Latest / The Interline Podcast / Mailroom #1
Transcript
- speaker-0: Hey, welcome back to the Interline Podcast. Grace is away at the beach this week, so in place of her quizzing me about the headlines, we're giving over the question askers chair to you, the listeners, for our first mailroom episode. To be clear, we don't actually have a real physical mailroom. We're a remote company. And if we did, I suspect it'd be like the one from Elf, the movie, full of surly drunks rather than anything magical. But what we do have are Email addresses and direct messages. And for the last couple of months, people have been sending in questions that they'd like us to answer to those inboxes. Today, we've picked six of them. Some of them come from a single named individual, while a couple of them are questions we got asked multiple times. So we've rolled them up into one question that I'm kind of condensed and paraphrased. But I figured it'd be a bit of a dull format if I just fielded all six questions myself. And there's also no way that I have the best answers to them all either. So I'm gonna pick up two at the end of the show myself, because I can answer them directly from our own data and by extrapolating from interviews I've done recently. For the other four, I wanted to pair each question with an industry friend who I thought had the best chance of giving a complete answer to the listener's original question, but who'd also be game for me using that question as a bit of a springboard to kind of overextend my host's prerogative and ask some follow-on questions of my own. Those industry friends are, in the order you're going hear from them, Hannah Crump, who's a director of BOF Insights at the Business of Fashion, Jason Malkosh, founder and CEO of Swatchbook, which is now part of Clow Virtual Fashion, Adriana Pereira, co founder and CMO at The Fabricant, and Vlad Mulhelm, who's the CEO of BoxLow. I'm gonna be the audience proxy and I'm gonna read out all the questions. But if we do this format again, and I would like to make it something we do a couple of times a year at least, if people like it, then I'd love it if you wanted to send us voice notes so we don't have to hear quite so much of me. I'm happy to take questions in writing as well, but voice questions are even better. I'd also like us to use this as a bit of a springboard for a request we recently had to do some office hours on a regular basis. We don't have that fully figured out just yet, but it's something we'd like to run for the Interlines community, provided we can bring on some fun and interesting guests every time. For now though, let's get Mailroom episode number one underway. And we're starting with Hannah Crump from the Business of Fashion, who's also joining me on stage in Paris about the same time this episode airs. Hannah's one of the contributors to the BOF State of Fashion series, which I know you all know, so she was my first thought when we picked this question. Hannah Crump, welcome to the Inside Podcast. Thanks for fielding this question. speaker-1: Thanks so much for having me, Ben. speaker-0: No, not at all. When this one came in, I couldn't think of anybody else better positioned to ⁓ to answer it from my roster of industry friends and contacts. Cause I think it's probably fair to say of all the people I know, indirectly or directly, you're the person who has the most kind of like top-level visibility into what's really going on in fashion. So this one felt like it was firmly up your street. And the question that we have had is this. speaker-2: What? speaker-0: It's the single biggest threat facing ⁓ existing mass market fashion brands today. And you can define mass market. Maybe it'd be useful actually to define what we mean by mass market and what that means to you. And you can then I will give you some leeway on this one and say you can pick one or two rather than having to confine it down to one. speaker-1: Yeah, absolutely. So the way I'd firstly define mass market brands is your typical, what we call in the UK, high street players. So those that have huge scale and then they tend to appeal to a wide variety of different customer segments. So if you think about the most popular mass brands globally, I'd refer to brands like H ⁓ and their various H ⁓ group brands like COS. as well as Gap, as well as your more sportswear orientated mass brands like Adidas, Nike. So that's how I describe the landscape to begin with. But what I'd also say is that this particular set of brands, this particular value segment hasn't had an easy time of it recently at all. And if I think about the last couple of years, they faced two major shocks. The first being tariffs and the second being ultra-fast fashion and the competition that those players are bringing. Now if we start with tariffs I guess to begin with they have really impacted your mass market players particularly those that have great exposure to the US market because of two things. One is their limited ability to pass on increased cost to their customers who tend to be a lot more price sensitive than other value segments like luxury. But also they have a lot more exposure typically to manufacturing countries in Southeast Asia. And those were some of the hardest hit when it came to tariffs. That conversation I think has almost settled a little bit more now as brands have adapted to that new environment. But what I'd say is that something that's very much on the minds of mass market brands that we work with is the competition that they're facing from their ultra low cost competitors. So by that, I mean your Shein's, your Tmoo's and they've really been undercutting your typical mass market high street player for a number of years now. And that's on both product development speed, but also quite crucially price. That's really the key thing that is threatening their business model. And that threat is still true, of course, but what I would say is it is not any more accelerating. So we saw a great acceleration of that threat coming from your Sheehan's and your TeeMoo's. But what we're seeing now is that things like the closing of the de minimis loophole in the US and now also in parts of Europe, as well as anti-fast fashion laws, they're actually making it much harder for your ultra fast fashion players to scale in the way that they have been in the past. So even if we think about Shien and their upcoming IPO on the Hong Kong stock market, their valuation is around 140 billion USD now. And that's quite a steep drop from what they were valued at in 2022, which was around, I believe, 100 billion. So that doesn't mean that the threat from these players has disappeared. We are seeing companies like Shien now start to go other routes to try to grow their business. like offering their manufacturing network to other brands. But it does mean that they are still a threat and there's still a challenge these typical traditional mass market brands in their market share, but also their market positioning. I'd say I think with that backdrop, really the biggest challenge today is finding what other leavers' mass market brands can pull to stay relevant when price is no longer their key option. speaker-0: Mm-hmm. And do you because I feel like I've read a few pieces around sort of counterpositioning on the price thing, where you will have brands that I think of as mass market, most of which crossover with your definition as well, kind of pushing upwards to try and maybe escape that bracket a little bit and to say, well, I'm targeting a more premium consumer and I'm targeting a higher price point because presumably there's no point fighting it out in the value bracket against the kind of competition you've just talked about. speaker-1: That's exactly right. And we've been seeing these mass market players over the last few years engage in what we call elevation strategies. So this is really what they see as an escape route of sorts to manage the ultra low cost competition that they're facing, but also the increased costs that they're facing at the same time. What I would say though about elevation. is it's not available to every single brand nor is it appropriate for every single brand. And on top of that, elevation costs a lot of money. So if you think about the kind of investments that brands need to make in order to move their pricing power but also their perception upwards, they're investing in things like runway shows, they're investing in things like fashion photographers that are traditionally associated with luxury brands. They're building premium flagship stores and all of this is of course to differentiate from the fast, cheap experience that you might get online, but it all costs a lot of money. The capex involved in elevating is a lot and many of these mass market brands, tend to be public. They have investors to report to. They need to justify that positioning with either increased sales or increased revenue. And I think we can't ignore the fact that we are operating in a slow growth environment. Our latest projections for how the fashion part of the segment is going to grow this year is still in the low single digits. We're not seeing the kind of growth that we did in previous years. So it's really about capturing market share. There's not so many opportunities to grow your customer base without taking it from other brands like there was in the past. speaker-0: I think that makes sense. And then what sort of levers would be available from the tariff and the kind of upstream and supply chain exposure sides of things? Because I can see there's there are avenues for mass market banks to experiment with their in their positioning and to define who they are and who they sell for. I don't know that there's a whole lot of options, or at least not ones that I readily see when it comes to addressing all of the upstream uncertainty. speaker-1: Yes, so I think there's a couple of different routes. We have spoken at length in our state of fashion research that we do in collaboration with McKinsey and Company about the agility that's required in your supply chain now to adjust to what is now a much more highly volatile environment where you're seeing what was black swan events like wars, conflicts happening much more frequently. And that's really accelerated since the COVID pandemic in reality. But what I would say is that these mass brands tend to have quite traditional supply chains. So it does require quite a lot of adjustment to try to compete with your ultra fast fashion players that have built their businesses from the concept of speed at the outset. Whereas the mass market players that are many decades old, have very kind of bedded in supply chains that have been restricted by the lead times of often 8 to 12 weeks, sometimes a lot longer, that mean they can't get product to market in the same way. Now, there may be some brands that don't have the option of elevating. They don't have the perception to be able to deliver a unique narrative that enables them to charge more for their product and differentiate themselves. These types of players, think the options for them are quite limited. So what do they do? Do they concede market share and accept that they will become smaller businesses? Or is it a kind of, if you can't beat them, join them approach when it comes to... supply chain. as I mentioned earlier, Shein is opening up its manufacturing network for other brands to enable to produce products on a five to seven day turnaround. Will there be brands that will say we can't compete with our own supply chain, we can't create a brand that is sufficiently differentiated, we're going to take advantage of this opportunity. to have a faster product turnaround. Now, I think the brands that have taken up that opportunity so far have been typically your smaller players. You're not seeing your mass, very well-known brands engaging in those partnerships. But will that change over time? I think it remains to be seen. speaker-0: Sure. And I think maybe the other thing that could move there for those kinds of companies would be more in the way of sponsorship, subsidies and support for domestic ⁓ supply chains, both material and manufacturing. But as I'm sure you would probably tell me if I refloated that in your direction, those things take a while to come online. speaker-1: There's been so much talk of near-shoring or on-shoring for many years now, but the data seems to suggest that the speed at which companies make those changes is very limited and the complexity that you face in moving your supply chain is multifaceted. Not only is it a cost question, the cost of producing your product on-shore might be incredibly different from another market they're operating in. But one part of that as well is also the human element. Something we talk about at length when we do our research on this topic is if brands upend their supply chains, move them onshore, move them to entirely different regions, they may be shoring up their ability to have a kind of more long-term approach to their supply chain. But what happens to those manufacturing regions that are known as manufacturing regions, that's how employment works in that region. Sometimes for better or worse when it comes to how employees are treated, but if you're going to be moving your business from one region to another, there are consequences of that, not just for your own P &L, but also the manufacturing country you're moving away from. speaker-0: Perfect. So I think one one final kind of maybe it feels like a positive note, maybe to me, but you might disagree, would be across both price and positioning and then the supply chain side of things. We've we've talked about Sheehan a couple of times here. And you mentioned their upcoming IPO being projected significantly lower than ⁓ than it was historically. The way that I've chosen to read that is that they're not magic. Because I think there's this I think there's this prevailing idea. It used to have swirl around Zara and IndieTech, this prevailing idea that they have something that nobody else has, and it makes them a fundamentally different type of company than a fashion business. Either in Sheehan's case, people think of them as a tech company to some extent. ⁓ and I think what that valuation betrays is that they're not fundamentally a different type of company than other fashion brands are. They're just Have excellence in certain areas and that other companies don't have. And that's the excellence that they're looking to kind of ⁓ rent out in the near future. If I was running a mass market fashion brand, I would take something positive from that and say, Well, at the very least, I'm playing the same game that they are. It's not it's not fundamentally tilted away from me. Am I being too optimistic? speaker-1: don't think you're being too optimistic. I think it's a bit of a reality check for companies like Shein that they are going to be subject to the same forces that other consumer companies are. And namely, when things like consumer appetite to spend shifts, they are going to feel the impact of that. Now, of course, when you're selling at such low prices, you might not feel it in quite the same way, but the volumes surely will drop and the same forces that are impacting your mass market brands are also going to be impacting brands like Shein. They are ultimately a fashion company as much as they would love to be positioned I think for the IPO as a tech company. They are subject to the same forces that other fashion brands are. speaker-0: Perfect. Hannah, we've more than answered that question and I think you've indulged my prerogative to quiz you a bit further on it. Thank you so much for ⁓ joining. ⁓ hope to have you on ⁓ for a full dedicated episode at some point in the future. speaker-1: Thank you, Ben, really enjoyed it. speaker-0: Okay, next up, I'm bringing in Yazin from Swatchbook, who I've known for many, many years, and who's always been the dude I would ask if there was anything I didn't know about how digital materials work. So you can see why he felt like the right fit for this one. Yazen, welcome to the Intime Podcast. Thanks for coming on and answering this question. I've known you for a good few years. And ⁓ when this one came in, I was like, I I know a dude who's gonna know the answers to this and is gonna have some things to say. So so this one comes in from a listener called Roz. And I'm gonna paraphrase her question slightly just to ⁓ put it in the right order and also to inject a little bit of context from the interline side as well. So her question is, is AI, and I think she specifically means generative AI images here, leading people to believe in what she's calling assumed drape rather than real math. So assumed drape of fabrics. Because she feels it's really important that people understand that proper 2D patterns have to be the basis for 3D apparel. And if I can just add a little bit. So we've just recently run ⁓ the survey, AI survey for the year, which is ⁓ captured in the AI report 2026 that the Interline published. ⁓ and people largely agree that image generation, both for initial kind of ⁓ creative and concepting and development use cases and for like downstream marketing, content creation, product detail page images, and so on. Is is the most mature and most ubiquitous kind of use case for AI, which is a good thing if you're selling AI solutions, but maybe less of a good thing if what we're trying to do is to put accurate, pattern accurate, fabric accurate representations of garments in front of people. I suspect you might have some thoughts about this, so hit us with them. So speaker-2: What she's alluding to is correct. ⁓ I think if you are the matter expert, ⁓ you probably have a lot more awareness in terms of the AI not able to give you ⁓ the re the real accurate results that people are looking for. But the majority of people using AI are using it for things they're not experts in per se. And so for them it looks real and it looks like this will do. And I think that's you know, that's the question is Whether this is real versus this'll do for now, ⁓ there are two different ⁓ objectives and I think people kind of mix them up, and I think that's the biggest challenge of using generative AI. There's certainly a lot of use cases for it that are, you know, ⁓ of use, but the reality of it is most of it is for visual product generation. Most of the manufacturing data to make physical stuff. ⁓ is not a part of that and isn't taken into consideration. So a lot of times what it's showing, we don't know if it's producible or not producible. There's two two primary things to kind of make sure I cover is one, ⁓ it kind of goes off the whole idea of the the matter expert. I think if you are a non-writer and you use generative AI for LMs and things like that to create stuff, it'll seem like magic. But if you're a writer, you'll start seeing the flaws ⁓ pretty quickly. And it's the same thing for design. It's the same thing for art. And you'll start even seeing where it's coming and picking things from. Which again, for a designer, it's like, ⁓ I n I I know what it's trying to I mean, at some point you can even you can even actually allude to what I think it's thinking and what ⁓ mashup she ⁓ it's trying to make. But you will quickly ⁓ see through those, whereas a normal user who doesn't do that for a living, doesn't have the experience and the years, ⁓ we'll just it'll just seem like magic. Like, ⁓ my gosh, look, it solved all my problems so quickly. And I think that's the fundamental reality we're in right now. Maybe it gets better, maybe it not. speaker-0: Okay, and that's interesting to me because the point about subject matter expertise is one I come back to ⁓ a whole bunch. Now I'm I'm no photographer, so I can take a look at generated images and unless there's any really egregious mistakes, which is kind of less and less the case with ⁓ the frontier image generation models these days, ⁓ I'm I'm hard pressed to pick a lot of those apart ⁓ from a pure kind of aesthetic judgment point of view. If you bring AI specifically LLMs into the area that I do consider myself to be more of a domain expert, which is writing, I can quickly see the faults. ⁓ and and what looks like a f complete and presentable essay to from the outside doesn't hold up to much scrutiny. ⁓ are we talking about fundamentally the same thing here? speaker-2: It is exactly exactly that. Yeah. So again, if you take that into this same example, if you're not a designer and you tell your AI agent to cla create a design or help refine their design, it will seem magical. It will seem like, My gosh, this thing can shave and I th ⁓ minutes and hours of my time. And I think that's Where we get a lot in the industry where there are executives and people in management positions that are looking at these AI tools and saying, you know, ⁓ maybe let's assume that their background is not design, but they use this and they're like, well, why do I need designers at this point? It's easy to assume it because you're not the designer. You're not solving design problems. And design is an iteration, ⁓ you know, unlike other things, and I won't say, you know, it it's not that you don't do that in writing. You definitely do, but you're probably solving, you know, storyline. ⁓ challenges. Design does the same thing. It's it's trying to to solve, you know, for you know, functional things, aesthetic things, form ⁓ challenges. And so if you're not a designer, you're not looking for those. You're just looking for the aesthetic. Does this look good or does this not look good? Which again is a thing that you need to pass, but there's form, there's fit, there's function, there's other elements as a designer, ⁓ especially in the fashion space that you have to take into consideration. And I think that's when you start Poking those holes as a designer into the AI and like, ⁓ this doesn't work, or how will this cut? Or the pattern would never function in this manner. And the draping is the same thing that Roz had brought up to kind of circle back, is that we don't have enough draping ⁓ examples of of all the materials digitized in 3D to kind of match everything that the, you know, that the AI has access to to do that. And I think. Whereas, you know, simulating it in 3D first to showcase the the broad strokes, I'll say. ⁓ that's not to say that there aren't use cases for AI to add that final level of detail that would that would be possible using a 3D artist level, you know, things that a person that would work in entertainment on movies and blockbusters that would add all those details. The unfortunate part we don't have that in terms of capabilities and ⁓ manpower at the brand level to do that for every single garment that you're pumping out. So you could use AI to do that final level of polish, but it wouldn't be for the final draping. It would be to kind of add the a bit of fuzz, a bit of just like the photography stuff. You add a bit of value, a bit of ⁓ tone mapping, ⁓ just to finesse it. puffer jackets are a great example of something like that where it You know, it looks good, but that those little tiny wrinkles, you just want them to add that level of realism, taking you from like a ninety percent to a hundred percent or a ninety nine percent. That is to me a more realistic use case ⁓ to get something real out of it. But then the patterns and the draping still require a good amount of simulation. speaker-0: Perfect. And you've you've said something that's just given me an opportunity to ask my own question here, which again, like host host prerogative. I got in an argument with somebody on LinkedIn, which is the worst kind of argument you can have. ⁓ and it was it was based on a bunch of generated images that they put out there, ⁓ where they were saying, you know, I've been really pushing the frontier of material simulation with this. And the argument was, well, there's no, there's no material simulation happening here. What you have is a an image model trained on the output of a lot of either real materials or material simulations. And those are not the same thing. Now, what you said is there is that there's we don't currently have sufficient training data to make that kind of visual approximation fully accurate. Is that something we could have? Do you think if there's enough training on Accurate drape that we get there. speaker-2: I definitely think so. The challenge is that there probably is enough data. The problem is that data is not clean. We don't we don't have the ability to confirm ⁓ that data. Let's take an example. If you showcased a draped ⁓ fabric, I don't know if that was draped, you know, consistently in a manner that is explained And tracked and when you put the next piece of fabric, you're draping it in the exact same way. So I have a way to combine and also to compare. ⁓ because I know a lot of times what happens is that you're most of the shots that you see online, it's to sell products. So you're just trying to get an aesthetic view of it. So people are moving things, just like you do in a Photoshop, right? You're pinching things, there's clamps behind it that you don't see. There's a lot that goes into a photo shoot of Products that don't necessarily represent the real draping because they're trying to make it look as good as possible. They're not trying to make it drape as realistically as possible, right? So unless you are doing some sort of consistent ⁓ style draping of fabric, and then obviously fabric is just one aspect of it, right? The fabric direction is another aspect, ⁓ warp, weft, diagonal, it drapes a bit differently how you orient it. Y you almost can't get the exact same drape on ⁓ like even on a cylinder twice. ⁓ just because the fabric is so nuanced. And so when you're cutting it and you're putting patterns in it, it whether if you you know, if you cut a triangular piece, but it's oriented differently depending on the warp and the weft, it will drape differently. The AI doesn't have that information. It has mass information that you can feed it, but it doesn't have that clean data tagged and labeled for it to understand how that draping would speaker-0: occur. Perfect. Yasin, ⁓ I think we've done a pretty comprehensive answer there. Roz, ⁓ hope we tackled your question well. Sorry for piggybacking in it with ⁓ with a couple of my own. ⁓ Yasin, thanks for your time. Thanks for coming in and field in this one. speaker-2: I appreciate so much. speaker-0: The next question is going to Adriana from The Fabricant. Adriana's one of those people who's been part of the debate around digital fashion and the digital transformation of fashion for so long that I can't remember a time when I didn't know who she was. Recently, though, she's been putting a lot of work into helping new brands start with the advantages that modern solutions, especially AI, bring with them. So she was a slam dunk to field this next question. Okay, Adriana, welcome to the entire podcast. speaker-3: Thank you very much, Ben, for having me here. speaker-0: Not at all. It's been a while, I think, since you and I caught up. And I've been meaning to bring somebody from the Fabricant onto the show. I know we we maybe have an opportunity there to do this something a bit deeper on the big interview show sometime before the end of the year. For today though, ⁓ I wanted to put a question that we've had from a few listeners actually. So I'm kind of synthesizing a few different perspectives into this. ⁓ and I wanted to put it to you because I know the work that you've been doing at the Fabricant recently and you Personally, you're quite passionate about the ability for technology to enable new entrants to break into the fashion industry and to build different business models from scratch. So the couple of questions we've received, I can wrap them up like this, which is people are effectively saying, if you were to start a new brand today and you didn't have any of the legacy of ⁓ systems and setup and technology and everything else that comes with it, how different would your operating model look from the historical one and what would modern digital tools and things allow you to do differently to the way that companies do it who've been set up for a long time? speaker-3: ⁓ I love this question. ⁓ it's ⁓ very close to our heart and the reason why we actually start the fabricant to begin with, which is create ⁓ a fashion industry that's more democratic but also more creative and ultimately more profitable and sustainable. And when we start in 2018, we're using 3D technology to try to achieve that. And it proved to be ⁓ not the right tool really, because it it's quite expensive, takes too long to market, and it's very technical. So the accessibility was very limited. Fast forward, three years ago, we saw the the change that ⁓ artificial intelligence were bringing to especially the creative industry, and we fully adopted we left 3D to embrace AI. And the AI in its early days ⁓ has been ⁓ very well used and it's I think the adoption has been incredible, honestly, in the last twelve months, for the creative part, for the concept development, for the presentations, right, for the old models, look books and so on. So I would say the design and marketing. And ⁓ what was missing for that entire pipeline to digitalize was the connection to manufacturing. And that is something that we actually have been working extensively in the last period together with experts in pattern making to develop that piece. So right now what what happens is ⁓ a ⁓ process that used to take months and several disciplines involved and quite a lot of capital To be brought to life is being condensed in a much short period of time with much I would say less disciplines involved and with a strong connection with the manufacturing. So to make a long story short, to launch a label, if you were off school last year even, to launch a label, if you put it all together, you probably needed around 150,000 euros in capital. to put behind marketing, content creation, development, the expertise you need, plus ⁓ buying your initial stock. Right? In the current model where all the disciplines are ⁓ pretty much replaced by artificial intelligence and where you can connect to manufacturing on micro runs and also sell before you produce, basically produce only on demand, the only cost that you need really is for your marketing and for building your community. So we when we estimate it from 150,000, you can probably launch your label with 15,000. And this 15,000 will be primarily for the brand building and for the awareness of your collection. Right? So for an if I would start my brand right now, I would start at fully AI. I would start with an end-to-end pipeline. And again, with the tools that we're bringing to market right now, that last piece, which is the pattern making, the technical information, is also being able to ⁓ we're also digitalizing it and ⁓ reducing time to market, reducing needs of expertise and create an on demand ⁓ model. speaker-0: Okay. So if you were to start a brand straight out university now, you said it would be AI native. Who would you hire? Like when would be your trigger point that you would say, okay, I'm a solo brand builder. I now need somebody else. I'd be I'd be interested to know who the most likely second hire would be. Would it be somebody would it be a pattern maker? Would it be a CEO? Would it be a marketer? Like I I'm I think it's a pretty open space and I'm keen to see what you speaker-3: Yeah, yeah, very good question. I guess it starts with your own expertise, right? ⁓ I think as as a fashion person from school, you might actually be more than a designer and red cover areas, for example. You you already build a very strong community of people that love your handwriting. ⁓ so if you already have a community build, I think the part that I would be that I would think you would be lacking the most would be the sourcing. And the pricing and more business model, right? So I would be looking for somebody with a more buyer or sourcing expertise. If you are somebody that has a very strong design handwriting, that you understand product constructory well, ⁓ you might be able to actually cover the gap, start your production with a local manufacturing, so you have that covered, then I would actually say you need a marketer. ⁓ and and what I see is that mostly ⁓ People graduating from school, that's what they need, is a marketer, is a brand builder, is somebody that's gonna really ⁓ put your brand in front of the consumers and be able to generate sales. speaker-0: Okay, perfect. That's that's a really good answer to that question. So my next follow-on would be which parts of that equation? So let's just assume you're somebody who has a a typical set of skills leaving school, which parts of that equation are where AI is the strongest now? To where you would say, okay, I think AI is mature here. You don't need to think too much about it. Just just go for it. AI is there. And conversely, Which do you think would be the least mature? The area where you think people will look at the current frontier of AI and go, Yeah, maybe, but I feel like I need to do that myself or hire somebody in. speaker-3: Yeah. I think it's very strong on the concept of development and visualizations. It's very mature there. You know, and typically is also the strength of a designer, really that more creative director role. ⁓ I think the areas where it is lacking and that's why we we're putting so much research and development is ⁓ is the pattern making, the accuracy on pattern making. Which again depends on how you build your own supply chain. If you if you have a dedicated small manufacturer, they can cover that for you. based on the proper information of the of the garments, right? ⁓ but definitely the more translation into manufacturing is the area where AI is lacking the most. speaker-0: Okay, perfect. And that aligns up with the survey data that we saw from ⁓ from this year's AI report as well. Then just personally and anecdotally, I think the area that I would maybe struggle the most would be in the business side of things here. So just and I mean that from a pure like costing modeling, margins, markdowns, like all of the levers that big established companies already know how to pull to get collections sold through at full price. ideally or with some planned markdowns and so on. Because if you come out of university, yes, okay, you've got a a lot of technology in front of you, you've got a mix of skills and things in front of you. I feel like that business expertise is what's missing. And I'm keen to see what you think would be the the best approach to somebody who would go, you know what, I know fashion. I'm a good designer, I'm a good pattern maker, I'm a good sourcing professional or whatever. I don't know if I'm a good business person. What would you say they should do about that? speaker-3: One is yeah, it's true. But what I tell all the graduates is start with a on demand ⁓ model. So basically just produce what you sell. It already reduced the complexity of ⁓ markdown, stock management, right? And we have quite a few examples in our network of users, ⁓ of people that start the model just like that. You actually create the entire product, you create even the patent, but you just produce what you sell. ⁓ that reduced the complexity tremendously. And it's more profitable and more sustainable. And as a starting point I think it's very valid. speaker-0: And presumably you produce what you sell. You do the selling through AI visualization is is what I'm guessing. I think you're right on a bunch of counts here. I think ⁓ I think we're gonna see this Frontier advance ⁓ in the pretty near future. I know we're talking to somebody else from the fabricant a little later in the year, but for now, Adriana, thanks for fielding this question. Thanks for fielding my follow-up ones. ⁓ it's been fun having you on. speaker-3: Correct. Yeah. Right. Thank you, Ben. To the next one. speaker-0: Now I'm turning to Vlad Mulhelm, who is the CEO of Boxalo, who you'll find included in our AI report twenty twenty six. Vlad lives not too far from me, weirdly, around Manchester, and on top of it being fun for me to pull in a tech executive with a similar postcode, Vlad's also just a straight up expert in a very particular area of computer graphics that made him eminently qualified to field this question. speaker-2: Okay, Vlad Mulhelm. speaker-0: Welcome to the Endline Podcast. Thanks for coming on and volunteering to answer this question. speaker-4: Hi Ben, thanks for having me. speaker-0: Not all. I'm looking forward to this. I've been thinking a lot about Gaussian splatting myself. So spoiler, that's the question that we have from ⁓ from the listeners here. And I'm rolling up a few different angles into one here. So we had some people asking about the use of Gaussian splats to t replace or augment or run alongside the traditional material capture workflow. So you're kind of take a material swatch, do a flatbed scan. bring that into a 3D design tool and so on. So goes in splating for the fabric and material capture. But then also, I know people are interested in the same technology and the role that it can potentially play in making it quicker, faster, more straightforward, more streamlined to capture full finished products and then to get those visualizations in real time in front of people that they can interact with in their browser. ⁓ when all of these questions came in, I couldn't think of anybody else that in my network who would have a better, more expert and more grounded point of view in ⁓ in all of this. So to give bring us up to speed if you can, on those two applications of Gaussian splatting. One in fabric and material capture and one in finished product capture and then real time visualization afterwards. speaker-4: Yeah, so I think to make a ⁓ a good jump into it, ⁓ I'm just gonna like lay out the two technologies ⁓ side by side. The traditional way of capturing surfaces or materials ⁓ is usually by breaking down the texture ⁓ into its ⁓ sort of albedo component and then capturing all of the texture surface height ⁓ using some ⁓ image information, ⁓ and that requires some specialist equipment, ⁓ maybe a bit of AI. ⁓ and it's usually then used further down in the ⁓ pipeline to apply to a 3D model and then relight in a 3D rendering ⁓ engine. ⁓ so that will not go away. ⁓ Gaussian splits have a slightly different strength and slightly different purpose. So go the way Gaussian splits work actually is more suited to soft ⁓ fabrics ⁓ and materials that have rich textual information, ⁓ but it's not ⁓ then that easily transferable to the traditional mesh pipeline, so it cannot be applied that easily to an existing claw model ⁓ or a garment that's been designed ⁓ for the traditional design process in 3D. ⁓ having said that, Golden Splats are basically ⁓ millions of blobs ⁓ in 3D space ⁓ that kind of change their appearance ⁓ as the viewer changes angle. ⁓ this is their strength, that's why they can capture ⁓ things like fabrics, soft fabrics, ⁓ down to really the fiber detail. ⁓ but at the same time they also capture the baked lighting, the lighting that was applied to the garment at the time of ⁓ of capture or shooting. ⁓ and this is the difference, ⁓ whereas all of the kind of fabric scanners in existence, ⁓ they tend to break that down, decompose it, so that then the ⁓ the technical user can then reuse them in a 3D rendering engine and relight them. So for specifically for flatbed or sample capture, Gaussian splitting might not be the right choice. Having said that, alongside material capture, what Can be actually a really strong use for Gaussian Splatz is to capture the the sample maybe as it drapes in 3D shape alongside the flatbed or the the flat sample capture so that there's a sort of a ground truth comparison between the two because the strength of Gaussian Splat, as we said, is it captures the environment lighting ⁓ but in a view dependent way. So as you move, for example, ⁓ around a a satin, it would change appearance ⁓ in the same way as it was captured. ⁓ so if it if you lit the sample in a specific way, you would capture those highlights moving along the surface, which is very challenging in ⁓ traditional 3D and CGI workflows to kind of visualize after the the the the capture. So There's strengths and there's ⁓ obviously ⁓ things that Gaussian splatts are not that good at when it comes to purely ⁓ material capture. speaker-0: And just walk me through what the capture process would look like and what the hardware requirement would be. Because my it's my understanding that at least as you practice it at Boxalo, k taking Gaussian splats and then having those intelligently assembled into ⁓ a a model, if you want to call it that afterwards, is something that's more easily achievable on consumer hardware and in a range of different settings as opposed to needing specialized hardware if you want to do high fidelity capture of materials in the traditional flatbed way. speaker-4: So the capture process is the strength ⁓ of Gaussian Splats. Actually it's a lot easier and a lot ⁓ less constrained than traditional photogrammetry or traditional capture where you require maybe polarization or ⁓ various tricks to even out the lighting. Whereas with Gaussian splats you can just use commodity video or photography equipment and you Take multiple frames from different angles, and then they get processed using an algorithm to reconstruct the Gaussian splits, so millions of ⁓ sort of ellipsoid blobs with ⁓ color information that changes with the angle. So actually it's a lot easier for the non-technical user without any specialist equipment to capture really high fidelity ⁓ detail ⁓ with Gaussian splits, and that's why we love it so much. speaker-0: Fine. So I if I was putting myself in the shoes of a fashion brand who is working with a mill or material supplier, if I wanted to do highly technical work with ⁓ a fully representative traditional kind of flatbed scan and something that was put through a fabric testing kit and everything else, included all of the visual attributes and the performance attributes and characteristics and stuff, that traditional workflow continues to make sense. If what I wanted to do was to have the the easiest on-ramp to being able to share the way a material looks between ⁓ a mill and their brand customer and you wanted to have low overheads in terms of time and effort and hardware investment and so on. It sounds like that's maybe an area where Gaussian splits would have an edge as well. speaker-4: Yeah. ⁓ I would imagine you would do your yeah, more scientific ⁓ capture with the equipment. You would then probably take a a larger sample, you drape that in different ways. You would then use even something like a a good phone with a high resolution camera, take a video from ⁓ various angles, usually three sixty walking around the the draped garment. And that will produce a Gaussian splat ⁓ in that environment. with some investment in, you know, cheap photography lights or or slightly more colour accurate daylight lights you can you can capture alongside ⁓ a really, really good, high detailed ⁓ sample. speaker-0: Perfect. My my other follow-on question would be about Gorsies Plats when it comes to capturing finished product. So if we think about the the typical use case I could think of would be a garment on a mannequin in a studio. and the goal is then to be able to put that 3D representation in front of someone else who is making a choice. So that someone else could be An internal audience, it could be if you're ⁓ you you're doing sample reviews or so on, it could be an external audience if you have a fully finished product and you want to get it in front of consumers. Now, my understanding is, well, two things. The capture process is going to be consistent there between the fabric and the finished garment. I'm not imagining you suddenly need to approach that differently, which is again very different from traditional photogrammetry. Then the visualization and real-time rendering. Part of it is where I think it gets particularly interesting as well. Because I've if you want to take a traditional 3D mesh and texture-based representation of a garment and you want to render it out and have it exist in different levels of detail for use in different scenarios, you're doing like mesh decimation and texture compression and assembling different render profiles and things for all each of these different scenarios. I've interacted with some Gaudi Gaussian splatting kind of big like environments and things like that as well and as well as like smaller product visualizations. They seem pretty lightweight. They seem much easier to run in real time. Explain to me why that is. speaker-4: Well, first of all, everything is baked in. So all the data or the lighting, ⁓ the color data is baked into the ⁓ sort of the point cloud ⁓ and those ellipsoid primitives just represent the detail ⁓ by overlapping with each other and as you spin in the web ⁓ rendering engine, they kinda change with the viewing angle. So they are they are much much lighter to represent because there's no re-rendering and re-lighting happening. ⁓ and when it comes to real time, it really shines. ⁓ because to do something ⁓ that looks really good in real time, let's say in a web viewer, you you actually are constrained by the user's hardware. Therefore you've have to decimate the mesh, reduce the number of polygons, you have to use ⁓ a flat texture and you're limited by resolution. Download speed then becomes an issue. Whereas with Gaussian Splat, you ju you're just ⁓ you know, streaming one file with ⁓ points that have all of the data attached to them. ⁓ and then commodity hardware, ⁓ consumer hardware is particularly good at rendering those in real time, with a simple web viewer ⁓ supported by any browser. But because the lighting doesn't need to be redone from scratch by the rendering engine. It's it's already there and it's dynamic. It changes with viewing angle. The the realism's much much higher ⁓ and then the fidelity ⁓ and the texture is much much ⁓ more readable and also because it's a fuzzy sort of structure, not a f not a solid ⁓ mesh. You can represent things like fibre, hair, fur, which was previously impossible in real time to be honest. You needed an offline rendering engine. to represent these things and you needed to simulate those. So the the the time between capture and full photorealistic rendering in real time is compressed to hours rather than days. speaker-0: And I think the final my final thing would be so if I extrapolate from what you've just said there, a lot of when people talk about like extending the value of 3D assets. So if I put myself in the shoes of a brand who has natively authored a bunch of products in 3D, I'll simplify it that way. ⁓ so they have pre-existing clo, browse wear, opti-text, whatever whatever those are and and they want to then get those in front of consumers, but ⁓ simplify it down to consumers. The the two ways to do that are either to compromise on the fidelity of the presentation if you're doing it in real time, or to render it offline as static images, ⁓ to do essentially like virtu virtual photography. Gaussian Splat seems to split the difference there and it seems to get you the the real-time value, but it also gets you the fidelity, which I could see being particularly important to the luxury industry. I can see it being particularly important to products where that Texture and feel and things are perhaps ⁓ more important. speaker-4: Yeah, I mean it's difficult to convey with audio the the difference between ⁓ the photorealism of a rendered item versus the captured Gaussian splat item. ⁓ and I've I've been in this ⁓ you know the CG and 3D industry for twenty years ⁓ or so. And I've ⁓ I'm blown away by the the level of quality Gaussian splatts achieve. But there is a hybrid approach as well. Potentially those brands that have assets that they have designed and generated using traditional 3D ⁓ and CAD methods, but then use offline rendering to represent a still images or video, they could actually generate Gaussian splats from that virtual item with the benefit that they could render those offline with all of the kind of rendering engine strengths ⁓ that they might get from offline rendering. The sheen, the fabric, qualities, ⁓ anything that kind of needs simulation. So for example, any fur or hair that needs to be added, that detail usually cannot be represented in real time by converting those meshes to let's say GLB, but you could render them offline and convert to Gaussian splats in the same way as you would capture them. You'd you just generate a virtual ⁓ video or virtual sequence of frames and process those. into Gaussian spots and then you get the best of both. So there is a hybrid in between method as well that could be used to visualize the garments in real time with the quality of offline rendering, but actually present them on the web for more direct consumption ⁓ interactively. So speaker-0: Perfect. Final one. ⁓ so there'll be a lot of folks listening to this who are in the the three D slash digital product creation community. That's a big ⁓ that's a big audience of ours because we've done the digital product creation report for a good few years now. If for those folks who are listening to this who are deep in the weeds on the established way of of doing everything, like these three D workflows that we've talked about, if they're interested in Gaussian splits, what would you suggest as a starting point for them to get in and start exploring. speaker-4: I think the easiest way would be to ⁓ you know, do a little bit of homework, a little bit of reading. ⁓ if they want to produce some Gaussian splats, there are a few offline ⁓ applications that they could download. ⁓ the key ones out there ⁓ post shot. ⁓ is a is a fairly well known one that they could run locally ⁓ on their machines. ⁓ there is also Lichtfield Studio. ⁓ I would ⁓ try and follow ⁓ a few of the creators in the space. ⁓ Radiance Fields are ones that post regularly about updates in the in in in that kind of field. obviously it's a manual sort of technical approach, but for that audience, this experimentation might be quite helpful for them to understand how this works. ⁓ they could they could capture with ⁓ DSLR or with their phones, ⁓ and then they need to do a bit of pre-processing. one of the one of the good candidates for this is ⁓ reality capture from Epic Games. It's a free piece of software, traditional photogrammetry, that's kind of the first step in goes in Splats. And then they can use one of these ⁓ offline bits of software, ⁓ if they have a a a decent bit of hardware to run ⁓ run that on and ⁓ just experiment. ⁓ there's obviously online platforms like ours. ⁓ there's there's ⁓ the ⁓ the ⁓ pioneers in the space, Luma, ⁓ Polycam also offer ⁓ Gaussian splits, even though the fidelity's ⁓ potentially a little bit ⁓ lower as that's not their specialism. Kiri Engine support Gaussian Splats. ⁓ These are just a few options out there for people to start experimenting. When it comes to using those files, a lot of the rendering engines are starting to support Gaussian Splats. Not as flexible as meshes, but they could be used alongside meshes or in environments. ⁓ V-Ray rendering engine, ⁓ octane render, ⁓ I believe. These support Gaussian Splats. ⁓ to some extent, also they could relight them. ⁓ obviously, if you want to go really technical, Houdini has ⁓ some implementations of Gaussian Splats, some Real Engine has some plugins ⁓ where Gaussian Splats can be used. These are just a few options, but the ecosystem's growing. And obviously, as we progress, imagine this is like day zero of this new technology. ⁓ it evolves, ⁓ it will become a lot more flexible and more editable, and also. ⁓ more integrated with new AI ⁓ generative workflows ⁓ as it kind of lends itself to that that sort of work. ⁓ but yeah exciting times ahead for for 3D ⁓ content creators and ⁓ designers. I don't think Gaussian Splats will ever replace meshes. I think they have a a place alongside ⁓ design workflows because you need to capture They're good for visualizing things, ⁓ but they're not that strong for editing and and creating things. And this is where the distinction lies, really. speaker-0: Perfect. Well, that's the great encapsulation of the answer to the question. Vlad, thanks so much for joining and fielding this one and fielding my follow-up question stuff as well. That's it for the industry friends for this first edition of the mailroom. But there's still two questions to get through, and I'm going to tackle them myself for reasons that I hope will quickly become apparent. The first one is this, which we got asked at least five times by five different people. So I'm combining and paraphrasing here. But the question is. In which teams and or functions are fashion brands seeing the fastest AI adoption and what's causing it? I can field this one from our first party data, which we captured pretty recently as part of the AI report 2026. If you're interested in going deeper on anything I mentioned here, you can download that report or you can read the essay I wrote called Arms Length AI, which is intended to be a more digestible way of understanding the story of the survey results rather than looking at the dry analysis. And that essay is also accompanied by our first toolkit for AI agents, which gives you a way to turn the essay and the survey data behind it into a live working asset to use with your AI agent of choice. So if you think what I'm about to say is interesting and you want to go a little bit deeper on it, you wanna have a multi-turn conversation about it that's grounded in your memory, you can do that. And if it sounds complicated, it's not. You can visit the essay page. And you just copy a one-shot prompt and you paste it into ChatGPT, Claude, Gemini, Hermes Agent, OpenClaw, whatever you use. We'll put the links in the notes accompanying this episode to both the essay and to the toolkit. As for the answer, the survey data shows us where fashion professionals perceive AI to be the most mature across the product journey. That doesn't directly map to actual proven maturity, but it's a good proxy. And the top two areas are. Trend, market, and competitive analysis, and marketing and content creation. Or, to put it a different way, the upfront analytical stage of the product journey and the visual creation part that happens once a product is ready to advertise. Neither of these should be a surprise, really, because they line up pretty much exactly with the things that people generally consider AI to be good at. Synthesizing insights from high volume, high variety, unstructured data. which is exactly what the fashion market's made up of, and generating pictures and videos that look increasingly like real photos and film. Now I don't want to sound dismissive about any of that, but there's no real secret to it. Every industry is doing everything they can to figure out where they can extract quantifiable value from AI. And I don't think it's a coincidence that companies keep finding those answers in the places that LLMs and generative image and video models are already proven and well suited to. Especially since those places both have a heavily influence on product outcomes and market success as well. If you know with the backing of novel insights what the market wants, and you can massively scale the way you tell visual stories across channels, e-commerce, social, traditional marketing, and so on, about that assortment, you're gonna see some return. As you might have guessed from the way I've been phrasing this and approaching it though, there's also a big gap in between those places. And that's where Product development, technical design, sourcing, and production happen. You know, the things that turn data into actual products that can then become the subject of image and video, either traditional or generated. Those are also the places that the industry tells us that they believe AI to be the least mature. And although this wasn't directly part of the question we received, That perception of what AI is less good at contributes to what I've ended up calling the non-virtuous loop of AI adoption in fashion right now. Let me explain that look quickly before we finish with this question. despite roughly nine in ten fashion professionals telling us they use AI every day at home and at work, only about a quarter of them trust it enough to base important decisions on. That lack of trust shows up in a few different ways. Primary one is that roughly three in five people fear that their companies will end up making the wrong choices, or at least suboptimal choices, because they relied on inaccurate, incomplete, or straight-up hallucinated information from AI. Now, I've spent enough time talking to tech executives and brands and playing around with frontier models myself to know that this year was a bit of a tipping point there. If you want an LLM to be able to return accurate answers that are grounded in your product data, your brand DNA, and so on, you can just give it the right tools and you'll very likely get the right responses. But because of that lack of trust, around two thirds of the people who took part in our survey, I think, told us that their AI initiatives were still either fully sandboxed, i.e., they live completely independently of the rest of the technology estate, or they were very minimally connected to other enterprise systems. Now you can see the loop here, right? AI isn't going to give you grounded, reliable responses of the kind you want if you don't connect it to your sources of truth. But if you don't trust it, you're not in a rush to embed it into your tech estate and connect it to those things. Personally, I think this is the kind of issue that gets resolved naturally, where the baseline just kind of rises over time for everyone. And then people wind up Trusting things as a bit of a domino effect. They see a brand they admire connecting their PLM to Claude through MCP, and then they go on and do the same, and the people who follow them do the same afterwards. I don't think that's guaranteed by any means, but I do suspect that we're going to see that perception of maturity shift more towards the middle than just the ends of the product journey through a kind of generalized swing towards acceptance. But also probably because the technology vendors will do a better job of explaining. The value of AI in the middle of the product journey. It's definitely easier to sell an image generation workspace or an AI native analytics platform to merchandisers and planning teams than it is to convince core Garmin engineers, product developers that AI belongs in the heart of product development. But I do think we'll get there. Then our final question is a quick one from Yvonne, and it's also about AI. So she asks: well, tells really this. We don't talk about alternatives to the closed source AI models nearly enough, which well, isn't a question for one thing, but it isn't incorrect either. I'm sitting down to record this show just after we published an edition of our Friday analysis all about some of the big news in OpenWaights AI and mean for fashion. So I'd encourage Yvonne and everyone else to go read that first. I can summarize some of it here for you though, and maybe give you some extra context as well from from my side and from what we see in the industry. First I want to point out that there's a difference between source and weights. I'm not going to get too deep into this because this episode would last forever, and I'm also not a subject matter expert. But basically every LLM is closed source because being the opposite, open source, would mean giving the world full access to everything they would need to rebuild the model from scratch, from the training data upwards and downwards. Both Claude and Deep Seek are closed source. The distinction we actually want to talk about is between open weights and closed weights. Weights, in the simplest sense, are numerical representations of information that lives inside a model and how it negatively, positively, or neutrally influences other information. Weights roughly correspond to what models quote unquote know and how they act. And training a model is in a really simplified way a process of adjusting those numbers up and down. Weights also correspond to a model's size as measured in parameters. It's not a one-to-one mapping necessarily, but it's roughly there. A small model that you can run on your phone might have a billion of those parameters, whereas something huge like Fable 5 is believed to have more than five trillion, although nobody really knows for sure. The models behind ChatGPT and Claude are closed weights, in that they're a finished and immutable product. You can't take GPT 5.6, Sol or Terra or Luna, and adjust the numerical values through fine-tuning yourself. You can't slice layers out of them, you can't run those models on your own hardware, and you can't, at least not officially, use them to train other models through distillation. OpenAI did release an open weights model series, GPT OSS, which actually gets a fair bit of use behind the scenes, but that was the exception, not the rule. There are some other American AI companies that have released OpenWaights models as well. NVIDIA does it. And one of the original C-suite members of the OpenAI team, Mira Morati, started a new company called Thinking Machines earlier this year. And they recently released an open weights model called Inkling. Google also does it fairly often with their Gemma series. But the big AI companies you know, and the ones that are embedded into your organization, OpenAI and Anthropic, largely release closed weights models that are defined by the fact that they live in the cloud, they exist. within either their applications or through APIs that they govern the pricing of, you can't change them and they don't belong to you. Now, most, but not all, models developed by Frontier Labs in China, by contrast, are open weights. DeepSeek, Quen, Kimi, ZAI, and other companies have all released really capable models that are, especially in cases like GLM 5.3 Flash, which is top of mind at the moment. Also incredibly cheap to run, as well as being yours to run locally. If you've got the hardware for it, you can fine-tune them and so on. There is some complexity in the licenses that come with OpenWaights models from China. Some of them have clauses in them that restrict commercial usage or contain separate agreements that mean you have to pay the model creator if you if you generate revenue above a certain threshold with the model. It's kind of similar to the licenses you see for Unreal Engine. Now. As the final piece, that model I just mentioned, GLM 5.3 Flash, was also the first highly capable LLM that ran on Chinese chips and infrastructure for its initial stealth launch. It's now available through a lot of different avenues and inference providers, but this was all a bit of a milestone in the sense that not only are Anthropic and OpenAI committed to selling you closed weights models, they also dictate where those models run. And the answer is they run in America, on American hardware. The EU, as you probably know, is currently working very hard to reduce its reliance on American technology providers, both software and hardware. So a lot of state-level sponsorship is flowing in a direction that's very favorable to models developed outside of America and especially towards open weights models. And brands that care about governance and data sovereignty are doing basically the same thing. So you can see that. Despite Yvonne's question sounding a bit dry and technical, it's actually a complicated geopolitical and strategic one. If you're listening to this in the US or the UK or Europe, the odds are that almost every text based and data based interaction you have with AI inside your company's walls has been powered by a closed weights American model behind the scenes running on American infrastructure. It might have been supported by some tiny open models that rewrote your query on the back end without you noticing. But generally speaking, AI in fashion enterprises has been a closed weight game dominated by suppliers from one particular country. Now that's set to change, I reckon. You're not going to see Western companies using cloud influence providers based in China anytime soon, I don't think. But I suspect you are going to see more of them either building their own server rooms up again and filling them with NVIDIA. DGX Sparks or Mac Studios as a stopgap, or you'll see them partnering with so-called neo-cloud companies and jurisdictions that align with their governance policies. And despite all that global perspective I just mentioned, I think the actual calculus is going to be much simpler. And it's going to come down to what it takes for an LLM to answer the kind of queries that everyday people working in design, product development, sourcing, and production, and so on would ask. Something like Show me the styles in this season's assortment that are below our target margins for their category. Now, you you don't need Fable Five to do that. You need a model that can understand that request, it can retrieve the right data, it can reason about that data, it can call whatever tools it needs to complete that, and it can give you a reliable answer. If two models can do that, and one costs 30 times what the other does for basically the same output, and the cheaper one is both ownable and self-hostable. you're gonna pick that one. Or to be more accurate, you're gonna trial it. You'd be silly not to. Now you might find in the process of doing that trial that despite the two models being very close in agentic capability or overall intelligence, the cheaper one doesn't actually really work for your workflow. Small differences in reliability between models and even model personalities can compound over time to make something that seems commercially compelling into a non-starter for end users. Anyone who uses Claude in their day-to-day life and interacts with Opus V is probably coming away thinking, I I hate the way this thing talks and I want to go back to Opus four point eight. And it's a it's a matter of personality as much as it is capability. There are also some other broader considerations to take into account here, but the key point that I suspect Yvonne was driving at with this question is that what we think of as frontier intelligence right now is going to end up being available for pennies on the dollar within a year. And that's something that you have to factor into your technology planning. I actually talked about some of this with ⁓ Francesco Bottigliero of Brunella Cuccinelli and Solome AI earlier this year in our podcast interview. So go listen to that if you'd like to hear his perspective on why that company's Calamacus platform used a mix of open and closed weights models and how they approached deploying them. It's a good reference case for some of the things that I've just talked about. Now, we recorded that conversation before the current wave of Chinese open weights releases, but I think it's still very valid. And that's it for mailroom number one. If you enjoyed this, send your questions in. ⁓ you can reach me at editor at theinterline.com, or you can get me on LinkedIn, or you can message the Interline's corporate accounts on Instagram and LinkedIn. If we do this again, which I really hope we do, I'm gonna bring in an even wider roster of industry friends. So I really want to make sure we get some incisive and ⁓ make sure we get some questions that get them really thinking about their answers. Grace is going to be back next week, so we'll have an edition of the edit for you. And my regular interview shows are still on their weekly cadence every Thursday, so come back for those through to the end of the year. When we reach the end of the year we'll be finishing what's been a really big, successful and unexpected season two of this show. And we'll move into season three for 2027. For now though, speaker-2: Yeah. ⁓ speaker-0: Thanks for listening to this first edition of Mailroom and I'll talk to you again really soon.