also try to be complete with your questions so that we make sure we're answering them appropriately. Today, we have 2 exciting speakers with us, and they're going to share their real-world experience with GenAI with you today. With us today is Brian Sathianathan. Brian is the Co-Founder and CTO of Iterate.ai. He leads the development of Iterate's enterprise AI platform called Interplay. Brian is serial entrepreneur. He founded Avot Media, which is a video transcoding platform used by Warner Bros, which Smith Micro then later acquired. He transitioned then to investments, contributed to 13 investments and acquisitions at Turner Media. Welcome, Brian. And we also have Ekta Chopra. She's the Chief Digital Officer of e.l.f. Beauty. Ekta brings over 25 years of technology experience from private equity and retail companies. Ekta joined e.l.f. Beauty in 2016 as the Chief Digital Officer, and she spearheads the digital transformation at e.l.f. Beauty that covers e-commerce, engineering, data ecosystems, security, metaverse and then, of course, AI. Welcome, Ekta. So we're going leading off today with Brian and Ekta, and then we'll proceed with the Q&A. So take it away, Brian.

Thank you, Sancha, and thanks for everyone being here. We are very excited. So let's quickly get started. So today's topic is the Beauty of GenAI for Retail. So let me -- before we go into the topic, let me quickly give you a little bit of an introduction to Iterate and what we do, and then we can quickly go into the topic. Next slide. We are an AI application platform like Sancha mentioned. We've been around for a while doing a lot of AI work, and we've deployed AI at scale. What we've done is we've done AI for data analytics, we do AI for computer vision, and now a lot of work in generative AI, and these are, of course, optimizing the Intel platform. We have over 4,100 edge deployments for AI across 12 countries. And we also have a lot of web presence as well. Customers use our applications to build capabilities, AI applications for end consumers. And we -- the applications built by the customers, we serve over 40 million monthly uniques on application. So we've scaled AI. Now we are doing a lot of work with Intel and our customers on the generative AI space. We have -- as you can see, this is some of our sample client portfolios. We work with very large corporations. Everything that we do is built on top of our Interplay platform. Our Interplay platform is a drag-and-drop AI low-code platform that has 2 sides to it. At the very top, you can begin to build applications. We build applications solutions that's built on our platform has been serving across many sectors, banking, retail, QSR, automotive. And then underneath the platform, we actually optimize it for in the Intel technology stack. So we have integrations with OpenVINO and then we have integrations with various Intel SDKs underneath the covers, right? So we could run at a higher performance. As you know, this is -- these last 2 years have been an amazing years with so much going on in AI, specifically with generative AI. It's been a year of growth, right? This is a survey that was done by -- from American Institute. Really interesting part here is that they surveyed retail leaders and technology executives. And it's -- 79% of them has at least used generative AI once. And 35% of them use it very frequently, right? They have a regular use. And then the chart next to me on the right hand -- next to you on the right hand actually shows you like are they using it at work versus how they're using it at the home. With the generative AI technologies, people -- the adoption is amazing. As you know, ChatGPT was one of the fastest-growing apps and had the fastest adoption compared to traditional social media, right? So this is something that is proliferating really fast among consumers as well as among enterprise users. So this is the -- so there is so much happening. And one thing that was really exciting is last year, I was talking to several senior leaders, Board members of various companies. And last year was the year where corporations were thinking about and figuring out what generative AI. And this year, based on all what I've seen, it's sort of the year for action. So a lot of new initiatives, prototypes, all that stuff is being built in generative AI. So this webinar is very timely because we're going to be talking about things that we have done with our customers as well as what Intel is doing in this space. What's the market saying, right? A couple of really interesting thing about generative AI is like everyone wants to be on it, no one wants to be left behind, right? And they also want to use generative AI in a very powerful way to increase the end customer experience, right, and also to create internal efficiencies. In fact, we'll talk a lot about internal efficiencies today, how to create internal efficiencies as well as increase better experience, right? And of course, every leader is like where do I start, right? And like what's the right thing to do? And what we are also seeing from a needs perspective is there is a lot of needs. One of the things with generative AI as it's evolving really rapidly, at times, accuracy has been a problem depending on what your use case is, right? So accuracy is definitely something critical responsible AI. You want the AI to provide responsible answers as well as provide results that are in a very responsible manner and also the ability to kind of bring things to market very quickly because still, people are figuring out how to productionalize generative AI, right? There's a lot of experiments going on, but how do you productionalize this thing at scale? Compute is really another big challenge because as you need more GPUs, as you want to scale these things, whether you are working in SLMs, smaller models versus large LLMs, the paradigms are very different. Especially one of the things within traditional companies, companies whose core is not -- the core product is not technology, they are used to like using CPUs and existing platforms. How do you actually use existing platforms, but upgrade them to a point where you can very quickly roll out large language model applications and then not go through that big learning curve, how do you actually simplify the learning curve. So these are all -- and cost as you know is a critical factor. So these are all things that leaders are right now talking about and a lot of leaders that Iterate has spoken to, customers we've spoken to are all thinking about. So these are things that you would want to think about as a leader in your organization when you're thinking about what my next step on an AI, generative AI project is? These are things you want to consider, right? The question is like, we get this all the time, right? Where do I get started? What do I do on Monday, right? Well, I think it all depends on use cases because don't think of generative AI and AI as a shiny object as a toy. But think of it like something that's very practical for you and that can change and can ripple through your organization. So think about use cases, look at the business problems that are very applicable and where do I apply AI there, right? When you think about the business problems, there are 2 or 3 paradigms, we'll cover them in the next couple of slides. But the idea is to think about like if I apply AI for this given business problem, along with it solving the problem, what sort of intellectual property value is it creating, right? Is this very core or very strategic to what I'm trying to do, right? Does my team have the skills and knowledge, right? And then how do I -- how customized is the problem? Can I just buy something off the shelf and get it solved, right? And also, over time, you want to begin to measure return on investment on it, right? Is it measurable? So those are some of the things at a higher level as a leader when you are considering generative AI you want to be thinking about, right? And this is like a very high-level frameworks that we apply and then we help our customers and our leaders think through this sort of a framework, right? As you know, generative AI is not just ChatGPT or just LLM or textual content, right? Generative AI is actually fully multimodal. You can use generative AI for text, you can use it for code, you can generate images with it, you can apply it on speech recognition models, you can even generate music with it, right? Nowadays, there's a lot of video generation content going on. There is 3D, right? And there are so many use cases. I mean we even have really exciting use cases where we are applying generative AI LLMs to extract document content right? So there's so many different type of embeddings and so many types of exciting applications that one can do with generative AI, right? So typically, in an organization like if you look at text, there are so many text-oriented applications. This slide -- I'm not going to go through the entire slide because it's very detailed. But when we send it to you, you can look at it off-line. Text-oriented application all the way from document standardization, content extraction to chat, all those applications are powerful, right? And the beauty of generative AI, unlike the other technologies, it spans across your entire organization. It's something that IT can use. It is something that marketing can use. It's something legal can use. It's something operations can use. There is something for everyone, which is one of the really exciting things about this whole generative AI tech, right? At a higher level, I want to sort of create some buckets that you can think about, right? So if you're thinking about generative AI projects within your organization, there are sort of 3 high-level buckets. One, the first bucket is sort of called the general purpose bucket, right? The general purpose bucket is, you want to enable chatGPT or like conversational capabilities for your internal employees. So that's a very -- I mean, fairly simple capability. You could actually get third-party applications from outside. You could roll out your own solution. There are so many ways to do that. Then there are these things where you want to use generative AI within your company, but you want to customize it for some of your needs, right? Like so you want to do your own RAG, your enterprise RAG for certain types of solutions, right? That involves taking a third party, but you need to do some engineering to customize it. Then the third one, which is the high strategic value where you want to fine-tune your LLM or you want to pre-train an LLM, prune and distill an LLM for whatever your use cases are, right? So those are like very high strategic value, right? In the next slide, we'll quickly look at a couple of use cases, right? I mean these are just sample. There could be more. I mean, this may not be applicable for you. You may have different use cases. But the idea is like, let's say, for example, you are a QSR, and you're looking at drive-through ordering, right? That's a use case. Now for a drive-through ordering, depending on the complexity of your menu, right, you may have thousands of permutations and combination. So you may not just be able to take something that's out of the shelf, you might have to work with the vendor or your engineers have to begin to tune certain things and customize it to get to the point where it becomes very accurate for you. That's why a lot of drive-through rollout projects take a while to get it right, right? And then the company -- the customer has to spend time working and modify, right? And in this scenario, the LLM choice is also interesting that you use larger LLMs versus that you use smaller LLMs because you are running -- you're doing processing on the edge, in the store, in the drive-through. There may not be Internet connectivity, you may not have cloud access, but how do you do edge processing, right, much more efficiently? So you want to think about LLMs like Phi-2, LlaMa and Mistral and other type of engines, right? So this is something you could -- from a buy-build equation, it's something that you probably buy, but work with the vendor to do significant customizations, right? Let's look at another use case. So you're using a ChatGPT-type capability where you want to roll out your own ChatGPT. There are a lot of tools in the industry. So you pick the vendor that can store your document privately and provide you the right security, but you don't need to do development yet. Maybe some configurations, but not real development. So this is a buy type of a solution, right? Let's say, you have some very ultra-private strategic documents and you don't -- you want to do some sort of security, some kind of customization or you want to do a scenario where say you're an insurance company and people are asking you insurance question and customer support, you want to take a policy number, look up internal databases, identify their transaction payment history, premium payment history and then look at the documents to look at their policy compliance and answer to respond back to the customer. You can -- those things you can buy an existing product, but you can customize with some integrations into our system, right? So that's a buy and a customized scenario, right? So then the other scenario is more strategic. Like let's say you're a beauty company and you want to build a beauty LLM where you've trained it on your own products, you trained it on your own ratings and reviews, you trained it on your own content or you're a bank or you're an investment bank or you're a trading company and you want to build a finance LLM, which is strategic to what you are doing, right? And it's a big part of your business or you're an insurance company, you want to build an actuary LLM, right? So those things are custom built and you probably want to do a lot more custom work. You may take a foundation model or an existing model and train them, right? So these are some examples. And in your industry and in your business, if you think through this type of framework, I'm sure you'll be able to match and map some of these examples that's applicable to you. One of the things that we provide is basically one of the -- on top of our Interplay is one of the applications called ZenPilot. This allows you to look at private documents within your company, and you can load them securely inside ZenPilot. And ZenPilot has the ability to do RAGs on them, right? Retrieval augmentation generation. So you can ask questions securely on your documents. The simple use cases you're an executive and you want to like search -- you have loaded all your strategic documents and you're trying to ask a question like based on the prior industry report, what are the 3 strategies I should look at? And it will look at hundreds of pages of industry analyst reports and come back to you things that are very customized to your needs, right? Or you're an engineer looking at some code that you upload, but you want to do some co-pilot experience, right. Same thing in across every industry. So this is the type of experiences like certain things, certain tools, you can buy off the shelf and begin to use. Some tools you need to customize. So you want to think about sort of the right tool for the right job, right? Because end of the day, the use case is critical and the business case is critical when you apply the technology around it, right? Having said that, I'm very excited to transfer it over to Ekta Chopra, Chief Digital Officer and an awesome digital leader. Over to you, Ekta.

Thank you, Brian. That was even very informative for me as well. So thank you for doing that. So first, I always like to ground people in who e.l.f. is? E.l.f. Beauty, it's a house of 5 brands, which is e.l.f. Cosmetics -- if you can go to the next slide right here, thank you. We are a house of 5 brands, e.l.f. Cosmetics, which is eyes, lips, face, which is our original brand, 20 years old. We just celebrated 20 years at NYSE as the #1 performing stock. And then e.l.f. Skin, which is our extension of our e.l.f. Cosmetics and then Keys Soulcare, which is inspired by Alicia Keys, and W3LL People, which is our clean beauty brand and Naturium, which is a brand we just acquired as well. So when I think about e.l.f., e.l.f. is a bold disruptor with a kind heart. We like to disrupt norm, shape culture and connect communities through positivity, inclusivity and accessibility. So if I think ground ourselves in that purpose, when I think of AI, I also have to make sure that it stays through to our mission, which is e.l.f. is for every eye, lip, face, paw and fin, which means that we're very inclusive. So when we think of sort of all the noise around AI and possible as ethics of AI, we want to make sure that we are not just picking up anything, we're building it very thoughtfully so we can serve our community very thoughtfully. So that was sort of the grounding in making sure that we're partnering with someone who can actually bring that lens to the table as well. If we can go to the next slide, please. Thank you. So if I think about sort of, as I just said, how will we bring our vision to life? We stand with every eye, lip, face, paw and fin, as I said. So when we build our AI ecosystem, there's 3 things that we want to ground ourselves in. We want to be transparent. What are we doing, making sure that the decisions that AI is making, whether it be personalization on the site, what sort of content is being served or what messaging is being served, you want to make sure that it is right for that member, for that consumer, and we're not just serving something that shouldn't be served. We want to be accountable. We want to do things ethically, and we want to be inclusive. And that's very, very important to us at e.l.f. Beauty, making sure that we are living to our purpose. So when I think of our vision for AI, we want to have them under these 3 guiding principles. Next slide. And if I -- e.l.f. Beauty is a -- it's a brand. And of course, when we think of like the consumer journey, you do have to go through many different decision points whether it be think of it like a marketing funnel, you have awareness, you have consideration, you have intent to purchase, and then there's a purchase. But then there's a post-purchase. And in our case, we have one of the best loyalty programs for e.l.f. Beauty with almost 4.5 million members. So loyalty is very important. And then advocacy to really make sure that, that consumer is going to come back and you're providing them a service that they become that e.l.f. fan and they've become that e.l.f. citizen who really, really loves your brand, right? So you have to really think through sort of that journey and decision points. And I think what we're building and how we are testing with AI, we will eventually impact every single sort of journey point from awareness all the way to advocacy. So we are testing products in many different ways to impact that full journey. Next slide, please. And if I think about our journey at e.l.f., we actually did a lot of research. And of course, we went -- we're working with it Iterate very closely. But the general themes that I saw in my peers at other CTOs, CDOs, really was there are kind of 4 barriers for adoption. Either you have too much data, you don't know what to do with it because you're swimming in data or you have no data and not enough data to train your sort of models to provide accurate information. There's lack of computing power, if I just think about what we are hosting all our models in AWS and our consumption cost has dramatically increased, right? But that's the cost of the computing power is, is still evolving. Then we have regulatory concerns, of course, especially with us, if we are going to start using AI for, say, detecting skin concerns or anything like that, like you have to face state laws, right? There's laws against like how you can store that consumer information and so forth. So there's like a lot of regulatory laws that we have to navigate. And then, of course, there's ethical concerns. As I said, we are e.l.f. is for every eye, lip, face, and it's very important that we take that very seriously. Consumers trusting us with their data. So what I don't want to do is serve them something that doesn't serve them as that individual consumer, right? So if somebody comes up certain shaded tone, we want to make sure whatever we serve is relevant information for them. So I would say that from our perspective, these were the 4 barriers of mass adoption that we are seeing, and we're also trying to navigate as well. Next slide. If I think about sort of our road map, it's grounded into 5 key principles. One is leveraging e.l.f. You, which actually is our internal learning management system, making sure that we are really upskilling. One of my goals is to make everyone a great prompt engineer because one of the biggest challenges when companies are trying to adopt AI is like how does your workforce think about it? Do they think their job is going to go away? Do they think like can you help them work with it better? Can you upskill people? And I think all of that starts with education. So it's really important. You're not just walking into this and saying, okay, I'm going to plug and play this AI thing. And everybody is just going to be like comfortable using it or I'm going to force people to use it. You have to educate them. So we have built courses or we're also building future courses that are focused on getting people comfortable with just working with AI. And I think that's a really, really important thing. So the foundations of what is AI? What is generative AI? And then just the basics of how do you write a great prompt to get the answer that you're looking for. The second is there's already tons of co-pilots, like even going out there and building your models, there's already co-pilots, Microsoft, whether it be Slack, whether it be Zoom, everyone is coming out with copilots. So enable them, start testing them, let's -- and start learning like how it works. Even the teams start feeling comfortable because they start seeing the value of it. And then, of course, there are specific things. As a public company, we have to be very, very cautious about sort of what are we training, where is it -- is it secured where training our models, like how are we protecting the data? So when you're building that private infrastructure, you just can't do that on your own, right? You really have to work with experts. And in our case, that expert really is Iterate, and they've been able to sort of really help and make sure that the infrastructure that we are building is really secured. We're taking into account sort of segregation of duties as well as like how people are accessing that data, who is accessing data, how we training that data, very important. And of course, as Brian rightly said, in his use cases, you can't just one-size-doesn't-fit-all. Not every single model fits every single use case or it's the right thing to do there. So you do have to do some testing and seeing which model probably works with what use case and really make sure that you're investing in the right road map there. And then, of course, AI governance and security, right? Like so we're building our AI policy right now because we have a lot of the Gen Z generation that's like, "Oh, my God, can I just go to ChatGPT and get an answer. But you can, but you can't put any sort of proprietary information in there, right? Look, can I just use this AI thing that I saw that's giving me all these answers? We can test it, we need to look at it from a security perspective and make sure that we can do that, and it is something that we feel comfortable with. So you do have to bring in some level of governance. So your workforce is not going out and selecting things to just test on their own. So this is like how we ground our 5 principles on how we're rolling out AI. Next slide, please. Of course, there's a lot of learning. But what I would say is ChatGPT, AI has been around forever. But ChatGPT really made it accessible from a user perspective, how quickly you can get answers. Now there is a trust component and not all information is correct. But that's -- but at least you're starting to see the power of it from a user perspective. So I do think that there are some things that we should sort of focus on. Of course, there's speed to market. Like there's always -- there are things that we are working on, which we won't talk about it here. But then there are things that I think people should be thinking about use cases. So speed to market is a really, really important component here. Working with trusted people. There's so much noise things that were just like a domain name. The people have added AI and now you're an AI company, right? And generative AI very different than just AI. When you think of personalization, right -- serving the right content to the right person and the right messaging, that's been there forever. You've probably heard that in every conference, right? But generative AI is different, right? So I think really making sure you understand the fundamentals of the differentiation between the 2. And then, of course, there's -- how do I make sure that when we do build, say, our version of e.l.f. GPT or whatever we're going to call it, it's serving the right information. It's going to help my teams. So if copy team is using it the right -- is it serving the right copy for the right e.l.f. products, right? So if we say, "Hey, give me a Halo Glow Liquid Filter, which happens to be one of our most popular products and write me a PDP information, right? It should be able to serve you with your brand tone, with your brand language. And that's a really, really important piece. The next is I go back to improving staff productivity, right? Like you really want to focus on when CFOs are thinking about AI in their mind, they're thinking what used to be called RPA, robotic process automation; business process engineering, which ultimately meant that you're saving dollars. What I will say is that if you go into this thinking, "Oh my God, I will save a lot of dollars, I won't have to hire any people." That's the wrong thinking. First, you have to educate yourself. You have to learn what the possibilities are and then build a strategy and then build like where cost-efficient things can come into play. And of course, you have to act responsibly. If you are a company that's handling consumer data, you have to think about data privacy. You have to think about like what is the impact and so forth. So you have to work very closely with the legal team. And for us, testing and learning is an e.l.f. Beauty's DNA. So for us, this is a test and learn. We haven't really rolled out anything at mass scale yet. We're still testing and learning. We're testing and learning with copilots. We're testing and learning with custom products, which I'm going to show you one of them. But then we are also testing and we're learning what's off the shelf that can help us. And we can get started with. So I would say that those are like really, really important things to consider as you are building your road map. Next slide, please. And if I think about sort of once again our implementation journey, we started with, you can't just like, "Hey, I want to build a custom model and now I'm going to do things with it." You have to think about the use case. So if I think about certain use cases. One was, hey, education is one. First, I need to make sure people are not scared to use it. And then the second is like, all right, what are the things that can give me an idea about the power of this. So one such use case was, currently, we have Instagram, we have TikTok, we have Facebook, we have a lot of great community managers. And we are the #1 favorite Gen Z brand. So we try to really listen to our community. We try to really respond to our community. And one of the things I was thinking about as I was thinking about use case is like how can I do that quicker? How can I make sure I have the most response rate for that community? How can I make it easy for the community managers? When they are like actually answering, maybe the answer is there, and maybe they're just tweaking the answer. I'm going to show you some of this stuff in a minute, but it was really about solving for a problem. What problem are we solving for, what is the use case and then how do we make sure we have the right people testing it to give us the right feedback and then how do we make sure that you're training the model correctly, right? So I think you have to take all of those things into consideration when you're determining a use case. And then there's a lot of great data scientists. Not all of them are AI engineers. There's a lot of great AI engineers, but they might not understand some of the stuff that goes into building data lakes and other things. So there are like a lot of things that you have to consider like who is the talent that you're going after? I can't afford, frankly, any data engineers right now. Even though we're startup as a company that's getting bigger and so forth, but that talent is scarce. And everybody is a data engineer, they take a course on Coursera, and now your data engineer. So I think you have to align yourself with the right skill set. And sometimes that doesn't come with just hiring a body. You have to work with companies, and we decided to work with Iterate. And then third is like you really have to think about your brand ecosystem, your brand sort of ethos, right, like whatever information that specifically is going to go out to my consumer has to take into account the e.l.f-isms, the brand tone, we use something called e.l.f-ing Amazing, e.l.f-ing Awesome, e.l.f. -- why the e.l.f. not? Like we use these certain lexicon that has -- that our community is used to. So how do you make sure you're training the models with the right lexicon, the right -- we use -- I talk in emojis. I can talk in emojis all day, right? Like -- so I think that's sort of how do you train even the model on those emojis and so forth. And integration with other partners like we just did -- we're actually testing IG right now, Instagram. We're building TikTok, and we're building others as well. So I think we are -- it's like one test is not enough. We got to make sure that we can serve the full community as well. So that's sort of been our implementation journey and the things that we took into consideration. Everybody is like, okay, okay, you talked a lot. Everything is on paper, show it to me. Like I need to see this thing what you're talking about. So I'm going to do a quick demo on sort of the -- the one which is our Instagram community response product that we built. So give me -- if you give me a moment. All right. I'm assuming everybody can see my screen. All right. I just want to make sure, can everybody see my screen? Give me a thumbs up. Brian, like if you can see it? Okay, perfect. Okay. All right. So I'm just going to quickly navigate what we first did was that it's not just like a product that's standing on its own. We needed to make sure that our security was part of it. So it is an SSO connected. We're working with Okta. So it is connected with our Okta integration, single sign-on, dual-factor authentication. We built this product. It's right here. So when the community managers go in, you actually get like, hey, this is the Instagram product, right? So once you go in, this is connected live to our Instagram channel. We've trained the models going back to many, many years of information. But here's all the comments, right, from the Instagram post. So this is like one post right here that was just posted. We had a collection that we collaborated with Liquid Death sold out in 45 minutes, which was insane yesterday. So our community is very happy and excited. So a lot of comments. And then so what happens is like the community is asking a question, in this case, the AI is like already sort of responding, right? And the response is not live. The community managers have the power in their hands to kind of look at that response and say, I like this response, but maybe I want to add a skull emoji because it's Liquid Death. So they would go in here and they would edit this post. And once they edit this post, the edited post shows here in this column right here and then they can post it, right? So this allows sort of almost this model, we're also training the AI to be smarter and smarter on our brand ethos. So sort of that's like one response. You can sort it like what's posted, not posted, comments posted all time, there's like basic statistics here. Here's sort of the history and all the post so you can imagine you can use this for a lot of analytics. So the more I'm capturing behind the scenes, we're also building analytics around it. And of course, then you have statistics, right? Like that gives you how many posts were responded? How many posts haven't been responded and so forth? And how do you make sure that you're also getting KPIs. It's not just about making this investment to test and learn. Eventually, that investment has to deliver some value. One value could be around. Our goal is to have no post should ever be unanswered, right? Like are we reaching that 100% goal? The other might be, I want to make sure 80% or more are answered by AI. So our community can -- members can focus on other strategic things and engage the influencers in a more high-touch way. So there could be so many goals that you have to start thinking about. But I'll pause, that was a quick demo of what we are doing internally with the product that we're building. Now I'm going to pass this off to Brian again to talk the details a little bit about how he -- their team brought this to life.

Thank you, Ekta. Okay. So here is basically the architecture that went behind it and how we kind of bought all these things to life, right? As Ekta mentioned, the brand ethos is very critical. So one of the things that we had to do was this is -- what you see here is a RAG implementation, right, retrieval augmentation generation, right? What it simply means is you take all your documents, e.l.f.-related data sources, right? And then you train -- you create embedding and put them on a vector database. But we also have to train a large language model and fine-tune a large language model in order to understand the brand ethos and also this follow emoji creation and to react very responsibly at the same time in an exciting manner to the social media respondents. So we had both the fine-tuning as well as a RAG action as the documents for the vector databases were created. And as the user searched, it would go into our platform Interplay because that's a platform all these things are built on. It's a drag-and-drop AI platform. And then what Interplay would do is it take the query, then it will go create the embeddings, perform a search on the vector database, and it will take the results and then it will send that to the LLM to create augmentation, right? And then, of course, the LLM itself was fine-tuned further on to -- on interplay to basically be very e.l.f. compatible and support that brand ethos and the actual e.l.f. experiences, right? So one of the things as you are thinking about LLMs, you want to think about both public LLMs that are available, like things like ChatGPT and all that stuff, plus foundation models, right? How do you actually think about fine-tuning foundation models and existing open source models out there for your specific needs, right? This goes back to the use cases, right? So this is an area, I think there's a lot of discussions that are out there. I'm going to take a quite quick minute, and I'm going to talk a little bit about -- now about optimizing. How do you -- both in generative AI as well as in regular computer vision AI, how do you begin to optimize things, right? So one of the things that we've done in our platform, of course, we can deploy this in the edge environment as well as deploy this in the -- across all the cloud -- agnostic to all the clouds. And then on top of the Interplay platform, we built a number of different applications for each sector. So we've done computer vision solutions, data solutions to actually generative AI solutions. And how have we kind of begin to optimize that, right? So this is another example where we have a highly scaled computer vision, license plate recognition solution that's deployed across some 12-odd countries, right? And at very heavy scale, right? We see hundreds of thousands of cars every day through the solution, right? But when we deploy this solution, initially, it was deployed with open source models. But over time, what we ended up doing was working with Intel's partnership, we begin to take the models and begin to reoptimize them on OpenVINO. And then the moment we started optimizing them in OpenVINO, we saw a mega performance boost, right? So what you see here is basically in a given server in a Core i9 type server, how do we -- we've done this on Xeon as well, you are basically feeding in video streams -- simultaneous video streams of license plate cameras coming in, right? And then we are watching CPU load. As you optimize -- on what you see on your right-hand side is optimized under OpenVINO open and the left-hand side is non-optimized. So you are seeing a 3.6x improvement as the load is getting more and more and more if the OpenVINO optimized models are performing very well, right? This is mostly for computer vision. But we've done a very similar optimization inside Intel IPEX toolkits as well as DeepSpeed to be able to run large language models on CPUs as well as on various Intel hardware. So I think the message that I'm trying to say here is that it's critical when you are thinking about, of course, first is to think about the use cases and fit it in and make sure it fits the business. But when you are thinking about the next step of productionalize it, you want to think about observing it in a very optimized manner. You want to think very seriously about the hardware you're running it on, the software toolkits where your models are optimized on and your methods of serving your LLM. So these are all the things that you want to think about, right? And I want to show you one final example in terms of talking about models and being able to switch from models. This is one of the applications -- one of our document retrieval RAG applications called ZenPilot, right? This is an [indiscernible], right? You don't need to develop in it. It's all like for end consumers, right? And here, we are enabling you to switch between many, many LLMs between bison, the private ones as well as public as well as even running a Mistral 7B, right? And here, of course, I can go into the LLM. And if I want to summarize a document, I can just click on it and bam, it will just summarize the document form very, very quickly, right? It will basically -- this is a cosmetic industry report. We are talking about beauty in retail, right? And then it basically very quickly summarize the report, and I can begin to ask questions to it, right? You can actually do the same type of capability, too, like on service pilot, you get an email and automatically it summarize your e-mail. And then it goes through all your databases, right, all your corporate databases, and it can take the data and return that data and compare that data with PDF documents and merge all that data together, right? But do it across many different LLM models, right, so you can do that. What's really fascinating that I want to share is nowadays, with generative AI, the programming language is English, right? So you can go in and configure here and then you can basically set your system prompts, right? And then once you set your system prompts right, basically, all the system prompts and all the things are done in English. You don't really see anything in code, right? So that's how things have got easier. Like Ekta said she wanted to have everyone to become a prompt engineer, right? With prompt engineering and like with a lot of the manipulations, you can get a lot of done with various LLMs. So that's was just a very quick demo I want to show it to you guys. Then I -- let me share the screen as well. Over to you, Ekta -- over to you, Sancha.

I will talk about Interplay running on Xeon core and Gaudi. So I thought I'd give you a quick background on the Intel AI platform. And I think everyone here knows that Intel has hardware, but we actually have a pretty broad portfolio of software. And what you're seeing here is a subset, actually, of all of our software. We have very open source approach, and we support many of the GenAI models that you -- that's already in the market. We also have toolkits that works with existing models in addition to optimization tools like Fast Track that we created and available on GitHub that will help you accelerate the RAG pipeline process. And you heard Brian talk earlier about the RAG process. And so we've also built optimization software on popular AI frameworks like PyTorch, which is what everyone uses to build GenAI and also DeepSpeed. And we have a huge community on Hugging Face with many optimization tools there as well and OpenVINO and oneAPI. You heard Brian talk about OpenVINO earlier. It's an -- OpenVINO's open source toolkit that we developed optimizes deep learning models and allows you to write once and then deploy on multiple hardware platforms. So making it -- when we think about scaling your GenAI platforms, OpenVINO really helps make that easier. And then real quick. In terms of our hardware platform, we have Gaudi -- the Gaudi accelerator. You might not have heard of Gaudi. It's fairly new. But that's our high performance equivalent to the A100s and H100s. So use that for large model training and inference. Intel Xeon has AI capabilities. It's our CPU product with AI capabilities built right in with AVX-512 for machine learning and AMX, which is advanced matrix multiplication for deep learning acceleration. So all of that is built right into the chip. And in Intel Core is our AI PC. It's got -- actually got 3 processors inside GPU, CPU and NPU to give you like high throughput, lower power consumption and faster response. And we also have a GPU product with performance that sits right between Gaudi and Xeon. And I want to mention here that Databricks, so third-party did independent evaluation of all the competitive processors for GenAI, and Gaudi had the best result for training and inference in terms of performance per dollar, and especially against the A100s and H100s. And there's a QR code there. So if you want to get more information on that blog that Databricks wrote, use that. And one of the challenges that you heard Ekta and Brian talked about is how compute cost can impede adoption. It's going to maybe even impact whether you -- how you use GenAI. So Gaudi 2 is your cost-effective alternative to A100s and H100s. And pretty soon in a couple of weeks, you'll hear more about the next generation Gaudi 3, which will even have -- which will have an even better performance. Okay. So real quick. Let's wrap this up with a summary and Brian and Ekta chime right in, if I missed anything or add -- you want to add anything else you want to add. But I think the most important thing that we talked about was choosing the right use case, whether that's making money, saving money, accelerating innovation and choosing a right partner who's willing to work with you. So it's not a -- this is not a transaction. It's a partnership that you want to build on for the future. And number two, your data is unique to your company, right? It's your competitive advantage. So fine-tune or use RAG to customize your large language models for your business and also to keep your data private. And three, it's important to bring your employees along the journey -- along your AI journey. And AI do the mundane work and your staff to do the creative work. And number four, we want AI to be inclusive and without bias to build relationships with every customer. And then don't forget the compute cost, look for ways to optimize and compute your compute options to help you with the compute cost. And success metrics are important. But keep an open mind throughout your AI journey, test and learn and iterate, right? Okay. So let's -- so that wraps up our presentation. If you're interested in testing some of the things that we talked about today and testing a large language model projects, get a free trial on Intel Developer Cloud. There's is a code there and we flash up our disclaimers. Okay. So let's go to Q&A.

Okay. The first one is about ethics. Where many of the methods to promote ethics in AI includes AI being used to check the AI as part of that. Do you think this is enough? Or will human intervention be a critical part of making sure that the model is ethical and unbiased. I'll go to you, Ekta.

Yes, I have a point of view. And the point of view is that I do think that until we -- people start feeling comfortable with sort of the responses that AI is giving, especially in our world, right, like where we want to be inclusive, we want to make sure we're personalizing that information for our consumers in the right way, I think human intervention will be needed. At some point, when you start seeing the results that you should be seeing, but at least there will be some moderation. So I would say that AI to test AI, I'm not ready to do that or at least take that on. I would say for the foreseeable future, I do see this as a copilot type of a scenario. Brian, I don't know if you think differently. Of course, you can disagree with me.

Absolutely, Ekta, I fully agree, right? I think at this point, where things are -- frameworks are beginning to develop, it's critical to use a lot of human in the loop and a lot of human-generated data, human-created content, right? But at some point, as the frameworks mature, that we would go into a scale scenario where there might be specific AI models that can allow to test other AI ethics, right? But sometimes like an automated testing, of course, don't always yield because that's why if you really look at like Hugging Face, a lot of the models that are coming in there, like people basically have a lot of different benchmark numbers. But when you apply them to the use case, you see a different experience. So I think it's critical to figure out the human element. Especially, at its early inception, the human element is critical, so yes.

Makes sense. So this next question is for you, Ekta. What is the most impressive beauty retail GenAI use case that you've come across recently?

Of course, the one that I showed you. No I'm just -- I would say that there's a lot on search that I'm seeing online search -- on-site search that is so powerful, especially when consumers are searching our website, how dynamically generative AI allows you to do that is something that we are looking into right now. It's different than personalization. It's different than -- this really is very much revenue driving, making that decision easier for the consumer. So from a consumer use case, I really am excited about the innovations that we can bring to our channels on search and dynamic search and other capabilities. I would say another use case that excites me is, there's always -- chatbots have been around forever. But I think where the chatbots can now actually really serve your consumer in a very conversational way? I don't think it's revolutionary, but I think the evolution certainly is very impressive, and I want to continue to learn and test and learn.

Great. So I don't know if you can reveal this, but the next question is about which model was selected -- which model did you use for the use case that you talked about?

We tested a few different ones, to be honest. But -- and we're still testing and learning. So Brian, I don't know if you want to provide any input there, but yes.

Yes. At a higher level, we used both public models as well as foundation models, right? We don't want to necessarily say the specificity of which model we used. But we used public models that are available. We did checkpoints type of fine-tuning on them, and then we also did foundation models as well, yes.

Okay. Great. Okay. I think we have a few more minutes. Is the AI workload -- I think this might be for you, Brian. Is the AI workload running on Xeon?

Yes, this AI workload is running on Xeon. So we've done a couple of scenarios. We've done Xeon as well as cloud GPUs. So this workload was running on Xeon, yes.

Okay. And I think this one -- maybe this will be our last question here. Let's make this our last question. So I think this one is for you, Ekta. How many human hours of e.l.f. Beauty staff were used to help Iterate train the models? What was that? Maybe talk about what that process was like?

Yes, I would say I don't have a big team. So I had just like 1 very enthusiastic AI product manager on our side who's like, "Oh, my God, this is a new world, I want to learn, very curious." They work directly with Iterate, and I was there from a strategic perspective. They've worked with our community managers to really document sort of the requirements or what the wish list is. They worked with testing the team. So I would say it was 1 human body spending time not just on this, but multiple products that we're working with Iterate on. So it's hard to say human hours. I would say I got 1 really enthusiastic, eager person who was willing to learn and be curious and can test these products and work internally and with Iterate and can speak both languages. So...

Perfect. Thank you so much. Thank you, Ekta and Brian, for sharing your experiences with us. Before we close, I want to mention our next webinar. Our next webinar is on April 23, and a guest speaker there will be from Landing AI. Landing AI has been really at the forefront of AI technology innovation. So look out for the invitation. Please take a survey at the end of the webinar and let us know how we're doing. And then I want to thank you all for coming today. And again, thank you, Ekta, and thank you, Brian.