Tim Warren: We've created our own and we think it's better. I call it the Harry Potter style of parenting of parenting. A chatbot. If you have a child and lock it under the stairs, it's not going to get very good at stuff apart from magic, maybe. This stuff is hard, man. It's so hard. I don't understand it.
[Voiceover]: Welcome to ADD to cart, Australia's leading e commerce podcast that Express delivers all you need to know in the fast moving world of online retail. Every week, Nathan Bush catches up with Australia Australian e commerce leaders to get all the insights, tips and lessons to keep you at the top of the e commerce game and of course keep your customers adding to Cart. Here's your host, Pushy.
Nathan Bush: You think your cousin is cool because he can turn any song into Weird Al Yankovic Lyrics well, today's guest has been focusing on AI since 2016 and is now reaping the rewards of the technology becoming widely available and retailers demanding it. Tim Warren is the co founder and CEO of Ambit, a company that started building a conversational artificial intelligence way back in 2016. Ahead of the game or as Tim describes it, just a little bit braver than others. They went all in on Ambit GPT, a retail focused chatbot that provides deeply personalized automated conversations that that delight customers. Kiwi born and bred, Ambert has clients such as the Warehouse Group, Glassens and Dermalogica and raised a million New Zealand dollars last year to accelerate expansion here in Australia. Tim admits that understanding AI is really hard, but he does a great job of demystifying it for some of us. He unpacks the different types of data behind the platforms, shares how retailers can get the best out of their chatbot and and how to stop the black hats from winning in the wild west of knowledge engineering. He also goes into fan fiction territory and explains how to train our very own AI Harry Potter Wizard. Just go with us here. Tim is offering Add to Carters a free ROI analysis and savings estimate for your contact center or support team. Listen out for that code at the
Nathan Bush: end of the episode.
Nathan Bush: So thanks to our partners Shopify plus and Paclio, here's our conversation with Tim Warren, Co Founder and CEO of Ambit. Tim, welcome to ADD to cart.
Tim Warren: Thank you Nathan, Great to meet.
Nathan Bush: Great to meet you.
Nathan Bush: I secretly lurked on you and saw your presentation at Imedia on the Gold coast which I really enjoyed. So I'm really looking forward to this conversation today.
Tim Warren: Yeah, fabulous. It's good to know that I've got some stalkers out there and that's only part of it.
Nathan Bush: Wait till I tell you the whole story, Tim. Let's kick straight into it. Ambit GPT. Are we talking GPT as in ChatGPT here?
Tim Warren: Yeah, GPT is the same, but ChatGPT is a product that's come out from a company called OpenAI. We've created our own and we think
Nathan Bush: it's better just taking down that little GPT. What does GPT actually stand for?
Tim Warren: Yes, GPT is a type of technology and it stands for generative, which means it can write things pre trained, which means a lot of the work's done for you already. And Transformer, which is the actual technology that's way down there in the technology stack. It's a technology that was theorized in 2017 by some Google engineers and they thought, well, it's so powerful that we should release this and see if anyone can get it working. And it's taken until last year before people got it genuinely working.
Nathan Bush: Well, so it came into a lot of our consciousness, what, October, November last year?
Tim Warren: Yeah, 30th of November last year. The world changed, essentially. Yeah.
Nathan Bush: And so was it on your radar before the general public kind of got aware of it?
Tim Warren: Yeah, so we'd been working with it for about 18 months at that stage, this generative type technology. And in fact, speaking of iMedia, if your listeners know about iMedia, it's a great event. We presented on it at the previous year in iMedia. And I don't know if you were following me there, but I created a few jaw drop moments where I was able to live create some pretty impressive copy for advertising for social users, particularly in that retail sense. And so we had 18 months time working with it and working with all different engines. And I think what Happened is when ChatGPT was released, it went from being this industry secret to being all of a sudden it was retail. And as you know, the rest is history. Fastest adopted technology of all time.
Nathan Bush: Yeah. So what's the difference between GPT and chatgpt?
Tim Warren: Yeah, so GPT is a generic type of artificial intelligence and anyone can build it. The papers are out there, you know, the research papers, if you can understand them. I barely can. The papers are out there for you to read. And a few companies actually started building what we call Foundation Model GPT engines. One of those is a company called OpenAI, with spending hundreds of millions of dollars to do it. And one of their products is called ChatGPT, which is a regular chatbot type interface to access this incredibly powerful technology. That's underneath.
Nathan Bush: Gotcha. So Ambit GPT is just like ChatGPT, using the same underlying infrastructure, but a totally different product.
Tim Warren: Yeah. So a different product. We can tap into a range of things, so we can use OpenAI for certain things, but we can use various technologies, including some of our own that we've built. What we do, we don't have that. We don't profess to be an expert in everything like ChatGPT does. We just do a couple of things really well. So in particular, we work really strongly with retailers. We also work in a couple of other sectors, but helping them have really great customer experiences so you can talk about their company and it feels just like having a discussion with a discussion with a person.
Nathan Bush: Wow. And were you in the customer service game before GPT was added to Ambitious.
Tim Warren: That's right, yeah. So we kicked the company off in 2016. Myself and call it the Brains Trust, were sitting around saying, hey, there's some amazing new technologies coming down the pipe. What can we do with AI? We figured that it was going to have a big bearing on customer service in particular. So between 2017, when we founded the company, and 2021, we focused on customer service, particularly in retail, but we also work with utilities and financial services in a few areas as well. So we were big in that area already. And when this, when we discovered that this GPT technology we built, it was going to be good enough for us to use in production, we started testing that with customers and I started demoing it, et cetera, in public and started getting a great response. So we, at that stage, we started building it into our existing product of customer service. But we really quickly worked out it can actually do way more than that.
Nathan Bush: That's a pretty big bet back in 2016 to say AI is coming. We don't know exactly what form or function it will take, but we know it's coming. It's going to be big. What inspired you to go all in on AI?
Tim Warren: Stupidity. Yeah. Rash bravery and arrogance about my actual knowledge. So no. But I could probably wrap that in a better marketing message. So let's say no. I like what we really good. So look, what we did is sitting around with a group of people. We had time, we had a bit of understanding, we had some sales experience, lots of technology experience, and I guess we thought, well, we might not be smarter than other people, but maybe we're braver for whatever reason, maybe it's stupidity, like I said. And so we said, if we get ahead of this technology, so we're Using it. We don't know exactly how it's going to work a few years in the future, but at least we'll have some experience doing that. And there were kind of two moments when our initial bit was validated and then when it was validated in a much bigger way. So with, at first we got to a stage, we had only seven people in the company and all of a sudden we started getting incoming calls. Now we never intentionally set out to work with retail. Retail came to us. And this is one of the things I love about retail. Because of the fast cycles especially, especially in fashion, which is where we started, people are prepared to try things and they'll try it and if it works, they keep it. If it doesn't, they drop it and they move on to the next thing just like a discarded piece of fast fashion. Well, we started in 2017 and by the end of that year we had, we were working with bank and then we signed Helenstein's and Glassons. You know, one of the, one of the large E commerce players. They were bricks and mortar. They launched E commerce a little bit under the hood and all of a sudden that customer service started hitting them and they wanted to automate it. So they were really forward thinking there. And we said, we quickly pivoted in our heads is what you do with a startup. We said, well hang on, we like these people because we are able to do deals really quickly. They're forward thinking, they want to look at the front of the stack. They don't want to be followers. Right. We want to work with people who are right at the front of things. And retail is actually pretty good at this. I go to retail conferences and I go to, let's just say conferences with other sectors. I won't name them. And retail's generally three or four years ahead. I would say they don't necessarily even know it, but they're doing really good things, really great experimentation. We just want to continue being part of that. So now retail makes up, I think more than 50% of what we do.
Nathan Bush: Ever scrolled through an E commerce packaging website for fun? Nah, me neither. Until today. Paclio is putting the jaw into the packaging game. So let's play a game. I'll tell you the name of the Paclio product and you have to try and guess what kind of product they are. Fairy floss. Compostable mailer. Queen bee Honeycomb padded mailer.
Nathan Bush: Here we go.
Nathan Bush: Gummy shark water activated tape. Now if my jaded self thinks that this packaging is fun, imagine what your customers Will think Paclio is also eco friendly. Australian owned and operated with same day dispatch and 14 day returns. Now that's pure joy for everyone. Check out the Paclio range of ecommerce packaging options@paclio.com that's Paclio P A C K L E O paclio.com Tell us
Nathan Bush: about some of the retailers that you've got on board today.
Tim Warren: Yeah, so a brand in New Zealand, or rather a group of brands is the Warehouse Group and they consist of Noel Leeming's very large electronics retailer, the Market, which is a marketplace which sells anything and everything from all over the world, including heaps of Australian retailers are on there. And then we've got Torpedo 7 which is an outdoors brand. So we work with all of them and they're all quite different with different needs. And that's the benefit of this kind of technology is we can adapt it to what people need. Dermalogica, Konica, Minolta are another couple of brands and you probably know this, we partner with Imedia themselves which is all across retail.
Nathan Bush: Brilliant. So I'd love to dive into a couple of case studies rather than us go through the feature list of Ambit GPT is can you talk us through a couple of ways how that one or two of those retailers are using Ambit GPT to amplify and make their customer service more efficient?
Tim Warren: Yeah, so I love thinking about the customer experience end to end. So once they've come into a retail experience, online retail experience and let's use Torpedo 7 as a great example and you'll know some examples in the Australian setting that Torpedo 7 is the dominant player in New Zealand. So someone comes to the site and now they can self drive, they can click on on links and if you think of it, a link or an area on a page is what the web developer or the designer wanted you to do. The difference with chat is it's what you want to do. So you just type into the chat bot, you click on there, the friendly little face pops up, says hey, how can we help you? Just ask it whatever you want. And this is a demo that you saw. So people come and they click on that little smiley avatar down there and it pops up, says how can we help today? And they just put. And this is the demo that you saw. I want a kayak or whatever it might be, I want to return my water shoes or I need a vest or something like that. So they put in I want a kayak. Now we can pop that up and we can find, we can look into their database we can find the. The number of stock items that they've got for across the range that matches that kayak. And it's only going to bring up kayaks. It's not going to bring up kayak accessories because that's not what they asked for. Pop that up in a nice familiar kind of carousel so they can scroll through it. But we're also going to only show the things that are relevant to their location. So, you know, within driving distance, if they want that. Or you could have a button there saying, do you want to see what else could be shipped to you? People normally go and buy kayaks and they pick them up, but if they want it delivered to them, they can look at the options for that. Now, say they choose a kayak and they check out, they just click on it, you know, touch to buy now, then they go through a payment process. Well, we work with payments providers as well, so one of those is lay by. So lay by, you know, you turn a purchase into six equal purchases or payments over a number of weeks. And we can automate that process as well, both from the payment itself, but also the service on the back of that. And then it needs to be shipped and that's where we work with the shipping provider. So it could be something, I think a lot of your listeners would have heard of, shipit or starshipit. So we would integrate into one of them, help them get a live. This is in the same transaction, Remember they've paid for it. Here's your shipping details and you can click on that and you can come back to the site and check on that at any time to find out when that's going to be arriving. You could even sign up for SMS reminders about when that's going to arrive. The funny thing is, with retailers, you buy something, it says that's going to take three days to arrive and they check every day, they will check three times. We could tell this from the data. Even though it arrives on the day, it might even arrive early, they still want to know. So whether or not you deliver on time, you still need to give people this ability to serve themselves. So that's that customer experience, end to end.
Nathan Bush: And are customers going through from that initial inquiry? Find me a kayak all the way through to transaction within the chat, or are you sending them back into the website and then back into the chat along the way?
Tim Warren: We love doing it entirely within the chat, so that's certainly our capability. And different. It varies with every retailer. They have something different they want to do sometimes you've got, if you've got something complex, it might be they want to show a web page or maybe there's some unusual options they want to pick. But we can achieve all of that within. Within the chat interface.
Nathan Bush: Yeah, great. I could imagine that just like ChatGPT, you know that the results that you get are only as good as the query that you give it. And we're becoming better at querying ChatGPT. If I take that as an example, do you find that there's a learning curve for customers to understand the power that is behind this chat compared to traditional search bars or traditional chat that they might be used to interacting with?
Tim Warren: Yeah, 100% if you think of that. Search was, you know, first you had, we had, you've had to fully navigate and then the big improvement was search but search was a bit clunky and we learned a little bit around there and it got much better and there's some great search providers now. Then the next step was the first level, I guess of conversation where you had to be quite good and yes, a few people got quite good at it and once they've tried the chat experience, people never go back to other channels. What's happened I think with GPT now is if you do like what Ambit does, which is to fine tune models for every individual customer. So if you're, say you're JB hi Fi and you know, you could give us your search history over two years, we can pull that in and say this is what people typically type and we train the model based on that. So this isn't, you know, this isn't chatgpt where it's all publicly available information and they haven't got any of that private stuff. We take the opposite way. We say hey, well give us the information that your customers do talk about and we analyze transactions and we analyze their databases and their stock and their product information platforms and then we can understand it and that the understanding rates with our pre GPT version we're usually 70 to 80% which, it's funny, it doesn't sound good but people then have another go at it and so it will get up to the 92, 98% kind of area. But now with ChatGPT type approaches, you know, you can type something to ChatGPT and it understands what you mean. It might not have the data though and that's where we come in. So we have that data. So if they ask about something really unique to JB hi Fi, what's the deal with your special where it's 10% of headphones. We can come straight back with that because you know what, we've got direct access into their database.
Nathan Bush: I want to come back to how we train up the chatbot and the data that you use to do this. But before we move on from customer behavior, are you finding that you're seeing that customer's queries are changing now that they've had 12 months working with AI and understanding this across a few platforms, are the way customers searching or communicating, especially in that first instance, is that changing?
Tim Warren: So yes and no. So some people, early adopters, are changing because they adapt. They're early adapters really, but now they're kind of changing back. And so the people who aren't changing are sitting pretty. They're like, hey, we didn't change our language, we just battled on through with the web page and now we can use the GPT powered version and it works a treat. So we're getting much more towards natural language and it's just a nature of something that you alluded to before. The massive amount of data that this stuff is trained on.
Nathan Bush: And as a general rule, what percentage of customers do you typically see going to chatbots to interact with as a percentage of total visits?
Tim Warren: Yeah, it's, it's enormously variable. So, and I'll talk about it with a couple of different areas. So one of the sectors, one of the other sectors we work with is utilities and most people who have utilities bills are older. Right. Than people that buy stuff online, so they are less likely to use it. So you might get a 5 to 20% of people coming to a utilities provider are using that chat service. Now that quickly goes to 100% when there's a problem because if the phones go out or if they can't get through, this thing never runs out because it's of unlimited capacity. Essentially when you're talking about retail especially that's focusing on a younger cohort. They're much more likely to use it. But there's a little subtlety in there which I'd highlight. So we will get 50 to 60% of first time users might try chat. It doesn't mean they'll exclusively use it and it's not a fail if they use different methods. That's a win because that's their experience. There's a bit of a misnomer. People talk about substitution. Oh, we want to substitute out of phones onto website and they think it's a failure if someone uses something else. No, that's a win if you've given people that opportunity to do what they want. That's powerful. It's only a failure. If they tried the phone and it didn't ring, that's a failure. But you never have that failure with automation. So, yeah, there's definitely a subtlety there, which is once people have a good experience, almost everyone, and we're talking in the high 90%, they come there and they do their entire next transaction through chat to the level that's possible. So that's what tells you about the future. So if it suits young people and young people, I've been doing this six years now. The people who were teenagers before are maybe buying houses and having families now. So they're now talking to banks and getting utilities where six years ago they were buying a T shirt and glasses. So the cohort changes and those people are coming back and we work with the big postal and big logistics providers and there is an increasing, I think we are sitting at around 60% repeat usage. And these are massive companies, right? They're sending hundreds of thousands of parcels a day and so much of their volume now is flowing into this trusted method.
Nathan Bush: So if we have retailers thinking about this and they may not have that younger customer, they may have older customers, but still want to look at how they can use GPT in their chat conversations.
Nathan Bush: Do you recommend almost like an education
Nathan Bush: piece when people experience that chat for the first time, like, hey, did you know that the chatbot can answer this?
Tim Warren: Absolutely. So there's a couple of ways of doing that and I'd liken it to when people first added email or websites. So you don't hide them away. If you hide them away and don't tell people about them, people won't use them. So what you do is you tell people. And we just had, I feel pain, physical pain, and I had sleepless nights. If we ever have a customer churn. And while there's been very few of them, sometimes they do and it means we haven't completely met their needs. But one customer who churned a little while back, they had their phone number in about 50 point bold on top of their website. And the only place that you could get to their chatbot was if you scroll down to the bottom and said, talk to a robot. And this is a business that its average age of user was 60.
Nathan Bush: I mean, full credit for not hiding your phone number. Like that's always. I love seeing that. But there's a bit of balance to be had there right there.
Tim Warren: There's a bit of balance. So what we say and this is where we take people on the journey and we talk about that multi channel, total experience, right? So give people access to the phone number. And I do applaud that. Give people access to any of the channels they might want. Unless you're going to retire it, like maybe hide your fax number, right? We have had examples where people will put a fax number in preference to showing their chatbot. Say, I call it that the Harry Potter style of parenting, of parenting a chatbot. If you have a child and lock it under the stairs, it's not going to get very good at stuff, apart from magic, maybe. So that's not a great approach. So we take people on the journey. So if you're a customer and you get something about a bill, then maybe on that email you could say, click here to talk to our digital employee. Now you do that, you're going to convert 20% of those people. Now you make the Avatar 50% bigger. You get 50% more people using it. Now, I'm not saying make up the whole screen, but that education piece you asked about, we can do things like pop open the panel and say, hey, I'm just like, chat GPT. You've probably tried that out. Do you want to talk to me about anything on this website? And we could pop it up a panel a third the size of the screen if they want, have some fun, ask me anything. And you can ask it something. And it's like, hey, look, the example we say is like, we can talk to you about Santa, but we're not going to talk about Donald Trump, that kind of thing. You can have a bit of fun. And the really leading edge players in this space, they are not just trying to get money from their customers, they're trying to bring them along on an exciting experiential journey. So you'll have them pop up something in a chat bot. Potentially one of the ways to move people into that chat channel is the old promotion, use the chatbot and get 5% or 10% off your next purchase. They just do that. And we always get the same thing. You'll get people who are upset at the business process where they wanted something and it wasn't available and they'll say, oh, that was dumb. And people say, oh, people think the chatbot's dumb. It's like, no, you ran out of kayaks, right? That's not a comment on the chatbot, but when it goes well, they love it.
Nathan Bush: I love what you were saying there around how you introduced the chatbot, because your first example was Come talk to our robot. That's obviously not appealing for anyone. Like, I couldn't think of anything worse than coming in and talking to a robot. Then you talked about coming and meet our digital employee. And I'm like, that's a little bit better, but it's still not that appealing. And then you talked about having a bit of fun with it, giving it a bit of personality, bit of pizzazz. Is that the approach? Rather than trying to pretend that it's a real human answering questions?
Tim Warren: So we never, not one of our instances have we ever pretended that it's a human. We always put some text in there saying that it's digital in some way. But what you're talking about there is in that education that entertainment is the essence of Persona. So you can have a digital Persona saying that. So you can say, hey, I'm your. I'm your store assistant. I live in a virtual world, but I can send you real things. You can say that. So we want to process with people. And this is what you don't get if you buy something cheap, off the shelf, right? So talking our book again here, which is if you engage with a, With a. A great partner, they're going to. What we do is we workshop that with people and we come up with a Persona and avatar. We can even animate that and have it running around the screen and doing some really cool things. So on our, on our webpage, for instance, when someone comes, you know, it bounces up and down. If you go to Contact Us, it puts a phone up to its head, and all of a sudden you've got a different thing. You're talking about experience. And that's priceless because people say, did you see that? I went to the JB hi Fi. They're not a customer, by the way, but I'd love them to be. I went to the JB hi Fi website and, man, have you tried their chatbot? It's so funny. And you're right. You can call that. This is your platform, right? So if you're a retailer and say you, you're. You're selling beauty products, what do you call that? Your digital beautician. We work with, believe it or not, a. A chain of vet clinics, and they talk about a virtual vet nurse. Now, that's totally up to the retailer. Have some fun. Some people, you know, I know there's a, There's a, There's a bunch of banks, they changed. They didn't have branches anymore. They have stores. And you go into the, into the bank store and you pick up a product off the shelf, which could be a personal loan, and you hand it over to them and it's got instructions in it. That's what they're people are, you know, having some fun with that kind of stuff.
Nathan Bush: Yeah. And then the bank bot charges you for the service. I'm just being silly. So let's talk about the data and the training upside of an intelligent chatbot like this. I want to start with the data that you require because you've touched on a few there. You've talked about inventory, you've talked about past sales, you've talked about customer records and actions. What is the key data that you need to get started with?
Tim Warren: This is where it gets complicated. I love this topic. Should we get a bit deep on this one?
Nathan Bush: Let's do it.
Tim Warren: Let's do it. You've got different domains of data, right? So the way we teach machines to talk our language is we give it heaps of stuff. And while it does matter to a certain extent what it is, the more is generally the better. You're not going to be able to teach a chatbot really well if you feed tons of medical journals in there, but it's still going to be able to be. Still going to be able to speak quite good English or other languages.
Nathan Bush: Does that data have to be live
Nathan Bush: data or could it just be archives of data?
Tim Warren: No, it's a snapshot. This is the archive thing. Right. Lots of people talk about, oh, chatgpt, can't believe it doesn't use live data. That's just not the way these things work. You build a massive. A monster model so complex that humans don't even understand it. Right. It's essentially code that's built code. And that's. That's the LLM, if you hear about that. The Large Language Model. Large is a euphemism. These things are monsters, absolute gargantuan monsters.
Nathan Bush: See, that's more catchy. Turn that into an acronym.
Tim Warren: Yeah, I'm sure we can. Gargantuan Language Model. That takes us onto another track. We'll leave that for now. But then you get onto. So you get the snapshot data. Now, very roughly, let's talk about a snapshot of the Internet. And let's cut out the images, let's cut out the medical papers, the things that are too scientific, the things that maybe there's a list of inappropriate stuff. We don't want to train it on things with an overly negative character or something like that tone, but you put lots in. That's how it Learns two things, facts and information. Well, it might be facts, it might be alternative facts too. It learns information and it has no judgment as to whether it's true or not. But most importantly, there's language. Now, the language is primarily what you learn from those very large language models. Models. Now, when people ask, chat, GPT or something like that, a question and they want data back, that's a bit different. LLMs are not specifically trained at being able to be brilliant at sending you specific data points. That's a different type of data and that's contemporaneous, live. It could be private, it could be public. So you've got this public data. Think of it as the whole Internet or all of Wikipedia.
Nathan Bush: Teaching it to speak.
Tim Warren: Essentially, that's just teaching it to speak. But like in the Harry, we'll go back to the Harry Potter model, you've let Harry out of the, out of the cupboard and he's hanging out with some kids and he's hanging out with his parents and one day Harry comes out with a swear word, doesn't he? That's because he's learned everything and he doesn't know, he doesn't know what's good and what's bad. I remember I've got twins and one of them wandered up on days, he's all of, you know, 19 months old or two years or whenever it is, they start talking and can I swear on here? Yeah, of course, yeah. He leans casually against the, against the door. Remember he's about 2, and he says, for sake. And so now, he never said that again. Well, he says it all the time now because he's a teenager, but he didn't say that again. But he didn't know what it meant, so he was experimenting with it. And that's what happens when you train these large language models. They don't have native understanding, but they experiment, they use things and they work out what does work and what doesn't and then they start to get data on it. So the reason OpenAI opened that up is to start testing the language model. They knew that not all the data in it is accurate. You can ask it questions, lots of questions, and it's plain wrong. They know that. Not a surprise. You also get the thumbs up, thumbs down. You click thumbs down, and then not a person, but a machine goes and tries to work out what's going on. That's the background. Think of that as the cosmic background data that you just need so you can talk. Now, incidentally, this, the magic that I mentioned before, the magic happens if you train it on multiple languages, it works out how to translate. This is not taught. This is emergent behavior. This is the magic bit. Right?
Nathan Bush: So we're starting to give Harry's wand.
Tim Warren: This is where Harry knows what to say with the wand and he's never been trained. Right? This is getting pretty exciting stuff. So that's where the magic starts happening in that layer. Now then you augment it with specialist data. Now that might be scraped from non private but non public information. Non private means it's not personally identified to a person, but it could be lots of transaction details for a large retailer, so they know how transactions work. That tells the. This is called fine tuning. This is where you build another layer of data. This is where the power starts really coming in. So when you talk to ChatGPT, it's really good at doing general stuff until you go and, you know, review it. What I say is it's excellent at writing fiction, right? Remember that? It's excellent at writing fiction. Harry Potter's fiction. There we go. We're carrying on with that analogy.
Nathan Bush: Oh, we're right in this analogy now, Tim. We are right in.
Tim Warren: I love it. So then you've got the stuff that is specific to. Now you can have data in here. Now some of it's. Some of it again is the snapshot data of what's happened in this retailer over the last few years. Not for too long because it won't be relevant from 10 years ago. People weren't buying the same product 10 years ago. They were probably talking roughly the same same with different emojis. So then you've got that current data, it's your own set of data now that makes you better than ChatGPT. Immediately you can do things ChatGPT simply can't. Then you augment it with personal data. Now this, for instance, could be all of those weird and wonderful things that Nathan or Tim has bought online. So what if I said to one of these big electronics retailers, go to JB hi Fi and I say, hey, I want to buy something cool for Nathan's birthday. You could actually say we had. You know, you do that. You haven't done this yet. I recommend it. You build up your birthday list and you send it through to me and I can see what you want. And I'll be like, oh, yeah, look, Nathan wants a new iPad or something like that. Well, with that specific personal, personal data, we could be recommending things to that level of granularity. Now, we haven't done something to that level before because that's really jumping the shark there. That's pretty exciting. But we can certainly read your previous transaction history. We can see what you've spent before and now you can have a rich conversation like you do with a great customer services rep. What did I spend with you last month? We can answer that question and is
Nathan Bush: that data held with the retailer or with you?
Tim Warren: So what we do then is we have for. So we have a mass language model and then we have a retailer dedicated model and then on top of that we manage the instance and the private data is only mapped in there as it's needed. But remember, live data, you don't really need to store it for the long term or cache it or archive. We can pull that out when we need it, add it all up, run the fancy GPT engine through it, give the right answer and we store that, say for 15 minutes. It's called caching. And then we, we delete it if we don't need it. Now for some we, we have just signed a, a proof of concept with a, with a large global consulting firm. I can't say who that is, but it's a very big brand which you would know. And they have data which they're so particular about security of that they run it in their own instance and all we do is send them the product for them to install it. We do not even have access to that underlying data. What we have, and we're literally building this in the room next door right now. We have a private set of example data that we use in our own environment. So you can go all the way along the line. Now if you're using something like Chat GPT and Nathan's pretty excited about this chat with Tim, so he, he copies in his, his history with JB hi Fi and he pastes it onto ChatGPT and says how much did I spend last month? It can answer it. But where's your data then? That's what you worry about.
Nathan Bush: And so then is there any live data being passed back and forth?
Tim Warren: The rule of doing things in a proper privacy way is you only process it at one level, process at that level, you don't pass down. So you never train off live data. There's a couple of reasons why. Because once you've trained it's historical data so it's not live. So you're better off going and getting an updated set. Now you might. We've always, you know, for six years of running we've had access to databases and we still have that data. If we pass it from a database into a trained, fine tuned GPT model. We can answer really good questions on that. We don't need to store that for the long term because you know what, your data doesn't actually help answer my questions anyway because your cart's going to look different to my car.
Nathan Bush: 2023 might be the year that you have to unravel everything that you know about loyalty. New global research from Shopify showed that Australian customers are highly likely to switch brands in order to save money. And here we were thinking loyalty points will keep us together forever. While the battle for customers hearts will be driven by pricing in 2023, they will not sacrifice customer service in the process. 2023 might be back to basic retail, but it doesn't mean that the opportunity is any less. Treat them lean and keep them keen. I say to view more resources to help with your 2023 planning and see how Shopify can take your e commerce business to the next level. Visit shopify.com au today.
Nathan Bush: So how long does it typically take to train up a retailer's GPT model?
Tim Warren: So, yeah, good question. So if you were starting this from scratch, like say you just had a couple of academic papers and a whole bunch of compute power, it would take you maybe three to six months of training and it would cost you probably $100 million. Right. And you're probably not going to go and do that. Oh yes, yeah, yeah. And then everyone's going to complain. It's out of date. When we've explained that, that, you know, September 2021 figure or whatever it is that we're ChatGPT's got. And by the way, I, we've laughed a little bit about ChatGPT. It's a stunning, amazing tool, but it's a consumer tool and it's not for every business purpose. So what are you going to do if you're a retailer today, what you do is you, you work with a partner, someone like Amber or someone who's got experience in this area with data security, security and they've worked with retailers and that kind of thing. And you, you don't start from scratch because you probably haven't, even if you're Myers, you probably haven't got 100 million to throw at this thing. Right? Maybe Apple's got 100 million to throw at it. In fact, that's what Microsoft invested in OpenAI just recently, I believe was about $100 million. And most of that they get back because it's all going into training on Microsoft computers. Right. This is the big end of things. Right, so, so what you do is you use a pre trained model like we use now. We go and get. Essentially think of, oh, I'm simplifying a little bit. You get a copy of a pre trained model and you host it in your own environment so you don't have to do the pre training of it. So you're not paying for that cost. And then you tweak it and you trim it down and you take out all the things that you don't need. So again, to use that one we used before for a retailer's model does not need information about medical procedures in it. So you can strip a lot of that out. And that's, that's called either D training or fine tuning or a phrase that I use is turning a large language model into a medium language model. I made that up. So I'm pretty proud of that. Mlm. That's my hashtag. Man, I'm going to be famous.
Nathan Bush: I'm surprised you didn't go from a gigant Gigantuan data language model to a large language model. But then it gets confusing.
Tim Warren: It's going to be one of those. But either way they're all mine. I own all those terms, hashtags all the way. And so I'm going to be famous on your podcast, I swear. So, so you trim them down, you fine tune them and then what you do, there's a whole lot of parameters that you don't see through ChatGPT when you're using these things raw, there is heaps and heaps and heaps of parameters. Call it. There could be, depending on the model, 10, 100,000. And then you tweak and train these and you put another data and you subtract other data to improve it. And that's what we do over about a week. And if you're one of those aforementioned retailers and you haven't done this before, it's going to take about a year. And the problem then is you're out of date because someone's done it. And that's the question. And something I'd encourage, obviously it supports what we do. When people are looking at technologies like this, you know, do you want to be an expert in building large language models and you hire a team of PhDs like we do and become a software development house? Or do you want to be an awesome retailer? Because I talk to lots of awesome retailers. But this stuff is hard, man. It's so hard. I don't understand it. I've got a room full of people out there for that.
Nathan Bush: We won't put that in the trailer quote, I promise.
Tim Warren: This is so hard. I don't understand it. This is the thing no one person understands, that we have moved to a fundamentally different paradigm where it was possible for humans to understand before these large language model transformers. It was possible, and now it's really not.
Nathan Bush: Where do you sit personally with AI and the evolution of AI? Beyond Ambit, obviously, but all the ways that it's being used. Are you generally optimistic that AI will be used for the greater good, or do you have legitimate concerns around where we're heading?
Tim Warren: I'm going to give you an example with hats, because I like hats, right? So when I'm. I'm a bit older than I look and I was around, I like to joke that when they turn the Internet from color, from black and white into color. And so in the 90s, I was coding in the Internet and back then everyone was awesome. If we wanted something, we just write to, oh, he's this famous person. I'll just write to them. You guess the email address, it would be like, you know, bobm.com or something like that. And you were probably right.
Nathan Bush: Steve@apple.com.
Tim Warren: yeah, it could be something like that. Steve at Apple. And everyone was very trustworthy. And that's what I call the white hat phase. White hat hackers, right, are the ones that they're hacking for good. They're trying to find vulnerabilities so they can fix them. They can't be exploited. Then you get the people who are saying, okay, well, let's do the right thing, but we're gonna, we're businesses, so we're gonna start building things. And before you know it, you start drifting into the, what I call the gray area. They're doing legal things, but they're really pushing the boundaries. And one of the things I remember, I don't think I invented it, but there was. I was around when landing pages first turned up. And you, you'd register a URL for one purpose so you could track people. This is before Google Analytics. This was before Google, right? I'm that old. I don't look at it. I've got an amazing filter here.
Nathan Bush: What moisturizer do you use? It's amazing.
Tim Warren: We'll talk about that later. It comes from one of, one of my prospects. We'll sign them, then I'll tell you, lifetime supply. It's a door beauty. So get that one out there. So all of a sudden you've got people doing, really pushing the boundaries. They're the gray hat people. Right. And then the black hats come in and that's when the first spam started turning. I remember I got the first spam spam email, email I'd seen. It was like 1997, 1998. I'm like, what is this? What is Viagra?
Nathan Bush: You know, don't pretend you don't know.
Tim Warren: Wow. I was in my 20s, I didn't need to know at that stage. This has gone an unusual direction, hasn't it? We haven't signed Chemist Warehouse. I'm just saying that, look, we've got
Nathan Bush: a whole hit list after this. I love it.
Tim Warren: Yeah, we, we do. So this thing came through. My first reaction as well. I believe I was one of the good guys. I was the white hat. I wrote, I reckon this again was the world's first spam filter. And it looked for too many capitals. These are capitals to small letters ratio. And if it was beyond that, I just moved it into the trash. Worked really well for a couple of years and then they got a bit clever. Incidentally, you know what's really good at filtering out spam is GPT is amazing at filtering out spam. So that stuff's gonna, it could single handedly kill spam. So when that started happening, we got into the black hat world and you got the, you know, the, you got the dark web and all this kind of stuff. I think AI is going to go in the same cycle. So at the moment there's been a bunch of the white hat people around again. I'm, you know, I'm at the forefront of the good guys. And it comes from this metaphor, comes from cowboys. The good guys in westerns always wore a white hat. Right. And the bad guys always wore black hat. So I added the gray one. I have all kinds of hats. I've got lots of other things. We'll talk about that on a different one.
Nathan Bush: Harry Potter hat, the whole lot.
Tim Warren: Oh, there's the wizard hat. I hadn't thought of that.
Nathan Bush: There you go.
Tim Warren: So then I think I called the time. When it was released in November, I was bowled over. Very excited, but I was bowled over. This was the secret that we had was this mindblowing technology and they turned it over to the world. I couldn't believe that. What they needed is they needed everyone's training data. Right? That's really clear. And now they're getting it because everyone posts in it. Unbelievable stuff. I'd love to see the logs, by the way, for ChatGPT. You know, people are writing, they're writing summary judgments of criminal court cases. And they're writing wedding vows and eulogies and the whole gamut, drafting questions for
Nathan Bush: podcasts, all that kind of thing.
Tim Warren: Oh, that kind of thing. Imagine if people did that never. So when that went out, I realized, hey, we've moved from white hat, where we're all good guys, to gray hat because there's a little bit of time now and I've already seen. What you've seen is a proliferation, an explosion in the number of fake ChatGPT signs. They do that, they just pass it through to ChatGPT that does the work and sends it back. And then they've got your data, they're just stealing your data. And all of the apps that are out there that they say we're the official chat now GPT. No they're not, they're just stealing your data. Right. So we've. It's not completely, they're not stealing it, it's not completely illegal. But it's gray. And if, you know, if there's gray people working, then you've got the black hat people using it as well. And there will be sensational uses of it that will blow me away. I won't have thought of it. But now we've got this weird tool because AI is very good at this. Do you know AI can find vulnerabilities in code? So now you've got the white hat people again, it's always us doing the hard work, the white hat people running over their existing code. Not only the code that's not related to AI, but the AI code itself. So now you've got AI improving AI code. So am I optimistic? Yes, I am. I'm realistic though. So I think there will be. Think of it like nuclear power. It can be relatively cheap energy, but when it goes wrong it can be quite a problem. Right. So yeah, so I think there will be some really bad issues that happen. I think regulation needs to run a lot. There's none at the moment. I think the industry is very smart to try and start self regulating. They haven't done a great job. I'm part of that, I'm encouraging that. So I'm on a couple of bodies for pushing for proactive self regulation that we. Where we can get governments to take on appropriate regulation. It needs to be global. There needs to be global accords on this stuff. So that's the stuff that I get. I'm not going to say compliance is exciting, but at least we can write the policies with the tool itself.
Nathan Bush: Yeah, yeah, yeah. You're right at the starting line, Tim. I'VE got a whole bunch of questions here that ChatGPT have. No, I've, that I've written that I wanted to get to, but we are out of time. I want to get your vision now on where you think E commerce customer service is going to get to. Say if we put a timeframe on it for the next three years, we obviously know what customer service online looks like at the moment. The really good ones have a great mixture of chat, phone, email, really great self help, huge amount of product data, description. You know, we all kind of know what good looks like. Where do you think the benchmark for good will look like in three years time?
Tim Warren: I think it looks a little bit like the car industry. So what tends to happen is something comes out on a Mercedes S Class, that's the first thing in the world. They embed everything like ABS. After 10 years, ABS is available on the base level model. So I think it's going to be accelerated. But if you want to see what the typical retailer looks like in three years, probably look at the very best retailers today. That's what I would say. I think three years is probably about there. So if we want to bounce from what typical looks like to what good looks like, it would be all of the great tools today that are being a bit underused. And there's going to be some things just around the corner that people don't necessarily know about. The leaders are going to be using this. And I'd say, you know, some of those brands that I've called out here, I've called out brands. Some of those are people that we work with, Some of them are people that I don't work with, but I really admire and they will be innovators. So let me look at, let me look, maybe tell a story of what a great experience might look like for a clothing retailer of which we've got a few. So I don't think you're talking to a nameless kind of chatbot that's over there. I think you've probably got it on your phone and it knows you really well. So it knows more than your retail data. I think, incidentally, let's break out of this first to the direction of assistance. Assistance work for people. So I think in 10 years, maybe even less, maybe seven years, I will have an incredibly smart, dedicated LLM assistant that works for me and it talks to everything else. That's where I think it's going, retail. We won't be doing that in three years though. There's a few things hardware has to Change for that to work. We can get onto that another time. But retailers are going to have a brilliant chatbot that knows you really well. It probably lives in your app and when you, you probably take selfies with it and you take photos of things, like things in your closet, it's going to match for you. It's going to look in your calendar and say, you've got this event coming up. The classic thing is, you know, Helenstein Glasson have said, you know, we want to make, we want to make this the. It's the shop attendant of the future. So it can, you can talk to it about an event you're going to and you can go to themarket.com and you can say, I'm going to wedding. And the Market.com can say it can look in your. Or the Market.com wouldn't do it, but it can give recommended top wedding examples by having a better understanding of your needs. But it's not really about the retailer. It's about the individual and it's about the customer experience of kind of the end to end. Think of it like the relationship. I'm talking to you on an Apple Mac with an iPhone here. And I have a relationship with Apple, everyone. Apple's the first company to pass 3 trillion, right? Like it was 1 trillion a couple of years ago. And that was amazing. People are going to continue to seek that. And I think they didn't come up with the word for it, but Apple created this experience just like just touch one of their devices and it's better. It's amazing. So how do you create that? I've heard of some great retailer. The names escaped me. And every time they send out a package, they're an Australian retailer, they send out a handwritten note from the CEO. It's just extraordinary. And then people post that and they get 100,000 views on a handwritten note from a CEO. You cannot buy that level of social. But stop thinking about clicks, start thinking about the actual user. And in that example, before, you asked me about the moisturizer, which I'll tell you about later, but you imagine the beauty app built in. You know, it's on your phone, it's with you. And it's much more about understanding your lifestyle. And if you buy a pair of running shoes, maybe you need something that stops you getting wind and sunburnt lips. You know, little subtle things like that. Do you know how much utility value that adds to your life? A lot. But you invest that back into the brand that brought it to you. So I could. I love talking to really innovative CMOs. That's where it's going. And so I'm not going to point at a single one off technology, but it's going to be conversational, it's going to be interactive and it's going to be personal.
Nathan Bush: I think you should go build the product that creates handwritten notes from CEOs and sell that into companies.
Tim Warren: I could try it. I'll tell you what, I will tell you. I do a little bit of that. When it comes to Christmas, I do like to send a handwritten card. Well, I sign it, right, I sign it. But from the CEO and the whole rest of the team, everyone contributes. But I do love to do that. It means a lot to me. Yeah, yeah.
Nathan Bush: If you could outsource it though, you would, right?
Tim Warren: I'm not going to make you anyone
Nathan Bush: that's got a Christmas card or present from you. Tim, thank you so much for joining us on Add to Cart. It has been such a great conversation and I've learned a lot, especially around the different types of data and how to train up a model like Ambit GB GPT. Really fantastic. Thank you so much. Now, if we've got people who have heard our conversation, want to pick up on a few points or potentially look at how Ambit GPT could help their business, what's the best way for them to get in touch?
Tim Warren: Yeah, yeah, absolutely. So look, I'm always around on LinkedIn, so that's one of the best ways to get in touch with me personally. I'm just the, you know, the guy that wears the orange shirt around here. If you come to our website, Ambit AI contact and send us a message through there, it's going to go through to the right person in the company. But I'll tell you what, if that person's me, then I'll be picking up the phone as well.
Nathan Bush: Beautiful. Tim, thank you so much for joining us on Add to Cart.
Tim Warren: Thanks, mate.
Nathan Bush: Now, if you're thinking about giving your customers a chatbot to call their own, you may want to take advantage of the free ROI analysis and savings estimate for your contact center or support team. Go to Ambert AI forward slash contact to get in touch and quote Ambit roi. Add to Cart to claim yours. Alright, here are the three three lessons that I took from the conversation with Tim. Number one, the GPT fundamentals. Tim did a brilliant job of clarifying some of the basics of what GPT or ChatGPT is and is not. GPT or Generative pre trained Transformer is the underlying technology. ChatGPT is the product owned by the company OpenAI. There are a of number number of levels of AI when it comes to chat, teaching it language, teaching it to speak, teaching it context and then giving it personal information for individual context. You might even detrain it to strip out information. This is just tip of the iceberg stuff, but I'm so glad he was able to explain it. Simply to kick us off number two no boring bots. When implementing a chatbot on your your site, don't think it's a pop up box and done. Tim shared some great tips on how to set up a chatbot for success. Give it an avatar, even a name for personality. Don't pretend it's human, don't hide it. Make it obvious to see. Tell the user what you can ask it. Make it fun. Boring robots are bad robots, White hats, gray hats, black black hats and wizard hats. Tim's analogy to the early days of the Internet and email made a lot of sense. While we are in the early days of AI and it is exciting to explore the white hat possibilities of the tech, we also need to be thinking like black hats and the vulnerabilities that technology might open up. Don't just think about all the opportunities, think about the defense as well. The black hats are coming to get the highlights of today's episode. Head on over to add to cart.com and sign up for our free newsletter each Tuesday. We will send Monday's episode summary links and discount codes for you to go next level on. And if you're looking to explore your next e commerce opportunity, come and visit us at E Suite. We're a dedicated e commerce talent agency connect connecting the best e commerce talent with the fastest growing brands in Australia. Head on over to esweettalent.com where you can download the free e Commerce Salary guide and sign up to our weekly e commerce job emails. Thanks for listening and until next time, keep those customers adding to card.
