Driving advertising and marketing innovation with generative AI

Good afternoon, everyone. One of the largest trends over the past few years has been an explosion in the use and growth of data volumes from customer use of digital for searching, planning, and buying products.

There's now a significant increase in the number of information sources being used to gain insights, and the utilization of first party data and data intelligence is increasing significantly with over 28 different sources now being used by the average enterprise organization.

This inherently also makes it more challenging for marketers to develop the right creative assets to deliver to the right person at the right time due to the corresponding increase in touch points that a consumer now has with a brand.

Now many of you may be familiar with a famous quote from Maya Angelou: "People will forget what you said, people forget what you did but people never forget how you make them feel."

My name's John Williams and I work in the advertising and marketing technology group at AWS and we help many customers resolve these types of challenges, to create a brand experience that ultimately drives the business growth for their organization.

I'm joined by Paul Montero Perez from one of our customers Launchmetrics and Jerry Lowe, our solutions architect lead from AWS.

So companies now face this conundrum when managing their brand perception, as a brand's perception is the sum of customers' feelings, experiences and thoughts about a brand product or service. And so you marry this statement with the consumption of brand experiences multiplying exponentially that I just mentioned due to digital proliferation and you could quickly begin to see what a profound challenge this is for companies to tackle, to ensure they remain relevant and admired in the eyes of the consumers of their products.

And at the same time as all of this, with the availability of virtually infinite compute capacity, a massive proliferation of data and the rapid advancement of machine learning technologies, customers across many industries are rapidly adopting and using machine learning technologies to transform their businesses.

Generative AI applications have captured everyone's attention and imagination. Our diverse customer Launchmetrics lives at this intersection of brand perception, digital touch points and the explosion in the number of parameters required to consider for ubiquitous brand experience for consumers and have a fascinating overview of how they've used AWS and generative AI to help brands succeed, to create a brand experience that ultimately drives the business growth for their organization.

Launchmetrics is the market's first AI powered brand performance cloud providing more than 1200 clients with the software and data they need to connect strategy with execution. Its brand performance cloud helps executives launch campaigns, amplify reach, measure return on investment and benchmark brand performance with tools for sample management, event organization, PR monitoring and brand performance, as well as voice analytics.

The Launchmetrics brand performance cloud enables brands to build a successful marketing strategy all in one place. So please welcome PP, the CTO of the company, to share their journey, lessons learned and technology they utilized.

Thank you John for this warm introduction. So guys, it's Monday at re:Invent. Welcome everybody. I really love this event. I think it's my sixth re:Invent. Every time I come, I learn so many things and I enjoy it so much.

So let's make this re:Invent the best ever. And today, we will explain a marketing problem that has been bothering companies for years or for decades. We will explain how supervised learning and generative AI is helping us to attack this problem. And we will introduce some innovations that generative AI technologies are bringing to the table to help us develop better products.

And we will deconstruct an architecture of a product that is using several layers of artificial intelligence to solve this marketing problem that we will share today. Finally, Jerry will join the stage and he will share some considerations that you have to have in mind whenever you tackle a project based on generative AI technologies.

So hopefully guys, after this presentation, you will learn some techniques or some patterns that you can use in your companies, your products or your organizations.

So at Launchmetrics, we work with thousands of brands in the fashion, lifestyle and beauty industry and we help them navigating a problem that has been bothering them forever, which is how do I measure my brand, how do I improve the value of my brand?

So the questions our customers ask to themselves are: What is my brand performance? How do I compare to my competitors? And these are the questions that we have been helping our customers to answer for years.

But what is exactly the challenge that we are trying to solve here? The challenge basically, in essence, is the gap between how a brand wants to be perceived and how it is seen by their customers or by their potential customers.

There's a gap between those two things, right? And in an ideal world, this gap could be like there would be a total overlap between the qualities and values that a brand represents and how it is seen by their customers. And the most successful brands of course are the ones who have been reducing this gap as much as possible.

However, brands often struggle to align customer perception with the intended values and qualities that they want to project. And the first problem they have is to understand how they are seen by their customers, right? And understanding this perception is absolutely essential to being able to start doing a better job of better managing your brand performance.

So perception depends on a variety of factors like customer experience, social influence, cultural context and of course in today's world, it's even more challenging, right? Because of the dynamic nature of online information, there are so many opinions of billions of people that they can change the perception of a brand in any given moment.

So historically, brands have been using focus groups and industry research and knowing their business to understand this perception and these tools have been somewhat useful, of course. But today, we will share how we are attacking this problem with a lot of technology and several layers of artificial intelligence.

So before we try to solve the problem, we need to understand the framework that marketers and brand strategists use to navigate this problem. And it's quite simple, right?

So we have here on an axis map with two different qualities, pricing on the right, quality on the top. So you can take a brand, analyze the brand and map it here in this axis map, right?

The question here is that every brand cares about different dimensions. And this is the point here, not everyone cares about pricing or quality. There could be a brand who is interested in masculinity or femininity or concepts that are really specific for that brand.

So we have here an example of some guys that arrived to the American market in the eighties and they were producing this ugly clunky, cheap, low quality cars. And they said to themselves, ok, we want to produce high quality cars and they reorganized the company, they invest a lot in research and development in their manufacturing processes and they introduced a 10 year warranty program in the nineties.

So in the period of 10 years, they really successfully moved from a low quality cheap cars to a high quality cars. It's a case of Hyundai.

So brands are constantly playing this game, changing the way they produce their products, changing how they focus their energy. But at the same time, being very careful about how they are perceived and to see if their customers understand what they are doing.

So the brand in that example may be a little bit counterintuitive to you. But in the fashion, lifestyle and video industry, the brand is the most important financial asset. So it's not so much about how many dresses they sell or how many shoes they sell every year. It's just the brand.

We are seeing valuations and acquisitions where the value of the brand is more than 60% of the total value of the company. This is a little bit crazy. It's difficult to understand, but it's strange that the brand is so important.

So having the capabilities of monitoring, understand how and why your customers see you and doing it through technology creates a huge value for our customers. And this is where at Launchmetrics, what we are doing, what we are solving at Launchmetrics.

We leverage artificial intelligence to distill the informational chaos where we live in for the fashion, lifestyle and beauty industries and we help them connect their strategy with their brand performance.

So we have been working with this problem for years now. And the first thing that we did is a lot of machine learning, a lot of supervised learning because classifying information, it's the key, the first step to start understanding it.

So here we have several examples. We can see how we are using text or images to detect if we are talking about a specific brand or it's just a random word. In this example, it's the brand or the singer Celine.

We are using machine learning also to detect products, industries to see if we are talking about beauty, jewelry, watches. And in the right, we can see how we are analyzing fashion images to see colors, patterns, brands, products.

So there's a lot of things that we have been doing through machine learning and it's piling up on top of these technologies that we will produce and we will discuss today how we are doing better to help our customers.

So we have this new player with generative AI that is changing the game, right? Supervised learning has been the king of inference for lots of years now. And it's in different products everywhere because it's cheap, it's fast and it really works.

And then we have generative AI that has a different goal which is basically to create content. It's not to classify and it's way slower and very expensive.

So if you try to solve the same classification problem with generative AI, that might sound a little contradictory also, but it's 1000 times slower so it doesn't make sense to attack the same problems that supervised learning was trying to solve.

But what we are seeing at Launchmetrics this last month working with these technologies is that generative AI is changing the way we produce software, also the way we produce artificial intelligence and supervised learning, and it gives us new possibilities to attack this marketing problem that I was sharing at the beginning of the presentation.

So today we will share a few innovations that generative AI is bringing to the table and changing the way we produce software. It's enabling us in ways that were out of reach before.

The first innovation is the acceleration of prototyping speed. So in the past to build supervised learning, you need data, you need a lot of effort to annotate and pre-process this information. And then you build the model, you check the accuracy on that model, you deploy the model and improving your algorithm. And that was a quite a slow process. It could take from two months to six months being optimistic, it can be even longer for complex cases.

So today with generative AI, if we think about generative AI technologies more as a developer tool rather than a consumer-to-consumer tool, it can help us building AI products that give AI results very, very quickly and we can move from months to days or weeks. And that's a huge change also because before deciding if we want to build specific models with supervised learning, we can just check through generative AI if these patterns are present on the content.

Right? So it's not also a change on the time to produce these projects, it's also the kind of skills that you need. In the past for this you need data scientists building those models and checking them. And right now, what you need is somebody who understands the data and has some engineering knowledge, but not complex skills.

So such a reduction in time and the skill required is a game changer when you are approaching new projects based on AI.

The second innovation that we will share today is the creation of synthetic data. So you can imagine that taking information, as we were saying, necessary for supervised learning in some cases is very, very difficult.

We are managing thousands of brands and some of those brands are barely present in the internet because maybe they are new.

so there are lots of different cases where having this data which is necessary to prepare your models is just not there. so what we are doing is using generative a i to create this synthetic data and you cannot of course create synthetic data out of nowhere. what you are doing is taking data that it's very similar to what you have and detecting the properties on this data and generating synthetic data using generative a i synthetic data data sets that later we will use for our supervised learning algorithms in large metrics.

this is giving us the capability of being able to manage hundreds of runs through a i to thousands, which is a huge change for us.

the other innovation that we'll share today is unleashing better interpretations. so machine machine learning models started in the lab like lots of years ago. and nowadays they are everywhere but sometimes the results that machine learning produces are suitable for lab environment. but a little, a little bit challenging to manage inside the product. you you really need to be a data expert sometimes to use the results that machine learning is giving to you.

so there has been always this striking balance between what supervised learning is saying in their, in their results and how do you interpret the results. so with generative a i, what we've seen is this, it is, it is an excellent res resource to help us improve the interpretation of the, of the model results themselves. and we will see that in an example now in the product that we will present today.

so in in these cases, you don't need so much data expertise, but you just need somebody who understands what is the expected outcome of the product that we are building. and the last innovation that we'll discuss today is approaching subtle concepts, subtle concepts are fuzzy ideas. words that are very difficult to define concepts like creativity, happiness, masculinity, luxury, feminine, these are words that maybe all of us have different definitions also.

so building supervised learning algorithms using machine learning to attack these concepts was very, very difficult because as we were saying, you need to pre tact all the information to build your models. but how do you prett for happiness? right.

so how do you see that the thing is love, language, models and generative a i technologies. they have an understanding of these concepts and that's a super powerful tool. it will help us to manage the dimensions that we will we were discussing at the beginning of the presentation.

so in long, what we are doing is taking all these innovations, a lot of technologies and we have been working in these two projects, topic modeling and topic scoring that they are leveraging supervised unsupervised, generative a i to unlock the brand positioning, to unlock the understanding of what the customers are saying to us as seeing about our brands.

the first thing is topic modeling, which is identifying emerging trends that are happening online. and the second is stopping the scoring. it's linking those trends to dimensions that are important for the brand that specifically the brand that is working with us.

so, but what what is the topic? um imagine everything that it said online, all these conversations are happening at any moment. we are gathering all this information and summarizing it in bits of information that are understandable by someone who doesn't know anything about data.

so here we have an example of nike and we give a narrative about the story that that it's been said and of that's the example, this the example of one topic, but we we are generating hundreds of them for each run, right? and then we link this topic, this story to a dimension that it's important for the brand. and we can score also this topic in this dimension and help us to measure how it is evolving this dimension for that brand through time.

so how does this work? we will basically, it's four steps. we play a lot with ll ms to generate embeddings and to understand the content, we look for content that it is closer to our dimensions that we care. we do the clustering to generate the topics and then we interpret the results.

so let's let's dive a little bit slower into them into these four steps. and the first step we took, i think it's six or seven millions of documents per day coming from different sources, online news, social conversation, print press everywhere that talks about our brands of our custom. and we got a all of that in a data lake, organizing this information.

and as we were planning a few minutes ago, we crunch the information and reaching it with lots of different machine learning algorithms, detecting brands, industries, products, valuation, detecting sentiment of different enrichments.

the second thing we do is because inside of every document, there are several ideas. not one document doesn't, doesn't talk only about one idea. so we split this content into what we call chunks, we dislike phrases or paragraphs and every chunk contains the first words of the next chunk. so because we don't want to split the ideas by half and then we move from millions of documents to hundreds of millions of chunks.

and then when we have all this information, we process it through lms and we generate the vector of embeddings of all this content. so the vector embeddings is a way to convert our chunks into numbers that captured the meaning and the relationship between those words. of course, the more advanced your love land will want that you use the more nuances you will catch from all these concepts, right? but the speed and cost is at the essence here.

so you can play depending on the volume of data that you have to manage and find a good relationship about the the nuances that you want to catch and the cost and time that you have to process this information. and here we are introducing a concept that jerry will explain a little bit more, which is the large language model operations that we are starting to have to manage because we will not always use the same models to process this kind of information in our projects.

in the next step, we call it sitting. what we do is create a definition of the dimension that it's closer to what the brand understands. so here we have sustainability with this an example of a dimension that a lot of rats scare nowadays. so we are having a definition and we take this definition and we process it again with a large language model and we create the vector of embeddings of this definition.

and now we have these seeds and we have all the chunks that we have been producing in the previous step. and we can look for chunks that are closer to my seat and chunks that are farther away from my seat. so i am distinguishing from all the ideas that we can find in internet, which ones are closer to what i care about, which is in this case, sustainability and which ones are farther away.

so then we are reaching our topics, we take all these chunks and we cluster them. what does it mean? cluster? it means detecting chunks that are talking about the similar ideas. and it's the first time that topics appear, we have several topics which at that moment are basically a list of chances, right? which is something a bit, a bit useless.

so we process this list of chunks that each, each topic through last language models again using the attention mechanism and the output of that is that each topic has a list of keywords ordered by weight. so attention mechanism is like focusing or detecting the most important concepts, i words and tokens that we have in the text and get rid of all the noise or not. so important concept that we have in these ideas.

so at this stage, we have a list of topics and what is the topic in our system? right now, the topic basically is related to a brand. we have that link, it has lots of different enrichments that we were discussing sentiment valuation, um products industry. we have also a list of keywords, important concepts and ideas. well, concepts and tokens that are part of this token and it's linked to what i mentioned.

and also we can link back to the original documents that we're discussing about those topics and we can link back to the images and the media that was connected to these topics. that we are also generating, right? so we have a lot of information around each of the topics that we have been creating.

so what we do is process all this information through generative a i again, and we create a narrative and a story about this topic. because if you give the data to the left of this image, to somebody from nike or any brand, they will not understand anything. so we give them an explanation of the idea that it's emerging online. and also we are scoring this idea in this particular dimension that we know this topic is about.

and then we can aggregate hundreds of topics in different dimensions and see the evolution how brand is evolving through time. so this is basically it, this is what we are doing to understand all the noise that is happening online and it really works.

so to finalize some key learnings from today's presentation through generative a i, we are producing a i faster, we are creating data that wasn't existing before we are interpreting results in a way that is understandable by somebody who doesn't know anything about data. and we are attacking and understanding concepts that were completely out of reach a few months ago. so that's the game change.

the second point is something that i really skim through my presentation. but our partnership with aws and the possibility of moving on with this project at this speed. it's because we are piling up in stacks of technology that we have been mastering through years. we started with the bobs and then machine learning operations and then we created our data lake. now we are start playing with this love language ma models operations.

so it's piling up on all these technologies that we can really move fast producing these products. and the the third idea i want to mention is that the combination of industry specific knowledge, a unique data set and working very closely with our customers. and then all these new technologies that are appealing are really creating an opportunity for, for us all, we are not creating these foundational models, but we the kind of products that we can produce are they are very specialized and you really need to be in this market to produce them.

and the the fourth conclusion or key learning is that of course, no one knows what will happen with generative a i it's i think it's too early to know, but we are seeing that what our customers expect from technology is changing and we are able to give to them content and results that are way more powerful than before. and the the human efforts behind producing the results is still very, very huge.

so i'm really optimistic about the kind of outputs and outcomes that we can have with our products and what the new technologies are bringing to the table.

so i have the conclusion in large metrics, we are reshaping how the marketing teams are in the fashion, lifestyle and beauty industries are understanding their customers. and we extend an open invitation to all businesses of all sides from learn and use these techniques. and of course, john jerry me will be available to discuss if we can help you in any manner.

so thank you and i will if the microphone is jerry. thank you. thank you. good afternoon, a big hand to power again, please. ok.

so in the next 5 to 10 minutes, i'm going to tell you at least one thing that nobody else will tell you. ok, to start off with.

so let's do this as an exercise. how many of you are actually running a gen a i product or platform in production today? if you could please stand? fantastic

You are pioneers leading the way. How many of you are actually running a POC today? Actively running a POC, please then? Ok, a few more. How many of you? Thank you. Uh how many of you are thinking about a POC? Just raise your hand. Ok, a few more. Ok.

Um fear is everything. So please stand. If you're afraid of AI, it's going to change the world, right? It's going to destroy everything. One brave man stands up. Ok. Uh and another one, right? So we hear this quite often and I think a lot of what we do with AI is actually driven by fear.

So uh let's dive right in, you can skip that slide. So uh the first thing is, you know, Gen AI is just a tool. Ok. It is really nothing new. It's not going to solve every and all problems for you. So be careful why you actually selected and, and know why you actually selected, right?

So, um it's, it's simply stated, I can't overstate that enough. It's not the tool for everything. It is a fantastic tool, it will change our lives, but it's not the tool for everything.

The second one is um because it's not easy, no machine learning or AI project is easy, it will take time, it will take effort. So be clear that you select this for impact. And what I mean by that is uh if you're going to spend your time and effort and money to actually, you know, launch a product that will generate three RFPs in a year for you, it may not be worth your while, right? So be clear in like what if it's successful, what will the impact be on your company?

Um consider your modality? So simply do you plan to actually generate text uh or a visual audio video and image? Uh I would suggest that you don't start with both at the same time, pick one experiment with that. So select carefully what you plan to do with uh Gen AI.

The second one is that at AWS we like to see, you know, privacy security as job number one. But I add to that actually transparency. And so transparency again is, you know, based on fear, people are afraid, how will you actually use my data to actually train a uh an LLM or a foundational model? Uh and how will I be able to remove that? Can I actually remove that? right.

So um for you as a company, your data is one of your most important assets. So when you use that LLM and you, you generate prompts and, and you submit data to it, is it actually updating a public LLM or foundational model? Uh or is it just your own custom model? So make sure that you know the answer to that, make sure your data doesn't leave your VPC, right.

So um clearly outline about how and why and what you're going to use, especially to the people inside your company, they need to be very comfortable with you and, and what you're going to do because they will actually have to participate very closely with you. And we'll get to knowledge experts in a in a second.

Now, the second to last one, establish a data privacy guideline. That's an obvious one. It takes time, right? So the one just below that is conduct an internal only closed loop POC. And what I mean with that is don't wait till the data privacy guideline is completed, right? Actually educate experiment internally in a very controlled environment with limited external exposure, preferably no external exposure so that you can gain that experience while the team that's actually doing or developing the data privacy guidelines can actually develop that. So that way you you, you're not waiting six months or a year and then starting ok.

Number three. So fear. So most people think, you know, Gen AI is gonna eat my job. It's gonna code faster, it's gonna generate creator faster. It's going to write, copy faster and better than I ever could, right? It's going to do better data analysis uh faster and easier and prettier than, than I ever could. That's not quite the reality.

More than any other project. you have to make sure that you include your knowledge experts about that data on day one. And so why do I say that any every step from prompt engineering to actually validating the content that that project generates are going to depend and need those knowledge experts to validate it, to give you feedback so that you can optimize it. So it's going to raise their level of effort significantly. And when I say raise their level of effort, the value proposition that they bring to the table, raise that significantly,

Um fast feedback loops. So uh raise your hands if you heard of like a disaster, a company that generated something and it went out there and it was an image that was really offensive or just disjointed. And it took like a week for them to figure out how did that actually happen? You probably heard some of these stories.

So um fast feedback loops, simple things. In, in terms of your UI design, you see some of the examples there. I especially like the one about a quick search pattern to validate the results because AI sometimes make mistakes, right? Uh I think if you use some of the the online ones, you've generated some of those interesting results yourself.

Um and then the final one is human in the loop. You are going to need humans to validate, to feed back to optimize. So when you actually create your operational design for your platform, make sure that you include a human in the loop step and optimize for that as well.

Uh communicate with stakeholders. So, uh again, you know, this is nothing new, you do it every day. Uh AI is is that sensitive topic. So the CEO fear of missing out, we must have an AI i project or, you know, we're gonna go bankrupt, the competition will win, right? Fear of the unknown, it's going to replace me. Fear of the risk, it's going to embarrass the company and it's going to expose, you know, PII data as an example,

Uh fear of change, it's going to replace me, they won't need me anymore. Um so, um and the way we overcome that is actually with knowledge with education and the best way that i know to actually get that is to provide an environment where people can play with it. right.

Um, so there is a actually a really good use case Thompson Reuters, um, open playground. I think it's called if you, if you search for it, it's a fabulous use case they did in six weeks, they created a large LLM environment for their team to play with. And so I think that lays the groundwork for safe adoption because people can understand it, they can see it, they can put their hands on it.

Um and it opens their mind for new ideas and, and how it can be applied. Um I'm a firm believer in telling people why, why do you actually need to have a Gen AI uh you know, project or center of excellence? Why is it important for your company, especially if you're going to establish it as a strategic capability in your organization?

And then the last one is, you know, a center of excellence for AI so sounds very obvious. Um but even if it's you and one other person, you know, create a playground, um start to experiment with it and then share that because from that will come your truly innovative ideas for your organization.

So fast experimentation. Uh Power actually mentioned this. So if you think about it, uh all of us in, in technology, you know, first there was kind of CI/CD so that we could actually release, you know, new versions of software faster. Then there was DevOps so that we could actually include, you know, infrastructures code and, and, and do that faster.

Uh then we got ML, you know, and then we had MLOps so that, that pipeline could go, actually go faster. So, you know, here we are so LLM Ops. So if you will think from day one about how you're going to automate your pipeline. And the reason for that is you're going to do many of these projects. Hopefully there are many, many LLMs and there will be more and you have to test with each one or many of them, right?

And in each one of them, you're going to have many, many iterations. So the sooner you actually spend some time automating those steps, the faster you're going to be the lighter weight and the lower cost those experimentations will be. So this is fundamental to your success.

Um you know, and i will say again, there's nothing new, you know. So i'll give you another example. Uh a few years ago, machine learning was something new and every CEO said we must have machine learning in our, you know, product and then we had blockchain and we must have a blockchain project and then we had the metaverse and we should all buy a piece of that. And now we have AI, right.

So new things come all the time and we should remember the lessons from the precursors to what we have today. And that's the same thing with CI/CD, with DevOps with MLOps and LLM Ops. These are not entirely new things, they're new iterations of, of new things.

So that brings us to the end. We're quite early. Um so first of all, thank you very much for coming a big hand to Paul and John for their presentation. Um i i know this is a tough spot right after lunch. So before you leave, i have one favor to ask you see the uh session uh app there to, to score this. If this was valuable to you, please open that app and, and, and give us a score. If it was valuable, please make it a high score. If not, please leave us feedback so that we can actually learn from that and improve for the next time.

So with that, thank you again and enjoy the rest of reinventing.