AI Is Building My Cellar | Episode 11

0:03Jonathan Stevenson:Welcome to Uncorked, a Cult Wines podcast where we take you behind the scenes of the fine wine market. I'm Tom Gearing, CEO of Cult Wines and your host of this podcast and I'm as ever joined by my co-host Jay Stevenson. How are you doing Jonathan? Very well thank you, very well. Good to see you Tom. Good to see you too. It's been a couple of weeks actually, we'll get into that a little bit later on. This episode we are very happy to have with us Damon Segal. Thank you very much for joining us. Damon is an entrepreneur, a marketing specialist, but in particular for the focus of this podcast, he is what we would call an AI specialist and technologist, especially when it comes to the idea and the concept of the future of wine.

0:42Jonathan Stevenson:But he's also a passionate wine collector and he's been exploring how technology can support buying decisions and portfolio building. As you know, with this podcast, each episode, we try and unpack the key trends of what's going on in the wine market, the trades, the talking points, and we often try and get industry experts who can bring a different and new perspective onto the podcast. So we're very excited to have one that's talking about technology, because I know it's something that's at the forefront of a lot of people's, you know, what you see in the papers and read online every day. There seems to be a lot of talk around AI.

1:14Jonathan Stevenson:So today we are exploring AI, we're exploring artificial intelligence, and in particular how that's starting already to change the way that collectors engage with wine and i think potentially today will be a lot around the future of how ai can advance the fine wine market the fine wine journey the fine wine experience um and how it will subtly start to reshape the future so hopefully that i've hopefully i did you okay with that introduction are you happy with that you did yes thank you and hopefully i'll be able to answer the question exactly yeah you've got to live up to the expectations now we've set the bar high so all good all good so thank you for that I'm very excited to be here and I'm looking forward to delving into the world of AI and how it will affect wine, which is not my normal discussions around AI, but they are changing everything everywhere.

2:02Of course.

2:02Jonathan Stevenson:Yeah. Yeah. I mean, the rapid development we're seeing in a moment, it's just it feels like something new every week to get yourself stuck into. Jonathan, I don't know if you wanted to kick things off. Yeah, I was just already starting to think about where, you know, where we can take this. I guess with regards to AI, was it a case where the wine aspects of how you're thinking and using AI came as a secondary? Did you say that you're already starting to get familiar in use of AI in other projects? Yeah. So, I mean, we've been working in AI for quite a while. So I have a tech business that had been working in AI about 10 years ago with a Google AI, which was arguably machine learning.

2:46And it was all driven by developers. It was really hard to understand. And then two and a half years ago, ChatGPT came out. And effectively, I thought, this is cool. So for my marketing agency side of things, I thought, let's see what we can do with this. And we started getting involved. and I used to be on the speaker circuit around 2005 for six or so years and some people asked me to present what we had done with AI and so I thought okay well I'll show them what I'm doing and they thought this is so cool you need to tell other people and all of a sudden I started getting asked to all kinds of places to talk about AI generally in hospitality quite heavily definitely within business processes and things like that and then how it can affect marketing as well and really understanding consumer behaviors and things like that so the wine side of it was just literally my own portfolio and starting to make decisions about what I like and you know we are in danger of outsourcing our brains for decisions so it's you know it's quite important to be careful about that but i'll generally take a list of uh wines from a merchant drop it straight into my i have a trained gpt which knows my preferences because i taught it all my ratings

4:06Jonathan Stevenson:from my vino using using an api in the back end with code or actually on the front end chat no even much simpler just a front front end chat so i built a custom gpt uh which if you have a paid account these are they live on the side and every time you start a conversation with them they know all the knowledge that you fed them so essentially i just downloaded my spreadsheet from vivino dropped it in there gave it a few other instructions about what i like and what i'm trying to do and now i can literally just take a list drop it in there and say do you spot anything good in here for my tastes and actually any good investment opportunities as well have you had any success so far with that yeah anything that comes to mind oh god yeah compared to my my portfolio that I bought probably at the exact wrong time, which was, I think, 2019 or 2020 or something.

4:53And fairly heavily Bordeaux-driven. Yeah. So you can imagine that took a lovely hit and my 8 % a year forecast did not exist. But actually, it's helped me diversify the portfolio really effectively. Yeah. And gives pretty solid information. I do double-check it with the various people that I work with. but ultimately it seems on on point so but there's always an insider knowledge that's you know there's still the human aspect of it um i was looking at uh getting some sassakaya so so after the the digging and the research you know it's come back and basically said tuscany is the place to invest at the moment uh whether you agree i don't know but we'll see ai might be right um and i was looking 2019 or 2020 it recommended the 2019 in the end again who knows if it's right and uh and then my contact came back and said well hang on for the 2025 on promote because actually the launch price could be good on that and that's a human piece of knowledge that maybe ai doesn't have yet yeah so you can't rely on ai for everything yeah but actually definitely builds a much more sophisticated picture for me than my own knowledge yeah yeah interesting i think i think obviously the whole thing with ai i mean we we're finding that in different kind of use cases within what we do i think it's it's all that ai is quite a broad uh you know it's quite a broad subject and obviously i think really the the true detail and power comes from how much information you provide AI right so you've obviously mentioned there that you have some prompts and and and that will build over time in terms of developing what you're looking for but I do think you're right I think giving you know giving AI just a complete free reign of of all the information that's out there in the internet is is one thing but also really fine-tuning how you're approaching the prompts over time I think is how that will evolve yeah you know better stronger I think from our perspective it's about making sure that you know all of the data and information that we are kind of starting to look at sharing with with ai and then and then the learnings that gives us back is is really like the the secret source to to kind of getting the most out of it yeah 100 i mean it's um it's interesting so a lot of my stuff is is quite corporate conversations around ai and uh i mean there's a there's a line that uh i had in a conference recently which was um the world's drowning in answers you know we we there as long as you have a question you can get an answer but the problem is everyone else can get the answer as well so it's a question is knowing how to ask the right questions and what problems to solve and really the only value now companies have is our own data so you got your own bespoke data means that you have a knowledge that nobody else can touch because everything else is out there and anyone can access it so keeping your own data detailed and clean and everything else now is becoming like one of the most important things corporately and for an app like yours you know the value is incredible because you're collecting tons of data i would imagine on people's trades and everything else which you're not then presumably sharing anywhere into the internet or accessible by ai so therefore you have an advantage with your data so yeah data is like gold 100 yeah i mean it's interesting i just pick up a

8:35Jonathan Stevenson:couple of things one firstly from jonathan your what you said i think i mean anthropic obviously um you know very good on claude yeah i mean obviously it's kind of on trend at the moment but a couple of weeks ago especially around this article that came up around sort of you know the industries are going to be impacted by um you know ai and especially agentic coding but you know they did an analysis of looking at all the different sort of industries and how they'll be affected and it's interesting because i think with wine there is an inherent protection i think around uh how ai will be used in the future and and around around the industry because fundamentally ai doesn't take a physical form as of today maybe in the future but you know let's say for now we find when it comes to wine ai will always be reliant on human data sets as the initial foundational points so i think to your point where the opportunity and the um focus is especially in the wine market is around making understanding a way you can collect the data and then storing it and cleaning it and have it in a really uh you know easily accessible way but i think it's interesting just kind of like thinking about the application of ai in the wine industry because you know an ai can't tell you what a wine tastes like but it can uh you know it can triangulate a lot of data and as you said it can then use your own personal data as well as external data to then maybe make suggestions by using unstructured data and coming up with these sort of connections to be able to enhance the customer experience, enhance the knowledge that a user might have around wine.

10:01Jonathan Stevenson:But from your perspective, do you think that, you know, the wine industry over the long term will benefit from AI? Or do you see it as also a potential threat to the industry? And I suppose if you do see it as a potential threat, which areas do you think are the areas that are most likely to be threatened by AI? That's a good question. And actually the autonomy of AI, so robots and things like that, we're really not very far away from that. We're literally 12 months, I think, before they start being commercially available. Price-wise, they're going to be really low priced as well. So humanoid robots, you'll be able to pick up for$25 ,000,$30 ,000, or you'll lease them for$500 a month.

10:45that's going to have an enormous impact on everything yeah exactly so i mean you can already go online and order neo yeah you know um but you'll have the uh optimus one and the the figure one will be available i'm sure in 12 months and i think we're probably two or three years away from them being pretty commonplace uh so i think that will change harvesting and things like that because you know you can't use machines on steep hills but you could easily use a humanoid robot to go and pick the grapes and they can use special cameras to look at the grapes and see how good they are and if they're ready and yeah everything else so i think um so that will affect things but i think primarily it will complement the wine industry i think in so many areas you know vineyard management now using ai for devices to detect whether a vine is stressed or not and cameras flying over to see if they've got diseases i think that's brilliant and ultimately anything that can kind of help wineries now is a good thing uh with unpredictable weather and things like that it's going to be a huge benefit as far as people buying wine you know generally the confidence i think is a really big issue you know you it depends where you are in your wine journey but if you walk into a supermarket or a wine shop or something like that you know you've got hundreds of bottles in front of you and if you don't know that much about you generally just go that's a nice label and it's on discount so all right let's try that which is great for the wine adventure and the discovery side of things but uh again if you've got ai that actually knows what you like and things like that you can literally just take a picture of it and go which bottle on this shelf do you think I should try tonight?

12:30Jonathan Stevenson:I mean, just to dive into that, I mean, you did mention the app Vivino already, obviously something that you use yourself, but is that not probably the best position to take advantage of AI, especially in that kind of augmented retail experience where you described? I would imagine it's already using AI to do that. So if you think, actually, there's a really good analogy for AI and matching like taste profiles, you know, because AI understands that a wine has a flavor profile of this, this, and this, and if it's this age, it will do this, and if it's this age, it will do that, and it's got enough data points to do that.

13:04And I think the best analogy I've come across is if you think about Spotify, Spotify basically uses an algorithm to decide what music you're likely to like because it has broken down the genetics of that music. It doesn't know what the music sounds like. It doesn't really care what the music sounds like, but it matches the profile of the other things that you do. So therefore, the likelihood of AI being able to pick a wine that's ripe for you is actually quite high. Based upon your user input. Based upon your user input.

13:34Jonathan Stevenson:But your user input is also relevant to the time and the place, the environment, and how long that data has been relevant for. A hundred percent, yes. We've all ordered a case of wine from a restaurant where we thought, that's amazing, after your third bottle, and then found the next day it wasn't quite so good. But, you know, your tastes change over time and as your experience grows and you understand why more you might start to explore different regions. And also that's the other thing that I suppose was a question we had potentially lined up for later on. But is there a danger with AI that it creates too much homogeneity in terms of if everyone's using the same models, the same ideas?

14:05Jonathan Stevenson:You know, as you said, because it's, you know, data driven, you know, recommendations. Does everyone end up in a point where everyone's just kind of getting concentrated? concentrated i mean sometimes you have it when you're driving around london you use ways and you realize that this little shortcut that it's giving you every single person has been given the same one that now has a massive i started doing the anti-shortcuts now yeah like you go into a wine shop and every single person goes to grab the same one bottle of the out of the 500 on the shelf it's particularly on a motorway when you just like it's just coming up and you can see the traffic jam i'm sometimes now just sit in the traffic jam because i'm just like everyone else is going to try and do that and i just to riff on this point a little bit and just come back to what we said on a podcast a couple episodes ago like and we're talking about restaurants and sommeliers and how like you know top tip is like you know when you're in a restaurant use a sommelier that's what they're there for that's what they're trained for don't be intimidated by a sommelier but one of the greatest things about a sommelier is not pandering to your to your you know not just going oh you like italian wine i'm going to recommend italian wine one of the great things a good sommelier can do is give you the confidence to step outside your comfort zone right and there is from my understanding there are very little ways you could build even a python script or an algorithm that could recreate that sense look of course you could say something like there has to be a percentage of your actions you take where you're purposefully recommending that the user goes outside their experience and let's say when you go out into that filtered space outside of what they know you're going to prioritize things that have got good scores from other users let's say so at least you know in safe round but exactly yeah to a certain extent what you said is amazing but also at the other hand it's kind of like you kind of don't want it to be completely perfect because you want this homogeneity not to exist in a way and that is the biggest risk because essentially if you think about it if i turn around and say you know i want a red wine to go with steak tonight it's such a generic question that the ai has to give you a generic answer and if everybody will go to the quickest answer yeah and it will go through you know go through 50 000 vintages and wines and everything else and it might ask you what's your budget or something like that but ultimately it's going to come to the same three wines that everybody else does so it's when i run my my work is that just another just like is that not just an inherent issue of ai which is it's probabilistic in nature uh no it's again it's to do with the input so the thing about ai ai is is you know garbage in garbage out kind of stuff and the way i kind of put it is if you imagine ai knows everything about everything yeah if you asked it how much is an expensive bottle of wine you're gonna gonna get if i'm in a room and i say you know to this young man here who's fantastically recording us what's an expensive bottle of wine he might say how do you know he's recording us well we haven't heard the outfit do that again 30 pound yeah exactly how much is expensive bottle of wine uh seven and a half thousand pounds yeah exactly so there's no right answer yeah so the thing is ai has to just pick some middle middle road generic thing so but what i'm saying is when it picks that middle road it's also picking based on probability it's it's it's choosing those probabilistic answer if if you're using a and that's based upon the data it's been given yeah so the data it's given it's going okay what is the highest probability answer to the question i've been given based on the data set i've been given yeah so it will go out and it work out most people spend 30 to 50 quid in a bottle of good wine and then then return that You know, I've asked my daughter.

17:33She told me$6.99 from Tesco's and I almost died on the spot. So it's like, Dad, you have to try this. You haven't been doing a good enough job at home.

17:41Jonathan Stevenson:No, that's what I was thinking as well. You had to have a long hard look at yourself in the mirror. That's a deliberately bad job. I was really disappointed. Catch you awake at night. Yeah, yeah. But that's the problem. So you have to give it the context. So, you know, I want to spend X amount. I like easy wine. I like full-bodied wine. You know, whatever you give it as a description, it's not judging you. that's the nice thing about it so you know if i say i want an easy i want an easy wine that uh is really soft and i like full body okay you know it becomes a bit more challenging um but it will work out the best thing it can as to what you're trying to tell it so so whereas a similar day might look at you going oh that's not actually a thing i think the other thing as well and it's maybe a slight misconception when it comes to ai especially on the data crunching side is i think the misconception publicly is that ai can handle very large data sets now i'm not saying it can't handle large data sets but in the kind of everyday consumer use of ai if you have a 50 000 row database you think it's going to evaluate every single row and give you the answer it's not doing that it because a lot of the llms don't have large context windows and therefore you don't have enough memory to be able to contact so it will get to the quickest best answer as fast as possible and it will forget everything else and it will throw the rest out.

19:00Jonathan Stevenson:So to your point, it's a combination of, I completely agree with you, rubbish in, rubbish out. So you've got to have good quality data to start with. Then on top of that, you've got to have context, which is what you described in terms of your GPT, like my preferences or whatever you can store locally. But then above that, you also need a framework from which you can make sure that the LLM stays on a guardrail so that actually it doesn't either A, hallucinate, B, skip, or C, actually crunch data. And then the last thing I'd say is, And by the time we've done all that, actually, there are other tools that are probably better than AI to actually do that large data crunching.

19:36And then I get to that point of going, like, where is the use case for AI?

19:40Jonathan Stevenson:And the AI use case for me is more around the natural language processing, the customization, the personalization. But actually, when it comes to large data crunching, are we still not in the era when Python scripting is still the best way to crunch large data sets? Well, ultimately, the AI will write the Python script in the background. Yeah, it gets that point to give you that ability, but it's still using a more normal statistical tool. So the interesting thing, the context windows are massive on most AI systems. You know, even if you're using, I think, ChatGPT. I think Gemini's got the biggest right.

20:13Jonathan Stevenson:It has, yeah. It's a million tokens. ChatGPT's 400 ,000. But even 400 ,000 is still the equivalent to something like, I think it was 600 pages of a book or something. Is it that much? Yeah, it's massive, yeah. So no one ever appreciates quite how massive it is. You're 100 % right. I think with Gemini, the most I've got it to do is 50 ,000 rows of data. So the problem… And that's the biggest one. When you… So it's rubbish. Well, it's not rubbish at data. That's unfair. It's brilliant at crunching data, but you have to use the right models in the right way. So, for instance, if I was going to crunch data, I would crunch data using an agent.

20:52I wouldn't do just a chat.

20:54Jonathan Stevenson:Yeah, of course. So the thing is there's an agent mode on ChatGPT which spins up a virtual computer and actually handles spreadsheets really well. There's actually a new plug-in available in the States. I think it's not over here yet, which allows ChatGPT to work with Excel as well. So actually you can... Yeah, and obviously Claude Cowork. And Claude Cowork can do as well. Claude is brilliant at crunching data. so it's as long as you're not trying to use a free version on a base model you'll you can crunch data or you build an api connection to do it which is starting to get more technical i think the best way i would frame it for someone listening who wants to explore this is if you have a really large data set and you want the lmm lllm to do a certain calculation repetitively over time each time it processes the information you won't get a good result you you are better off creating an agent or a structure that does the calculation do you want it to be done if it's something that has to be done every time repetitively you might as well build that into the pipeline or workflow first then once you've got the output that is repetitively done and you want the inference the inference is where you get the value from the llm and i heard someone say the other day that really actually it's going to be interesting to get your view from a marketing perspective um you know that inference level for llms is really where the value uh proposition is if you feed it um your guard lines your guardrails like you mentioned a minute ago um in order to get it to give you a repetitive output uh you need to have that training in it in some form interestingly they launched i think it's only a week or two ago a thing called skills have you seen that yeah and mark down files yeah so uh well skills you can actually um you create something called a skill that will then sit there and be called any time in any chat yeah to process something in the way you want it processed so from a marketing point of view for example there are there's a methodology that we follow for certain things like seo and the like or auditing a website for example and we want it to use certain tools and look in certain places for things and now when we ask in a particular channel for a client for an audit of a site or seo recommendations it will literally pull this skill and use our methodology repetitively every time so that works really well um i think for an everyday person the real core thing is you know if i'm a wine merchant or uh just a retailer who wants to understand what's going off my shelves fastest and maybe what i should be ordering in and where the market's moving i i would use just a higher level of thinking uh so when you're picking your model and it starts with auto uh just don't use the auto one just go straight to thinking extended or pro which is just on the paid models but the output is immeasurably better so that's where i would go with it so i mean for for someone sitting at home wine enthusiast collector investor whichever your kind of main motivation is you know of course you know the tendency is now to do what you kind of described in terms of you know just going into one of the llms and kind of start building your kind of structure your your guardrails but do you see that as being the big unlock and if not where do you think the biggest opportunity exists in wine for the use case of ai and let's say for the betterment of the consumer experience for the consumer side of things so so for me that that well there's two elements there so one if we're looking at it from an investor point of view um for for me it's having the confidence to have more data more information about what i might invest in so um you know there'll be apps out there that will give you data and information i can go and understand but actually i can't have a conversation with it right um you actually really value the natural language conversation you're gonna have 100 yeah so for me you know i would give it an instruction a prompt essentially um and the prompts you might start off a really easy way to write a really good prompt is just to ask your llm to write the prompt for you yeah so i might say uh i want a prompt to um research uh where would be a good area to invest in wine at the moment for my portfolio.

25:21And then it will come out with some really clever prompt about, you know, go and look at all LiveX data, have a look at auction sites and what's been sold recently, look at all retail pricing on the internet, and go and do that research and assemble it and then tell me where demand is showing but the price hasn't been noticed yet, that kind of thing. And then it comes back with a whole bunch of suggestions. and at that point you start to explore it with natural conversation so for me it's like you know as i say it came back basically uh uh tuscany wine is is good for investment bordeaux is still soft uh we're we're still living in a world where wine hasn't recovered yet it's getting better um and these are the areas that you should look at sassacaya 19 was it sassacaya 19 yeah yeah well placed um sasakai 19 so interesting because it sounds as though you're to tom's point you're you you like the back and forth but then i feel i just hear all of that and i think that it almost requires additional context from your side to actually go well maybe don't use this or that even though it's suggesting certain areas that you know i guess you know working in the market as we do it's it's kind of that's maybe the the missing part of knowing exactly what parts of the data to use from LiveX or whether to use WineSearcher or to remove certain aspects of those.

26:51But I think it's, I mean, it's certainly interesting. And I think going back to what Tom was saying around that sort of homogeneity of like. Easy for you. Just ending up. Yeah, it was, wasn't it? I still, I wasn't sure you pronounced it right. I wasn't sure. Anyway, so yeah, I think it's a case where, you know, we work with it. in the market long enough where you know when you have conversations with some people that are working with certain merchants it's almost like you're going to get delivered certain wines that you know they're going to have suggested so thanks well i just think yeah you know what i mean so i think that that has kind of existed and it does lots of these kind of things that we we might pick apart for ai as failures or potential um you know challenges for what ai would would eventually give it's like well sometimes that exists in in the in the market anyway so i've

27:43Jonathan Stevenson:But yeah, I think what's really interesting about your point, and maybe I think what would be interesting to sort of see the kind of output you're getting on your GPT is, I think the biggest thing that wine investment or wine collecting has always struggled with has been the accessibility of the data, right? The data isn't accessible. You can't get transactional information. Obviously, we've been in this market for 15 plus years. And if I go back to my sort of family story, my dad started Financial Wines in 2001, which was the first online price aggregation tool for fine wine. So, you know, 25 years later, I'm still trying to solve a lot of the things that he was doing.

28:12Jonathan Stevenson:But, you know, if you look at Colt X, you know, we've got 300 million pounds of transactional data. Anyone can log in. They can see prices have been traded at the historical price data. You know, it's a bit off a platform. You know, you can trade with other people. You decide the price, you set the price. And, you know, for me, like we were trying to solve certain issues that have always existed in the wine market. But I think one of the benefits now that AI has come along to your point around sort of, you know, rubbish in, rubbish out is that that data now with the application of AI, it's like a tool that can really give you that that leverage and to jonathan's point it's you know i was actually you know i was speaking to our cto over the weekend an idea came to me which was like so we're you know we're working on like an like it's really interesting to hear what you're saying about the ai chat because i feel that some people you speak to don't like the idea that they're talking to ai so then you get into that problem of like going oh do we have to kind of pretend it's not ai but then i get to that point of going well everyone is always on chat gbt or claude or you know google gemini whichever one you're using or a combination of three and you're comfortable because the quality is good and i think really for me i find for me that's the angle it's like i don't really care that the ai has given me the answer back as long as it's being well tuned yeah i'm getting good answers back i'm not sure every consumers at that point no i'm sure there's still some consumers that will get something they go oh i think ai created that therefore i don't like it yeah whereas i'm like an ai has created that and they've told me something that the human couldn't have told me great fantastic they've used ai really well to give me an insight that i couldn't have got otherwise that's the way i think about it So it's interesting to hear your point.

29:42Jonathan Stevenson:But the thing that I was talking about to our CTO was, yeah, it's great having an AI advisor in the platform using all this data, but what would be better is if we had local skills file for each user. So when the user's on there, you can set your own skills file within. Because to your point, Jonathan, around sort of like, yes, you go to a merchant or you go to a chat GPT, it's going to probably, if 50 people asked the same question you did, maybe they all got told Sassacaya 19. And what we're going back to that homogeneity yeah maybe we'll go with it we'll go with it i have no idea yeah um you're up next that's for me kind of you know when i think about how we're using ai that's how i think about how we can avoid that issue where you're giving someone their personal agent within the platform yeah so you come onto codex every user has access to all this massive store of data we give you an ai tool that's connected to an llm that has a base level of knowledge but then on top of that your user experience you set your agent up you you give it your preferences and then each user's experience of that agent within the platform that's what access to unique data is then personalized so for me that again gets us to the point where you're getting the right answers for you yeah i think that really covers the trust issue as well like you said i think it's important as i mean we all are aware of how quickly it's developing right and to tom's point i think that's something that you see every day when you speak to different people right some people just are still just miles away from using it other people are really trying to explore it it's amazing everything in between and and i think to just falsely assume that everyone is going to be on the same pathway is probably going to be one of the biggest challenges or potential failures if you'd get that wrong um and i think back to when we kind of explored using blockchain um for storage solution and actually it kind of goes back to that same thing where the blockchain is only as reliable as the initial input and the ledger itself that was yeah was recorded at the outset and then from there it's it then becomes that you know reliable you know sort of tokenized element but i just still think it's a kind of similar thing with chat gpt and and ai is making sure to tom's point that the the foundations of what it's getting and where it's pulling from and how it's learning are are incredibly important so so it's funny because because blockchain i think is a really cool concept so i i was uh very into the whole blockchain thing quite quite a while ago and um i think what you've got is your platform has got this amazing opportunity to not just be a marketplace but to be a really trusted advisor as it were to people and you know the the biggest thing is so i get a list of um backdated vintages probably two or three times a week from one particular merchant of which i struggle so hard to not buy something every time i see it coming because oh my god that's amazing a 1983 this or you know and it looks fantastic but then you kind of sit there going but that's that's been around for you know 30 40 years a bottle of wine where's it been you know was it always in bond did it live in someone's garage for two years of its life you know that kind of stuff so without i think some and at that age it's less about great wines it's about great bottles yeah exactly and i think the worry is you know i wouldn't buy uh a wine like that without really understanding its providence and the idea now that you can you know whether you you're using computer vision so ai is really heavy obviously in fraud detection so use computer vision on the labels to check everything about a label high resolution images to print dot gain and all that kind of stuff.

33:21And then you've got QR codes or serial numbers, which can all ultimately be copied and used. And then, you know, NFCs and the capsules and things like that that can then also put it on the blockchain and have AI identify it. I would love there to be a really easy way for that to be like a passport that travels with a wine and that everybody uses them, as it were. because if that was on your platform and I saw a case of Chateau Mouton trading on there, but I can see the whole history of where it's been, suddenly the price is not as sensitive anymore. Do I want to go and spend all that money on a case if I'm not sure that over the last 20 years it's not been in somebody's garage getting hot and cold and hot and cold?

34:10I think it's going to become more and more, we're already seeing it, I think, particularly with some of the dealings we have in the US where obviously the market there is completely different when wines are moved across there they effectively become duty paid and exactly the you know what you said it's not just likely to have sat in someone's garage but may have just sat in the back of a lorry halfway across you know Texas in on a hot day so it that really just does prove to be a massive kind of gap between the general circulation of wines that have been stored very poorly and have very bad provenance to the versus the you know the inbound market that is more you know much more familiar here and across emia and into asia so i do think that that premium that people are going to be willing to pay um more for essentially just well provenance wines and we've seen a lot of that i think now with with the fact that's been i would say arguably our most important aspect of how we operate and making sure that we're incredibly um you know watertight on provenance for all wines that we hold um and that's i think going to be one of the biggest and most powerful selling points to a lot of the stocks that we the biggest challenge that's

35:22Jonathan Stevenson:always existed with the with the blockchain element well actually the two biggest issues and obviously jonathan knows that we explored this quite extensively a few years ago it's one the system of record who owns that is it open like you know is it a shared network a node network where you know everyone's got a responsibility but benefit for updating that single source of record i mean at the moment where we've seen blockchain come in at certain producer levels or even kind of third parties you know it's very bits and pieces you know and you kind of need that full commitment from let's say a producer yeah and everyone to say and then have that kind of value accretion from maintaining it and where does that value accretion go to so that's probably the biggest challenge that hasn't quite been landed on firstly and the second element is even once you get that is the connection between the physical and the digital you know obviously we know that a lot of highly expensive wines especially collectible wines people prefer to have them in their original cases so here's a great example right do you to maintain the physical record with the blockchain record do you open up a banded owc of domain de la romana quantity latash to make sure that you're updating the physical appendage on each bottle of latash or do you maintain it as a banded case and therefore does then the physical appendage that connects to the blockchain layer sit on the case or the bottles how often do you check it what's the lifespan of that bluetooth if you're using bluetooth detection if it's normally long life at the moment is only 10 years a lot of those wines are held for 20 30 years so once you get I know I'm going into the weeds with it, but we've done a lot of research and we probably ended up wasting a lot of time and money on it in terms of trying to investigate how we could do it.

37:01Jonathan Stevenson:But it got to the point where you kind of don't really want to be the first mover in the market. You kind of want someone else to go and do it. And you go, great, the technology and everyone's thought about all of the edge cases of all this scenario. To your point, do we all collectively, and as Jonathan's mentioned, see the benefit of it? Absolutely 100%. but you need a singular leading force to kind of create that ripple effect that then you can just jump on top of because I don't see it as a first mover takes all market. No. If everybody's not involved, the problem is you only need one part of the chain to break down the whole thing was pointless.

37:37Yeah. So, yeah.

37:39Jonathan Stevenson:And then the other issue is, I saw one company in Europe that came along that was very technology focused. I think they store their wines in Luxembourg and again they've gone really heavy on this so like all the bottles have got um the how was it low energy bluetooth yeah tags yeah low low energy bluetooth tags and all the bottles and the cases and they have a sensor in the warehouse updates every i don't know day or whatever it's got temperature humidity whatever but the flip side of their marketplace that they've created is you can buy a cheval blanc 2019 from us from the market from the most trusted most established wine merchants in the uk like berry brothers for example uh let's say two thousand pounds a case or you can buy it for them for three thousand euros so the other issue you have is you're paying a 50 premium that this case has got a low blue tip energy thing that updates every 30 days isolation you've decided that i'm almost certain that i'm happy to not pay the extra 50 and know that it's in the bonded warehouse in the uk where it's pretty safe and certain and that's the other So it's also from a cost perspective, you've got to get it down economically at scale where the costs get to a point where there isn't a differentiation in terms of what you're paying for because consumers aren't going to pay an extra 30, 40 percent for the sake of it.

38:50Yeah, it would need to become that as market bottom. And then the rest would almost become like a slightly wider set of like you're either buying on the blockchain or you're not. And then the not would be kind of like almost more, well, I guess like auction houses, those type of like, you know, duty paid. And yeah, I mean, it's interesting. So is it, you know, picking up that fraudulent stuff, you know, marketplaces, I don't know if you have problems with people trying to put stuff on there that shouldn't be. But how are they being verified at the moment? I mean.

39:20Jonathan Stevenson:Yeah. So, I mean, obviously our marketplace is an inbound marketplace. I mean, you can only offer wines for sale that are already in the facility. So we have a photography studio. We have obviously trained staff there that know, you know, for example. Yeah. I won't give any of the trade secrets away, but certain producers have certain things you can look through. whether it's um uh what's it called uv light or um certain things that you can see under a microscope when it comes to the ink that they use on certain labels but there's ways to identify stuff like you know the high value wines um so that's what we do we pre-verify pre-check everything and i'd say most of the time bonded wines you've got a full track record of where the wines are where it started where it was originally purchased from where it's come in how many different warehouses it's been into so this is the other thing for me when you actually look at the uk bonded warehouse market you kind of cover 90 of the issues that blockchain would provide and i end up thinking to myself as well kind of like blockchain is a really good solution for like a full duty paid market like the us because us doesn't really have this idea of like excise held wines everything is like physical you know even collectors haven't really got their heads around having wines in a third-party warehouse everyone has it in their own homes they've got enough space yeah in america to have their own their own wine sellers it's not an issue you know it's an issue you have in Europe not so much in America so I think to myself like maybe that's the much better market for this product and we know that a lot more of the counterfeit issues that have existed have come from to your point it's like otherwise you're sort of solving for something that already exists with the bonded market right that's the whole by and large that's the paper trail that you need and to Tom's point there's enough places to reference that you know exactly where those wines have been as long as they're in bond but the duty pay market it is that's a really interesting point because obviously if you can then even to the point of show you know showcasing what time it's how long it's been in the country um that was the point i was sort of making with this so working with a a friend of the business really who who now also is a quasi uh collection manager like basically helps you know wealthy uh collectors make good decisions on what to buy for their home sellers and actually almost he's introducing to this um new client of ours the whole kind of reasons for buying you know buying stock from from bond um and i see that as being a big part of the the new market there because because ultimately he's had loads of bad experience of wines that he has bought from the 70s 80s 90s at various points in his lifetime and a lot of the examples of those wines when he's opened them have proven to be really poor um you know really poor examples of those wines because they have probably just been poorly stored circulated too often so i think yeah you're right i think that would be the better place for that type of ledgering that doesn't exist currently um and it's interesting because you know you talk about an advisor because there's lots of good advisors to wealthy people to do their sellers as it were but they're always biases with advisors because advisors either they've worked with wineries or they uh they have their own preferences about wine or they've had a bad experience with a spanish wine that they bought or something like that so so it's quite interesting if you bring bring it back from an ai point of view ais have a lot less of that bias in them because that bias would have to be on massive scale within data.

42:48So it suddenly becomes... That's kind of what I was leaning to as well. It's like, is it better to... Will the market end up being a better place for that because that slight lack of bias... I mean, I think we saw that shift a little bit with when... You know, look at the market now with regard to critics and the much broader selection of critics that you would look at for reference as opposed to how polarised it was when Robert Park was around and people were chasing Parker scores, us included, to try and seek that price movement when they achieved the perfect score.

43:20Jonathan Stevenson:I think it's a great topic for AI. Yeah. You know, because you think of everything in wine, like, you know, the tasting notes, the scoring, the critical opinion, obviously the emergence of stuff like Vivino and Cellar Track over the recent years around community, which I think, to be fair, a lot of the circles that I know in terms of collectors, a lot of people will preference or prioritise community tasting notes. But, Damon, if you thought about the next two, three years, do you see AI ever replacing wine critics? No, not really. And I think there are kind of two elements. Or augmenting them.

43:55Augmenting at 100%, yeah. Even augmenting a sommelier, looking up information and choosing wine for an event or something like that as well. It's really difficult. If you imagine, if I wanted to go through the whole marketplace and identify a ton of wines and bring it down to three wines for an evening, it's a lot of work and a lot of knowledge I need to carry, massive amounts of knowledge to carry. And, you know, I've met some wine people who are phenomenal. You know, the weather on that field on that year, the weather was, oh, it's a little bit wet, but actually in September. And you kind of say, how on earth do you know that?

44:34Yeah, yeah, yeah. And maybe it's not true. Who knows? maybe they're hallucinating as well but um but the reality is it can analyze a lot more stuff a lot quicker to at least narrow it down so it does 80 of the work for you yeah you don't end up drinking you know tons of coffee trying to go through lines of spreadsheets working out what you're gonna