Workshop: Re-imagine E-commerce with Generative AI - AI PM Community Session #36

Key Takeaways

[Music] well thank you for spending your day part of your day with me today and I wanted to have this session be all about reimagining what the world of e-commerce can look like with generative AI now this session is going to be mostly focused on target which is one of the brands that I really love just to make sure I don't get sued by any Target lawyers out there big disclaimer none of this information is sponsored by Target I don't work at Target or have any association with Target before we jump in a little bit about myself I am a PM with 10 years of experience here in the Bay Area I started my career working on the online shopping World on the Microsoft store team and then I transitioned over to work on the Xbox team where I led their XBox search experience for 3 years I then decided I was done with Wasing moved over to California and worked on the PowerPoint team and that's where I started building and learning about machine learning and AI fast forward to where I am today I now work on a small startup of 500 people and working on a zero to one generative AI product since my background is in online shopping and search I wanted to to have this session and be fun and interactive so I thought it would be really cool to think about how we can reenvisioning with Target by the end of the session you're going to have way way too much information about this company what it does and the potential future ahead of it since I do have a search background I'm also going to be teaching you about search how it works so hopefully you'll leave the session with some understanding of that world as well and at the end of the session we're going to go through a little mini exercise so that you can put your learnings to the test I know you all have been participating in this Cort you want to get into the AI and machine learning space so why not go ahead and put what you've learned to practice so I want this to be engaging so please feel free to leave comments reactions I want this to be helpful for you for me I don't want it to just be me talking to a camera to you all so with that I know you're here to have me talk about generative AI it is what everyone out there in Tech in the Bay Area is talking about every single company in Tech is building generative AI features it's starting to sell these features so I thought it'd be fun to have to have a play on that since you all have been learning about generative AI I'm not going to go into you know the benefits of what is possible here here a lot of the inspiration that I took for this presentation is based on one core benefit of generative AI which is the content creation Now if anyone here can name the three content creation benefits what are the three types of data that content creation uh is possible one of the first obviously is text I'm getting a little L ones exactly images the last one music videos exactly exactly so what you're going to see is all around this idea of content creation the inspiration was taken from what can be generated so let's sit the stage a little bit let's say Target is on a hiring frenzy they've been hearing about generative AI they want to incorporate generative AI they're paying top dollar for a PM to come and join their team let's say we've actually taken on that challenge and since we have experience with generative AI we're the perfect person what do we do we pack up our bags we sell our house and we go over to Minneapolis to go work for Target for those of you who don't know Target and the brand and what it is let's learn a little bit about the evolution of Target so that you can understand how to build generative AI features target has been around since 1960 they started as a brick and mortar store they actually started as a discount store they open their very first store in Minneapolis from there they expanded to new locations they built more stores and with new stores the next thing that they had to do is create their very first Distribution Center so in 1970 they put together a distribution center now they have different locations they needed a place to store all of their items that's was one of the biggest moves that they did in the 1980s something else was new that they introduced now if you're familiar with Target you probably know that they have their own set of Brands they manufacture their own set of products in the 1980s they introduced their very first clothing line now fast forward to the 90s one of the next big moves they did was related to the era of credit cards for those of you who have a credit card this is one this is a very this is a really big deal Not only was were credit cards sort of all over the world but the fact that Target decided that they wanted to have their own credit card they wanted to be part of that movement was huge anyone else can anyone else guess what H what else happened in the 1990s this is probably one of the biggest things that groceries groceries website the website Ecommerce and website exactly they finally introduced their very first website so now they could start selling the products that they sell in store online and that that website of course is still still around it's evolved a lot now fast forward from where they've come to where we are today in 2023 of course no surprise this is where the rise of chat gbt came to light and really sort of is where we are today but if you look at how far target has come and all the different opportunities and decisions it's made it's been around now for 60 years and in order for it to continue being relevant in a company that can continue to stick around it needs to evolve with the current world so what does that mean for them let's start by looking at their ecosystem target has built this impressive brand this impressive company and it's employed hundreds and hundreds of people the first type of people that that of course run Target are the retail workers these are the people that are in the stores the cashiers the people that are working in the aisles the people that are preparing your orders when you're placing a pickup order of course you also have the Shoppers these are people like me and you who are going to Target to buy things that we need and with an online presence comes the need for a customer support Advocate someone who can help you you navigate the site who can you can ask questions to who can help you with your online order someone that can answer the phone in case you have any questions of course we have to acknowledge the corporate employees this is not one we're going to be focusing on but just to give you a sense of who all the different personas in this company are let's not forget the factory workers the people that are in these distribution centers that are manufacturing all the goods that are packaging orders that you're placing online that's another very big important person in this target ecosystem and of course there's also the people that are driving the trucks full of merchandises that are taking products from one store to another or the warehouse to stores they've built such an impressive ecosystem over the 60 years that they've become a company that today we're going to be focusing on shoppers and the reason we're focusing on shoppers because again her goal is to introduce geni into Target Target ultimately cares about Revenue they're running a business they need to make money who are there people that are buying products well those are the Shoppers so now when you think about the Shoppers think about the needs and what they do in a dayto day whenever they go to Target how many of you go to Target for the sake of just going to Target maybe all you're doing is browsing maybe you don't need any anything you just want to go to Target because you have 30 minutes you have an hour to kill you you love Target you're obsessed with it what is one of the things you first do when you go to Target for those of you who um might have never been a Target or uh because I know primarily they're in the United States they have this uh little area at the front of the store where you can kind of go and browse and it's sort of this like deal section everything there is a dollar $2 $3 that's one of the quickest ways that you can browse the store another need that Shoppers have is shopping for a specific item some people go to Target because they're out of shampoo they need more toothpaste they need coffee they need they need to buy a shirt they're going to an event they need a gift for someone these are all reasons why you need to go to Target you don't just go to Target for the sake of going maybe you only have 10 minutes you have to go in and out and get exactly what you need the third need that Shoppers have of course is returns you bought something doesn't fit you no longer need it you go back to the store and you return it and you leave now feel free to use reactions which of these activities do you all think is the most common when we think about the Shopper Persona do most people just go to Target for the sake of going to the Target do they go to Target because they have a specific item that they need they're out of something or do they go to do Returns the most common is the second one where you need a specific item you have a reason to go to Target so that's what we're going to be focusing on uh for this session so now that we know we're focusing on the Persona Shopper and specifically the need around shopping for a specific item the next question you have to think about as you look to see where you want to incorporative generative AI is whether you want to start with their online store target.com or if you want to focus on their Brook and mortar stores the physical locations now there's pros and cons to both of these suck about the pros and cons for the right hand side which are the broken morm stores now this is their bloodline This Is How They started the company and nowadays a lot of stores physical stores are dying but target has found a way to continue to succeed here I'm sure they've started testing generative AI in some stores but some of the cons for incorporating such new technology into these stores include number one it's expensive imagine having to buy all these different devices and thinking about where you're going to place them and and how you're going to enable all of those online all of those instore retail people you got to train not only the employees for the store in Minneapolis but you also have to train the employees in Washington and Texas and California there's a huge burden when it comes to introducing new technology and stores when it comes to enablement another con is the audience the people that are going in stores are probably not as Tech savy everyone here in California is Tech savy but that statement might not be true to other states in the United States the easiest way to enter this Market of generative AI is by starting on target.com an online shopping experience that's already there you can experiment you can try different features that's your easiest way to entry and that's what we're going to be focusing on for this session but in order to think about how we're going to introduce generative AI what is the one way that people shop online search right you go to target.com you have a specific item you don't go browsing their search website you start with that search box and you type in exactly what you're looking for so in order to think about what we're going to do we need to learn a little bit about search so let's talk about how search works today it starts with an input box right when you think about what you can search search has to be based on the products that Target sells products like clothing Target sells clothing for women clothing for men so there's shirts there's pants there's socks there's shoes everything has to be categorized in a specific way so for now we're just going to say all of those products are in the clothing department they also sell kitchen items pots pans cups bowls spoons forks everything in the in the world of kitchen is is in that department of course they've got plenty of other departments like the health department I just went to Target last week to go buy some medicine for the flu for a cold I'm getting over a cold right now but they also sell mouthwash and toothpaste and deodorant and all things related to health but they have they have like close to a 100 different departments they also sell food they sell bicycles they sell all sorts of things like shopping uh gift items like shopping bags so let's call all of these different departments product catalog so when someone goes to to target.com and they type in something the input box let's say the word purple there is this ranking and retrieval layer that I'm not going to go into it's very complicated and I could spend an entire hour talking about that but basically what happens next is it looks at everything in the product catalog it looks at the clothing department and it figures out what items it should return for purple it looks at the kitchen department and determines what items it should give back for purpose maybe there's a purple pan a Purple Spoon a purple water bottle and it does this for every single department and search has to be instant so the next thing that happens after it's gone through all those different cataloges and it's determined all the eligible purple items it then has to figure out how it's going to rank all these items we got back 50 different items but what should we show first should we show the purple t-shirt the purple water bottle the purple pen that's all part of the ranking and retrieval logic that machine that has to determine what to return and ultimately based on some data perhaps clicks number of purchases it ranks those products in ordered so as you can see here if you go to target.com and you you search for purple the very first product that comes back are purple pens followed by purple water bottle and so on as you can see all of these are actually from different departments this is from like supplies to workouts to gifts to clothing now once you've built the basics of search you've built your catalog you've built a ranking and retrieval you're able to build a search page with items and ranked and all that good stuff you then start to build some of these more like quot fancy features these are what I consider Baseline you go to amazon.com you go to Home Depot RI name your store every single every single online store today has this concept of autocomplete when you type in text into that input box you expect them to give you a list of different options based on what you're looking for now another one of those features that have become Baseline these days are filters when you press enter and you're on the search page you expect these companies to help you refine what you're looking for so that you can quickly find the item you need to make that purchase for Target their filters include different brands like these at the top then you also have the other type of filters like sorting by Price sorting by Department sorting by size and so on this is how the world of search works today you got a catalog and you've got all these sort of like Baseline fancy features but the world of search as we know it today I'm calling it old search has its own challenges for one it's highly dependent on the product catalog what do I mean by that when you go to target.com and you search for let's say hat or Beyonce card if Target doesn't have any of these items it's not going to give you back anything in the search search results page that's the opportunity especially with geni and the ability to create images if they don't have a card for a birthday card that's Beyonce style couldn't you just generate an image of Beyonce I mean of course there's like some other legal rights to generating an image about an artist but that aside the point here is sometimes you're searching for something that doesn't exist that's an opportunity right there the other reason why search has its challenges is that it's dependent on product attributes in order for you to get results for purple a human has to specify some attributes for that product these attributes include the title the images the price that you're going to sell it and the different colors if Stanley had a purple cup but forgot to add a purple image or a purple color as an attribute that item will not appear on the search site of course you have some other attributes that can only be determined based on usage data like ratings and whether or not it's in stock so search has to maintain an understanding of these attributes to know how it's going to be ranked because search was built in this way where it's dependent on on items being in stock and attributes the way that we all learned to search became this turned into this concept of keywords so today when you go to amazon.com when you go to any store you're really trying to figure out what is the keyword that I could put in that's going to give me the result that I'm looking for I need a toaster I'm just going to go ahead and put in toaster I know it's going to give me back items because I know there's going to be toasters that they sell most of the searches that we do are somewhere between one to two to three words here's another example you need a black dress what do you put in that inut you say black dress very simple you know that they sell clothing you're trying to see if Target sells any black dresses and sure enough they do sometimes you even add some attributes to what you're looking for you need some shoes I'm not just going to put in shoes I'm going to say boots and I want these boots to be pink and I want them to be sparkly and because I know that Target sells shoes those are the keywords that I'm going to do that I'm going to add into the input box to try and find exactly what I'm looking for the world of search has evolved with generative Ai and very few companies are taking advantage of this generative AI has taught us to search in a different way now if you think about trbt we're all the way that we're interacting with jbt is through natural language we're writing complete sentences we doing followup questions if you were to do a search like this one here I need supplies for my camping trip what Target is going to return is a set of items that have the attributes camping or trip associated with them perhaps it's in the title perhaps it's in the description but this isn't really what I need I want snacks I want a tent I want a sleeping bag and as you can see here Target is not has not taken advantage of this it's still working based on the way that old search works and it hasn't evolved to the latest patterns of how search of how we're searching today through these new generative AI capabilities now going back to the initial topic that we're all here to learn about we said we were focusing on shoppers and our goal is to think about how we can reimagine the world of generative AI to start let's break this down a little bit further we know we want to focus on shopping for a specific item and we said we were going to focus on target.com because that was the easiest Bearer to entry so now let's talk a little bit more about some scenarios that arise for this Persona the first could be that you're shopping for a specific item that has very price specific criteria maybe you need a laundry basket and you need it to have wheels because you're going to store it in a specific corner of your house so you need to be able to roll it out every once in a while but you also need it to be in a rectangle shape and not circular because it's going to be in the corner of your space and you need it to have handles because you're going to go up and down flights of stairs with it so as a shopper you know you need a laundry basket but you have specific criteria in mind the second scenario is really all about shopping for multiple items I just moved into a new house and I don't have anything for my kitchen I need a toaster I need pots and pans I don't really care about the brand I don't have a lot of criteria I just need a basic toaster the third one is around planned activities when we go shopping at Target we have either criteria need multiple items or or have an activity that we need to shop for maybe it's for someone's birthday maybe you're going on a hike and you need to pack some snacks a water bottle some socks whatever that activity is these are all reasons why you need to go to target.com to shop for an item let's start with the very first one around specific criteria now let's pretend that we go to target.com and there's some new generative AI capabilities what I'm going to show you next are a series of different images that have been generative generative generated with generative Ai and they're really just here to kind of talk through different concepts that potentially can be introduced for inspiration so let's say we have a very specific need we're looking for a kitchen appliance that can bake air fry and fit in a small space we live in SF we have a small kitchen we need a tool that can do all these things if you go to target.com today we you're probably going to get back is 20 different appliances maybe one of those is an air fryer and that's all it can do maybe the other one is a toaster oven and it can only bake it doesn't give you something that fits all these criteria what you have to do as a user is go through every single one of those products read the description look at the images and decide if it meets the criteria that you're looking for that's the old search we don't want that in this new world generative AI has made it possible for for for for Target to give you an answer so maybe in this new world Target knows that there's only two items or maybe one item that fits the specific criteria you want it should just put that front and center for you it should present it as an answer and tell you exact ly why this is the product you should buy and might even pull in internet data and give you recipes that are possible that you can make with that product that you're looking for so this is net new today Target can't give you recipes because there's no online recipes that it sells in its catalog this is all from internet data right it wouldn't it be nice if it just gave you you know it acknowledged the fact hey you said you're looking for something that can air fry something that can bake and can fit in a small space this is the one thing you're looking for don't waste your time going through all the other products because you're not going to find it and again both of these were generated generated with generative AI but the whole idea here is just present what the customer is looking for right there on the search base don't make them go through all these other products especially if you have ratings data you know that this product has been well Reed by other customers so you feel good about this recommendation let's talk about a different scenario now in this case you have multiple items that you're looking for let's say you just moved to a new place and you need to furnish your living room you don't have any furniture and you have a specific budget in mind perhaps it's $500 so why not specify that right there in the search box so that I can get Target to tell me what exactly I should buy but let's take that a step further we have a picture of our living room why don't we provide that for them in this picture you see that I have some windows and then I have one wall when you think about the basic items that you need for a living room you know you need a couch perhaps a coffee table some side tables and some artwork you can see that I have nothing in this living room and all of those basic items that I just talked about are things that Target sells so what if Target could offer some inspiration of what I could put in this living room here and it could tell me the specific products that it has in stock that I could go shop for that are within my budget and what if it took that a step further and gave me different mood boards here's the aesthetic of if you want a modern aesthetic here's the aesthetic if you want a beachy Vibe maybe if you want everything that's Peak you could decorate your living room in pink this is what generative AI can do when you Incorporated it with uh some of the latest capabilities let's do one more in the planned activity space when you have a specific hiking trip you you're going to take you want to get a list of supplies right but we could take that a step further today if you go to amazon.com if you go to Google you type in a some text in the input box you've probably seen this concept of followup questions where there's another input box that appears that will ask you a follow-up question I said I'm going on a hike the next question could be where's your hike where are you going to that's one way to do it but what if Target since it has store all across the us could tell customers the different hikes that are near those stores the location by using internet data and what if it went a step further and told you the top products for different hikes if you're hiking in the mountains of uh somewhere cold and snowy you probably need a different set of shoes as if you're going on a hike in Florida where it's hot the thing that you need for a hiking trip sure some of them are Universal but what if you but what if that collection again was curated based on location and was right there front and center for you I'm going to go hike in California or I'm going to go hike in the east coast here are all the top products in the east coast here's what the weather's going to look like this weekend and so on now I just went through a series of different activities just to give you an idea of the art of the possible happy to answer any questions I uh have this sort of exercise that I thought would be fun to have to work with you all on and and see if uh you could put what you've learned to test through your through your course that you've taken so uh feel free to unmute yourself feel free to send in chat uh I see there's a question here we got also a few questions um that are being posted on LinkedIn um so maybe I'll kind of go through them um as well um before you answer the ones that raise hand just because um there are a few questions that were asked earlier would it be okay if I read a couple of them that were typed earlier yes go for it okay so one of the questions that were asked was with um with sashita she said um can you please explain how generative AI found the right product from a feature building perspective maybe sashita you just want to kind of explain your question in more detail feel fre to just jump right in yeah I think when I wrote that we were at the air fryer so essentially there was a lot going on uh so the data that we have to train is the targets data we have obviously their user you know actions interactions and obviously on the other side you have the product their attributes and the history of the product details uh how what exactly like which generative AI capability did we use to come up with this of course there's a little bit of I believe you know there is a search involved in this looking for the right because there's no image there's just a description what was the underlying layer that was working yeah I can talk a little bit about that so target has let's say 20 different appliances all of those appliances have description they describe what the product does when and I was reimagining what could be possible here I was thinking about leveraging rag the retrieval augmentation generation piece of generative Ai and really through what that could do is we could analyze all of those 20 products we could look at the descriptions we could leverage large language models and say okay uh read the descriptions tell me what uh like this customer is looking for air frying baking Etc narrow down the products based on that and through a process of elimination you can determine okay there's only one or two products So based on that information that you did offline you can determine oh this one product that's in my scenario there was only one product that was applicable so that's one part the second piece is the acknowledgement part so if you've gone to Google search and you SE like the instant answer I think this concept of acknowledgement is also important that is something that the Shopper needs to know that what they're buying is the right decision so in my in my view of this concept I want target.com to say hey this is the one product that can do XYZ go buy it like don't think about it don't go look at other products this is it you could spend your time looking at other products but I'm telling you I don't have anything else that meets your criteria does that answer your question yeah it does thank you yes to oh I was just going to say I uh flew through those exercises so if you all find Value in me explaining how potentially these could be built I'm happy to go through and talk about that as well awesome I was just GNA add uh one more thing and say that uh for those of you that you're wondering what Rags are uh we do have a session actually next Saturday same time um somebody's going to go over Rags with like some exercises so that's actually a pretty good addition to what Anna just talked about um and in case you want to kind of dig more into the AI product management cohort stuff feel free to just like um apply to our upcoming cohort I also submitted the link on um on LinkedIn as well okay so passing it to M now go ahead so uh thank you Anna for the great presentation I had to things in mind uh first of all I so you already touched one of them the groundedness of the answers of the uh search result so I was curious so I I know there might be some low hanging fruits in this space of like maybe we can easily uh you know show some result that we couldn't do that before and you know you know sell more but when we go deeper and find information even through multimodal searches like not just the this description of the image and products let's say we can we can interpret the image of the products and then show some results based on that so I'm curious um what you know what do you think about these two challenges of gathering information from let's say images and videos to to show the results of the you know search queries based on those and also how to kind of show and interpret the results to the customer so they can make sure the AI can you know you know does show there them the right product and it's true so because one piece of information came from the review of a customer and other piece came from the image it is it is kind of challenging but it's interesting for me to see how you know search can can you know uh mitigate these challenges uh if I'm understanding your question let me read it back uh one of the benefits that generative AI offers is the ability to introduce images what are some of the challenges that Target could run into especially given that today some of their products already have images as part of the attributes I think another challenge they face well first off let me when I think about the space I think today products have images and if you don't have an image an image then it can't show you an image today that process is very manual that's one of the way that uh generative AI can really help as humans type in every single metadata information some of the challenges especially if you don't have imagery or thinking about incorporating imagery I think for one it's the um concept of um whether or not the image is real like the the the the potential impact of you showing something that might be harmful like maybe um maybe you're shopping for a Christmas tree and there are limited Christmas tree pictures for the product that you're selling but because we're using J of AI we could show you what a Christmas tree could look like if you were to place it in a snowy location or inside your home or somewhere else but at the same time you might be overselling you might be the Shopper wants the tree that you're you're you're advertising if you start showing them pictures of products that don't look exactly what you purchased or thought you were purchasing you then run into this whole like you lose trust with the The Shopper right they might no longer go to Target I think that's that's one obvious one another one is maybe around like uh licensing of images I mean it's not really regulated today today you can generate all sorts of images and you don't have to worry about whether or not you're stealing someone's idea or someone's art or concept I think that's going to be another big one one opportunity that I think target could potentially start with is um um in the cards like the birthday cards which is a big department for them it's usually three rows in a store if you are looking for a card for a birthday for something specific and you want to you know make your own card I think that's the easiest way for entry with some disclaimers on top of it but um I'd have to think a little bit more about that one that's a great question does that does that help answer how I'm thinking about it so yes so actually but you brought up another challenge as well so my main question my main CH you know question is for the rag solution that you just said I'm curious how multimodal rag can you know mitigate the challenge of uh ground groundedness for the results that we show to the customer and how to you know relate those results to the actual uh you know description or image or video that we have that our model could you know draw those and show the result to the customer so I'm not sure if if any startup or big companies started to to you know experiment these challeng this challenge yeah so here's one idea on how you can use rag for this this scenario here that I have on the screen so Target sales couches it could analyze all the different images it has around couches and it can use internet images and data that's out there around different uh living rooms for these couches and it could try to decide based on that okay like these couches are L-shaped couches these couches are all against the wall and and that's one that's uh what really inspired this concept here of like okay we've got a 100 couches this image has Windows I cannot do an L couch here so I don't know if that answers your question but that's one potential approach for how it could use image multiprocessing to identify back to its product for groundedness all right looks like MOS is Happy okay one more question and then we'll continue because there's an exercise that and has planned for us fishy you were talking about uh the data that uh the model is trained on and you were talking about the descriptions that each product has and the rest of the structured data that's already available uh but that is still you know a limited set of data uh and when Shoppers are asked to search using descriptive terms uh many of those terms may not be there in the structured data or even in the description so for example I me say you know I have seen a roundish appliance that actually uh cuts my vegetables into SP spirals like you know like noodles and then the model may have difficulty coming up with the appropriate items even though they are there just because my terms are so uh you know nebulous uh so in your experience do companies also utilize external data so for example you know if the large language model was trained on forums Etc uh then there's much more likely then the terms that I am s you know searching for there'll be a greater probability of doing a match uh and I read that you know companies are limited to using the amount of data that they can train the models upon what is your experience what are you seeing regarding um your question is specifically around when you're training a search model do you leverage other data on top of the product attributes to train your products is that is that your primary question yes uh the short answer is yes uh so On Target there's I can speak to it from my experience at my current job where today the way that uh actually let's talk about it from this example here so this is one one scenario on how the product attribute data is not enough and you need other data to annotate so in order for a Target to be successful at this you probably need human annotators or other ways to do this sure you could leverage a large sandwich models and you can put some Logic on top of these products and add additional attributes uh for example for a tent you can start adding user queries a tent can now have I need a tent and that's now an attribute so that the next time someone says I need a tent that product can now be surfaced that's one immediate thing that you can do but that doesn't necessarily doesn't you need more than that to make sure that the results that you're going to get back are the right ones generative AI can label things incorrectly so now you need to introduce human annotators who are going to look at some of the top queries and look at the products and validate that for this tent did it add the right user query if it added a query that says I need a shopping bag I like a this product can also show up for snacks and that's a potential risk that generative large models introduce if you were to take that approach so you will need to have other ways to add attributes through either large Sandage models and if you do that then you need you need human annotators to validate okay great thank you all right uh let me go back to the exercise that I wanted to have with you all all right so we have roughly 15 minutes we could probably spend 5 10 minutes and then I'm happy to answer more questions but given that you all are looking to go into the AIML space and generative AI is a hot topic one of the hardest things that we have to do as PMS is come up with ideas so I thought it'd be cool to have you all sort of pick one of these personas that we didn't talk about that are still related to the online shopping world and think about some of their needs and some ideas that could help solve their needs so feel free to Noodle on this if you have immediate thoughts feel free to unmute yourself we can have a discussion around that around this specific Persona and also happy to answer some questions as you all kind of think through this as well do we want to give a few minutes to people are are you thinking like you know how many minutes would you like to give them uh we can give you all three 3 to 5 minutes okay we don't have any music to play but for the next time I'll make sure we have some sort of music to play during the way time e all right just in case you guys are wondering we got three more minutes so continue writing and I know I think people are also typing their ideas on the chat does that work for you as well let me take a look see they they're typing it seems pretty good I can also read them for you if you want afterwards but I can see them we seen this oh I see them I see them let me okay great bring back the slides no I don't think there's a way for me to do both okay well we can give two more read them for me yeah what that yeah sure I can read them for you yeah no problem okay thanks so two more minutes and then we'll get started I'm reading I I was able to pull the chat up so I'm reading some of the comments right now A yeah that's a great idea the using the models to find wrong images especially because humans can make mistake when they're entering these items into the catalog retail workers also really um fun to think about also I'm just reading one of the comments here yeah somebody on LinkedIn also wrote down the retail workers MH all right one more minutes and then we'll continue oh this is a great idea the for a product for our specific product because you know we have a lot of different videos that people create on product reviews offering those uh available would be cool okay great I think we're pretty much right on time so feel free to um just so that people the audience on LinkedIn is also aware it'd be great if you are discussing any of them you can kind of um read them just so that people that don't see the text message they can also participate on the LinkedIn side any does anyone want to unmute and volunteer and talk through sort of like the Persona they picked and some ideas they had I can give it a try uh so I picked up the Persona of a customer support like uh most of the call usually they get is like when they have problem with the website or their account and things like that so if you build a board that can supply them with the knowledge based articles that can quickly annotate the answer and then Supply to them then that reduces the the uh time that the response time uh that is one idea I got and the second one is like sometimes uh there are like I heard like the chat Bots are built based on like uh analyzing the mood of a person based on the tone right another one like when customer is speaking like based on the tone like if they can analyze and give that perspective to the support agent then he or she can manage how to uh structure the answer to pacify the customer if they are angry or to pleas them something like that and uh that helps to convey the solution like if uh if there is any additional things like uh if the customer came up with an question then additional uh solutions that you can provide okay I do not find an uh item uh for my camping trip uh is there anything that you can suggest then uh we can give a chat Bard that give like a bunch of suggestions so yeah I think uh customer support is Big because you have years worth of online data Target already has a bot you know they've I'm sure they've evolved it since then but your ideas really resonate because you can process all of the years of data they have to identify what are some of the top problems and find ways to find to make it easier for agents to quickly respond to things or even in real time you know when you're going to the bot and asking for questions you can quickly summarize large amounts of information for the agent so I love that thanks for sharing anyone else yeah sure I can share Anna and the first persective I was coming from is just the way people are purchasing now is different right and how do you really how can Target uh start feeding into and getting value from the customers pre-journey before they come to Target so kind of like Amazon why Amazon is doing really good is people are doing the search you know they have so much information about customers which Target doesn't really have so by offering a way for people to provide information like the videos or pictures or wherever else they are going to start getting those ideas it does two things one it's um gives new orders to Target that maybe it might not have had because someone else could have gone to that Instagram link or wherever and purchase items from there but the second thing is now it also starts getting more information from its customers so it starts capturing more additional information with the customers and thirdly it's really meeting the customers where they're doing their shopping because no one goes and says like at least I don't go and say I want a pink shirt I'm really looking at something before and that's how I think I want to look so I think it really people are shopping differently so it's really start supporting that new mechanism yeah definitely I think in the next year or so we might all be searching in a different way but those are also great ideas for sure um we can have one more person sh I'm happy to also answer any other questions you might have about this topic of generative AI Arch or anything that might come to mind I have one question so what are some of the challenges of building these kind of systems so rag can can you touch upon a little bit rag um so rag is this uh stands for retrieval augmentation generation and really what that means is you have a lot of data so there's the that's like the r piece what can you do with data the augmentation is you're doing something with that data and based on what you're doing you're generating something so uh the product that I'm working on right now is heavily using rag uh I can't speak into the specifics of it but some of the challenges that you run into is uh the qual the quality of the data that you're using so if you're processing large amounts of data the data is not always clean so there's always some sort of cleanup that you have to do manipulated maybe you filter out specific data if we talk about it from a Target perspective Let's uh let's talk about the problem of of attributes and how attributes perhaps are incorrect not everyone types in the titles in a specific way sometimes you forget to add attributes so you have to figure out how you're going to filter and clean up that data and there's many different ways you can do that you can do that through building some rule-based machine learning rules filter out everything that's that has jargon and here are some examples like jargon it starts with XYZ you could talk about leveraging uh the large language model GPT and giving it a fuse shot learning I don't know if you all are familiar with these Concepts but there's a this concept called fuse shot learning where where you're teaching the large language model okay I need you to clean up data this is the type of data that I need you to clean up here's an example here's another example and here's another example based on the examples that I've given you go and do the clean up for me that's one challenge another challenge I see with the manipulation is uh with the large sandwich model gbt not giving you exactly what you're looking for especially when you're generating information and you're using you know just if if you were to just use the open AI uh GPT 3 or four model it has a mind of its own sure it might work the first and second and third time but at some point it's going to do what it wants and it might stop following the rules that you asked it to do this is something that's ever evolving so you really need to add more things on top of that this could be like it's this could really be its own topic like it's it's all brand new uh large language models have been around for many many years like eight 10 plus years but the world has only really started thinking about it in the in the last year with jpt and and everything that's come up since then I I hope that answered your question I didn't want to get to in the to in the weeds with it I see there's another question would you consider image recognition or finding quality issues from images a gen features um I think there's different ways you can think about gen features it's whether you're optimizing for internal internal needs or whether you're optimizing for uh external customers and external I mean the Shoppers so you have to think about it uh based on what your goal is is your goal to make money is your goal to clean data internally so that you can Surface different products and the quality of your products are good so I think image recognition can be applied to both depending on what your ultimate end goal is um I can also try another person go ahead so I wanted to focus on the retail workers I thought maybe their main jobs are Inventory management uh customer support payment process and maybe some segment of them are the new retail workers so between all of these work you know tasks I thought maybe training new retail workers through generative AI uh by you know teaching them some some of their tasks by just some conversational AI might be a good way to train them uh but another solutions could be for example uh maybe for the customer support card you you already touched on for example putting some H and mobiles and Hardware to like for the users to ask them but there is more cost for the inventory management I'm not sure if gener if we can use generative AI or not but we can use some prediction models to help retail workers to go and uh you know manage the inventory easier than just manage it manually and for the check out process also uh there could be some like uh Innovations in maybe a s checkout with uh you know cameras to to look at the products and you know uh speed it uh you know the the basket of the customer and the all the products that be taken through the visuals but these these might not be Genera but yet these are things came to yeah those are some great ideas I think the immediate one that uh definitely like resonates and sends out to me is the one that you talked about regarding inventory you think about Target they have hundreds of stores across different locations imagine a world where they have cameras on every single aisle and that's a lot of data for them to figure out like from an invent inventory perspective what are the products that you should be sending to stores in the east coast versus stores on the west coast and so on especially like for different aisles different departments that's one one immediate way that you're the goal here would be optimizing for uh the inventory and and you could argue that if you always have the products that people need in different locations you're ultimately making more money well thank you everyone for your time I hope this was helpful I hope you learned a little bit more about Target hope you were able to take some of these ideas and think about how you can reimagine the world for the area that you work on and as we think about the future for Genera generative Ai and the Art of the possible uh an thanks for the wonderful presentation and it is awesome and we like especially I like the way you presented it very clear and crisp I hope you feel better soon tha