← Just Now Possible
Just Now Possible
From COVID Pivot to AI World Building: How Snapbar Reinvented the Photo Experience
01:14:21
Loading summary
Transcript111 lines
- [00:04]
A
Welcome to Just Now Possible with Teresa Torres.
- [00:09]
B
I'm Sam, I'm the CEO of snapbar. I've. I started it with my brother Joe who's on this call as well. 14 years ago, totally different company, not AI, not tech. We have pivoted multiple times to land where we are, but that's where I am.
- [00:22]
C
I'm Joe, I'm the cto, the Chief Product Officer. Basically look at all the our products, see where things are going in AI now that we're doing all software and yeah, that's my role.
- [00:33]
D
My name is Patrick, I'm the cto and it's been a ton of fun to take the original DNA of Snap R and really reinvent it both in the virtual like software world and then again as I'm sure we'll get into in the AI native world, both from a product and from a production perspective.
- [00:50]
A
Yeah, I am floored to Hear you're a 14 year old company because when I saw what you're doing, it was so creative and so AI native that I just assumed it was a young company. So I would love to hear a little bit about what in the world were you doing for 14 years.
- [01:05]
B
Yeah, tell me the back. It's a story I've told a couple different times. So 14 years ago we started as a traditional photo booth company and it was like just a side gig. So my brother and I had a wooden box that had a camera in it, cords running out of it, hooked up to a Mac running aperture.
- [01:26]
C
You would stand in front of it
- [01:27]
B
at a wedding, click a little clicker, it snaps your photo and shows it on an iPad that was screen sharing the Mac monitor. It was super old school. I think it was fun. But yeah, side gig. So on the weekends we'd go to a wedding, make a couple hundred bucks. We did that for about three years and then went full time with photo Booth's work in 2015. So three years in and grew it from weddings we migrated into parties parties turned into corporate parties, like holiday parties and then holiday parties turned into B2B gigs, conferences, trade shows, employee appreciation weeks. And so we started a single photo booth. By 2020 we ended up having, I don't even know, hundreds of photo booths. Our business model was like let's be the Starbucks of photo booth companies. So we would try to set up in a lot of different kind of locations where companies might have trade shows, conferences, these types of events and yeah, do these photo booths. At the time we tried to elevate the term. We talked about photo experiences because some of it would look like what you would understand a photo booth to be. Sonvo is a little bit more involved. It started sneaking towards being more technical. We had a backend gallery that we built out and stuff like that. But so that was 2020, pandemic happens. And we go from physical, logistical, heavy events, like we're physically going to events setting up cameras, computers, screens, everything, to can't run that. And that is when our true kind of tech journey began, was back in 2020. Patrick had come on board by that time and he developed our very first tech product that we just simply called a virtual photo booth. And then that's really where the tech story began. So we're five years old in that sense, 14 years old in as a company.
- [03:12]
A
Oh, I love that this was born out of COVID I love this for a lot of reasons. Like, I. Okay, so photo booths. I love that this started as like a very physical, like, lots of Adam's side gig, side hustle. Like, what an amazing, just starting story. And then you had the gumption to go full time. You grew it big. You were everywhere. Conferences were. And then the pandemic hits and everybody's staying home. And that could have just been the end of the business. But you also had the, like, wherewithal to reinvent yourself. Tell me a little bit. Let's just go back to 2020. What was the first virtual photo. What does a virtual photo booth mean?
- [03:53]
B
So it was not our idea. It was a client, actually, who had said, love that idea. That in the middle, when everything was quote, unquote, burning down at our company and we were refunding events that weren't going to be happening anymore, they said, does any of your photo booth stuff work online? And we said, no. That was just the truth at that time. And then at a retreat, at a company retreat that we happened to have scheduled months before the pandemic. And this was really early. This was like March, so we still didn't even have the full lockdown policies. We were like, is this worth exploring? And Patrick, you can correct me if I'm wrong, because my memory is almost fuzzy on this, but I'm pretty sure that we found something on the Barbie website that.
- [04:34]
D
Do you remember this?
- [04:35]
B
Literally I had been searching for, like, how to create, I don't know, like, photo booths online or something like this. And I'd sent it to Patrick because I don't know if this helps, but check this out and I'll let Patrick take over because that's when it was like, okay, could we actually do any photo booth stuff if people, I assume at this point people know what a photo booth does, but like online. Patrick, why don't you take over?
- [04:57]
D
Yeah. So we found ourselves in a co crazy situation in 2019, right before COVID Sampar had done its best year ever from a revenue perspective, team growth, all of this. And then all of a sudden, overnight, we had an amazing relationship with all of these tech companies and other providers and companies that we could sell to. But our product basically was illegal in the sense of you can't go to an in person event where all of these boosts would be. So we had to figure out, oh my gosh, how do we take all of this brand equity, all of these relationships and this core value prop of be able to provide this experience at an event, be able to capture photos and bring that into this new virtual world that we were throwing into. So all of these event platforms where you would go and be a part of a seminar or whatever, virtually. And then we were thinking about that and then of course it dawned on us. Gosh, WebRTC, the ability to stream video and audio from a webcam across any device, it's not just the native iOS or Android app, but also on on the web, made a ton of sense because we could give QR codes, we could give links out to these virtual events and then allow users to open that on any device, totally cross platform, and then engage in the same way that they would with a in person physical photo booth, be able to take photos, interact, and then have a camera or a photographer, if you will, at that virtual event, collecting assets so that the brands can then share those like they would an in person photographer. Just thinking from first principles about what are the assets we have, what is no longer usable. Our warehouses are having folks in different cities actually operating these massive pelican cases that we had shipped across the US and across the world. None of that worked. But just thinking again from first principles about the technology that would address that same user problem with the help of our customers, which has just been a thread throughout to today, hearing a lot of these roadmap ideas and feedback from our customers, putting all together that birth, the virtual photo booth, which has led into all of our AI native products today.
- [06:56]
A
Yeah, it's so funny. Photo booth feels like something that, like how in the world would that work online? But then I think if you start to get it, like why do holiday parties have photo booths? Like, why do conferences, like corporate events have photo booths? This feels even a little more weird. But Then you think about it, it's like such a nice brand opportunity to let people capture a memento that they get to have fun with and have a little bit of creativity that's on your brand. Like, what a cool thing. And the other reason why this really resonated with me, 20 years ago, I ran this in person, like, weekend scavenger hunt event thing, and we made trading cards as part of the event so that, like, people that were participating had to interact with each other because they were trying to collect all the trading cards. And I know that's like part of your AI photo booth. There was just so much that resonated with me of. I immediately thought of that meme for a while where people created AI generated action figures of themselves. There's something about this, like a brand plus an individual that's engaging with that brand, and you're letting that individual be creative in the context of that brand. And then I also go to a lot of conferences, both as a speaker and an attendee. And man, the normal stuff on the conference floor is so boring. And. And like, when I saw your stuff, I was like, this is really genuinely different. So let's tell me a little bit about what does AI photo booth mean today? Help people understand, like, what is it that you're offering and to who?
- [08:32]
C
Yeah, so similar to what sort of virtual booth did, in the sense that you could operate on a webcam, on your computer or someone, or even a physical iPad in a booth or something in a trade show. The difference before, we were just doing digital outputs, maybe background removal, because that was like a thing before generative AI stickers, that kind of thing. And then generative AI came around and all these new models came out. And so then we integrated that into the image capture taking. So you just take a selfie on whatever device, and then we're using these models to transform them, using basically like AI generative portraits, a lot of the time involving brand components as well. And then whatever theme the event or marketer wants things. So it could range from a space cowboy to, you had your dream property in the background and you're holding a new product that came out. It just spans the spectrum. And so then we started figuring out, okay, how can we actually create new products within this? We have all these tools out there, and so then we've segmented different pieces of this so we create more of a Mad Lib style, a quiz section, all using generative AI. And then video is a whole new world as well. So if that makes sense. Anything you can Imagine we can essentially create as an output now, which is hard, or trying to productize it a little bit more because really, it's like infinite options out there.
- [10:08]
A
I will say your website was super fun. I have not. I am being 100% genuine. It is rare that I land on a site where somebody applies. And I looked at everything. I like every product. I was like, this is so fun. And just the analogy that I'll give you just talked about space cowboys. Okay, so physical wedding. You go in person, they have a photo booth, they give you the cheesy costume stuff that everybody puts on and creates a silly photo. I always hated that, to be honest. But what you're doing is so much cooler. You take a photo with your camera virtually on your phone with a QR code, whatever, and then the AI pieces are turning you into the space cowboy. There's no silly costume that all gets layered on afterwards. And like a corporate party, they can have a really fun theme. You take your photo, you get added to the theme. And then I saw pictures of some people are putting up a giant board so that at the booth of the conference, you see everybody's silly pictures, which is really fun. Like I said, I loved the trading card idea. So you're not just doing take a picture and integrate it with the brand. There's a whole bunch of things you're layering on top of this. I'll share the use case that I wanted to do and I could not pull it off in time, but I will do this down the road. I'm giving a talk at a conference next Friday and I've been toying with this idea that there's this idea I have in my brain of we're getting to the point where software is getting cheap enough. The average human that's not a software engineer can see a need in their community and they can create software for that need. And I don't mean they're going to start a business. I just mean I refer to it as mom and pop tech. We're now at a point where we can just create little tools that our communities need. One of my ideas for this talk was I was going to try to motivate a whole bunch of tech people to think about things in their community. And then the photo idea would be. They would integrate their idea with their photo. So it's like I could collect all these people, have joined the movement, put it on my website, like, we're doing
- [12:03]
B
this, not actually ready to give.
- [12:05]
A
What's that?
- [12:06]
B
We could do that. We could do that by next Friday.
- [12:08]
A
That would be fun, right? It turns out it would be fun.
- [12:10]
C
It's a good idea.
- [12:11]
B
I already know the product I'd try to fit you with.
- [12:13]
A
Yeah. The challenge is that's not the talk I'm giving on Friday. No, I'm not prepared to give that talk yet. It's just not ready. Okay.
- [12:22]
B
It's a good idea, though. I love it.
- [12:24]
A
Yeah. I am going to start a movement around this. The thing that makes me most excited about AI is there's all these little annoyances that would never be a good business, but are perfect problems for software to solve that are now possible. So I want to do this, and I definitely want to do this with you, launch a campaign with you. So we'll get into this later. And my point is, I can see so much potential in what you're doing as a way for people to connect with an idea, as a way for people to connect with a brand. And it's very. What I love about it is it's really participatory. And if I think about conferences, like one of the hardest things is like getting people to talk to each other. And I think a lot of what you're doing with this, like the individual contributes some of the creative elements that goes into the photo, but then so does the brand. These are natural connection points. So I think that was the piece that like jumped out at me is, wow, this is really fun. And I can see so much potential. Thank you. Which is why you don't have to pitch your product. I'll just do it for you. Okay, let's get into the technical bit. So it's Covid you're trying to figure out virtual. You have like early virtual AI photo booth and not even AI yet, right? Just virtual photo.
- [13:35]
B
Yeah, just virtual, Yep.
- [13:37]
A
Where does AI enter the picture?
- [13:40]
D
So technically we started with traditional ML based AI applications, not just the generative side. So that looked a lot like background removal. And then also, I'll never forget us, we were all excited about this white paper that came out about some new methodologies we could use to essentially relight an image. As you can imagine, with WebRTC based images, one of the biggest challenges we have is you're typically working with really poor camera quality because we're just taking a still from the stream of frames that are coming from a device which is typically an iPad on really bad event, wi Fi or another mobile device. Now, of course, this technology and just the quality of cameras have come a long ways over the past six years. Since we originally built this out. But being able to relight and just overall enhance an image has been a really core focus of ours. So we had some traditional ML applications of that background removal, touching up images. But what really got exciting, if you can think back almost four years now to the release of stable diffusion 1.5, which again feels like an eternity in the generative AI timeline. And that actually pre ran ChatGPT just to give a sense of where we were at the time. But what was really exciting for us is this was the first model that could really come out with interesting outputs. A lot of superheroes and other outputs like that for us back then that were really creative. As you're kind of mentioning earlier, it's a big focus for us is how can we allow brands to really world build, to allow them to take users to and to intimately kind of bring them into a world like a marketing activation, both from the engaging perspective to help draw attention to a booth. If you're on the trade floor or at a trade show and there's a thousand other booths around you, how do you stand out and really grab somebody's attention? So be able to allow them to be entertained through that process, but then also to really bring them into the world that you're trying to create or that you're trying to kind of connect them with your to your brand with. So that was a huge powerful piece of using the gen focused models. And of course now stable diffusion 1.5 seems like absolutely terrible compared to nano banana 2 and GPT image 2 and all the current models that are out. But that was where we originally got our focus. Also at the time it was really hard to get good outputs. We had to actually spin up a bunch of H1 hundreds and H2 hundreds Nvidia GPUs in our service. We didn't rock them ourselves, thankfully. That would have been a ton of work. But through a couple different vendors we'll access those. But we had to create a whole process for actually creating fine tunes. So generating loras and actually doing 10, 15 minutes of fine tuning on a model in order to then create predictions or generations of different images of somebody. So we take different brand assets alongside the user's likeness and then would prompt and bring all that together into a fine tune and then could generate these really great looking photos. And this is something that nobody else in our space had the technical capacity or time and energy to really put into building out. So that was a really fun piece for us to be able to really focus on working with those early models. To get compelling outputs that people couldn't see elsewhere.
- [16:57]
A
Yeah. Like what's coming to mind is I just want to remind people where we were at this point. Zoom had stickers, right. Like you could put like we didn't even really have good virtual backgrounds. We just had these. Put horns on your head and hopefully they were somewhere on your head. Right. Like it is funny to see how far we've come, especially with image generation. But like I just think about the original stable diffusion models and like I very early on, because I give one or two big conference talks every year, I force myself to start using AI generated images in my slides. So then I would build up this natural progression and have that evolution of. It's a little bit like if for people that follow Ethan Moloch, it's a little bit like his plane test, you get this like natural evolution of how these models perform. And I have some really crummy slides, slide images that are coming to mind as you described this like really just broken. Right. Not great. Patrick, what is your background? Do you have an ML background?
- [17:57]
D
I don't have a traditional ML background. I. So I studied at UW Computer science and then was down in San Francisco for three years working at two different early stage tech startups as a software and product engineer. So I got a ton of full stack experience actually originally taught myself how to program in middle school, helping out different friends with creating their websites. I just Learned everything from YouTube which is it's been a lot of fun to kind of see YouTube throughout all these years, just help people in different capacities. So that was. My background is part of formal education at uw, but a lot of just self education. I'm just incredibly curious about any sort of new technology and really the definition of kind of that t developer where I'm specialized in a few areas, but I just go super, super broad. And of course anything AI related has just been absolutely fascinating. And when you have legends of the industry like Andre Karpathy who led self driving at Tesla and was essentially a founder at OpenAI and it's just done a lot of amazing things who are creating four hour long YouTube videos for you to learn and absorb or all these white papers that these labs are publishing, anything that you want to learn is out there. And of course AI you can summarize and kind of zoom in or zoom out as far as you want or find adjacent ideas that really connect between a white paper that you're reading and the exact code stack that you're working with and the exact problem that you're trying to tackle. So the ability to just be incredibly agile and learn quickly anything that you need and figure out how to apply it, it's just incredibly effective now. So no traditional ML background, but a lot of curiosity led self education.
- [19:30]
A
I love that I'm old enough to vent through the first wave of the web. I was actually in College in the mid-90s in the Bay Area, and it was like one of the most fun time periods of my entire life. I remember like a new version of HTML would come out and everybody would go read the spec and see what was new and what can we do with this. Like, I specifically remember when frames came out and all the crazy things people were trying to do with frames like it. And it was a similar idea, like, it's new for everybody. So there was a lot of openness and a lot of sharing and a lot of people trying to put together standards like we're seeing right now. And it's actually what really motivated me to dig in this time around and to start a podcast around this and start sharing and start getting people to tell stories. Patrick, I think what's amazing to me about this, that was my long way of saying, your curiosity resonates with me. I love digging in and just learning a new technology. But what I've heard in your story so far, for all three of you, not traditionally a tech company, started to move in technology really because of a big pivot required during COVID but to go from that to bleeding edge, the image generation, that's a giant jump. And Patrick, I was expecting you to tell me. Oh, yeah, I've been doing machine learning engineering for a long time. And so this was just like the next level on that. The fact that's not the case is pretty remarkable to me. So I'm curious from all three of you, what was the impetus to let's push on the AI part of this? And then maybe, Patrick, you led the charge on this. But I don't want to assume. What did those early days of just learning and figuring it out look like? I'm just thinking of all the people that are hearing this and being like, I could never do that. They probably could. They just don't realize how small we start.
- [21:21]
B
Yeah, I will speak to the technical side. I can tell you one of the things that we began to realize again and again over 14 years, which actually the length of business helped us in this regard, specifically because one of our core verticals is events. You could say it's not even a core. It's like it's been the main vertical, now it's actually expanding more into marketing. But in the world of events, if you're a brand, you want to do something new. And because events are inherently like trend based in a sense, like if you did something successful in 2020, you don't want to do the exact same thing in 2021. Like you're trying to offer something new, right? You've spoken at conferences and this is different for speakers. You might have a killer talk that you give at a lot of different events. Generally speaking, if you're invited to the same conference two years in a row, you're going to try to do something different even in a talk, right? So it's the same thing in the world of events. So by the time AI had, let's say that 2022 is when we really started pushing into AI. This product came out in 2020. Patrick built it, I think at the spring of 2020, the very first non AI virtual background. So two years in, we had done genuinely like thousands, thousands of events of virtual events over the kind of pandemic period. And we began to see a decline in the repeat business that we'd normally get. Because I think just from a pure business standpoint and in the world of events, like they'd done it before, two, three, four times. They've done it at their virtual trade show, they done it at their virtual conference, they even did an employee appreciation week. And now like it's 2022. What new things could we do? And so we were even before AI, trying to come up with new exciting ways to engage people again virtually via tech via the WebRTC method. Our roofs were obviously in photo video. And when it was probably Joe or Patrick, that kind of where it clicked Joe is. We always joke around that Joe sniffs out new tech like earlier than anyone else it feels like sometimes, or at least not even new tech. I don't know if it was like AI specifically, but he has an actual photography background. And when we noticed just cool images being generated, I remember really early on, before we even ever thought of let's be an AI company, it was like, oh, wouldn't that be so cool. We would do all this stuff to try to fit it into a theme or to the world that the brain is trying to create, as Patrick said. And with AI, this like glimmer emerged. You could just, you could do that. If Microsoft is launching something that's really fast and they wanted to try to put people into a speed based world or somehow communicate speed and which Is like sometimes the general direction we're given from a creative standpoint, AI could do this so much better than we would have to like piece it together with a background and an overlay and all these other things that we do to try to create your output. So that from a business perspective, that's when we started asking that question was this would be new, this would be cool. If we could integrate it. I think it'd be, it'd do really well. Brands would really get behind this because we'd be able to bring people in and it does the creative heavy lifting for us instead of us having to do it all the time. So yeah, that's part of the answer.
- [24:29]
A
Yeah, this is a piece too where I think AI, like people say AI isn't very creative, but I actually completely disagree. Like I look at some of the I use, I'm a big fan of Nano Banana Pro and I am just blown away by the variation in images that I get. And of course you have to learn some prompting skills and understand the boundaries of the tool. But what you're describing, that you can just start with a concept again, a bunch of instantiations of that concept and you're almost like an advertising agency, like an AI driven, maybe campaign agency is a good way to put it. I think it's just such a fun way to apply AI. Okay, so Joe, do you remember was there something you saw that instigated that we got to bring some of this generative image stuff in?
- [25:21]
C
Yeah, I think when we first saw some of the outputs in stable diffusion, when you compare them to today's outputs, you just think they were horrible. But back then when there was nothing to compare them to, they were awesome. And we stable diffusion was that first click where it's okay, we're all in on this. I think what helped us actually was the fact that we'd been in business for much longer because we'd seen a lot of these things. And so instead of what Sam was alluding to, you have green screens or whatever to create these concepts, you could do it in a way that was way cheaper, faster and very creative at scale. And so we saw that and so we doubled down on it. And then as new models were released, pretty fast actually. And then Black Forest Labs introduced with Flux and everything. Even before some of the GPT stuff, you could quickly see where it was going. And so we were thinking back then will be able to create full videos as like an advertisement where you're part of this creative ad and there's music and it's a full production essentially. And that's where we are now and it's just getting even better. And so I think we just saw those glimpses early on, specifically with stable diffusion. Yeah. And that's where we just put our eggs in the basket.
- [26:35]
A
It sounds like you had already as a company had this aspiration of what we deliver through virtual AI photo booths or virtual photo booths. Forget the AI part. There was already this aspiration of let's help our customers world build and give them an opportunity to put their customers in that world. Is that true? Is that sort of already a driving like an inspirational direction for the company?
- [26:59]
B
Yeah, I'll let Patrick speak to that more. But 100%. It started in physical photovis with green screens literally putting you in a world. And then the very first use cases of the virtual photo booth that Patrick built out was background removal which again was AI but more on the ML side because every brand would be like at the base level they're like put a step and repeat or put a branded logo background behind people. But a lot of the time it was like it's face. Right. Or we'd tie it to this. Like there'd be some, again some world. It's very. That's pretty minimal when it comes to world building is just put them in front of a background but. But generally like you mix that with the stickers or the props with an overlay with. And then you get like a cool branded piece, branded output. And so that was something that Patrick developed, had developed already and when. But it had been out for two years and naturally we still run the virtual photo booth product today, albeit it's been rebuilt, it's way better. It's really, it's cooler. So it's not like it just disappears off the face of the map. But when AI hit it was the AI wave that took over and it really just accelerated that ability a ton.
- [28:10]
A
Yeah. And clearly accelerated it because I will say looking at your website, it looks like you're an AI native company. Okay, so Patrick, let's go back to the like there's this seed stable diffusion models come out. You look at it, you're like whoa, this is a game changer. This is new technology. You're not familiar with it. You didn't have a machine learning background. What did you do? Like how did you start to learn? Like how are we going to deploy this? How are we going to take advantage of this?
- [28:35]
D
It's, it's a great question. With all these new pieces of technology, there's so Much curiosity as you're alluding to with I remember like HTML5 that's coming out and just getting so excited about all the native easy to build pieces. And we've done that too with WebRTC, with HTML Canvas, with a bunch of core elements that sound kind of boring like HTML Canvas, but when applied in a really creative way are incredibly exciting in terms of again, what we're able to build and what we're able to enable our clients, our brands to to create. So in terms of the technical piece of that, obviously there's a lot of creative ideation on what that looks like from a product perspective, thinking from first principles, applying customer needs, which we have an unending list of with the technology. But from a technical perspective, again, what's amazing is the number of resources that are just available for free online. And I'm a bit biased here because my background, again since I was a middle schooler, was creating and consuming a ton of content on YouTube and on other platforms as well that have just allowed me to continually learn so much of what I'm building today. So YouTube, AI, Twitter now is a huge help, but back then it was more like reading the documentation of different providers, looking at, okay, who are all the folks that are hosting these diffusion models like stable diffusion and how can I reverse engineer the way that they're running and fine tuning and doing other sorts of lower generation and like all the technical pieces and really piece that together or finding talks From Stanford or on 3 Carpathy or other folks that are sharing these resources. I'm also lucky to be connected to a couple researchers in Seattle. So that was really helpful having a few conversations with them especially early on to get a better sense of the really more in that case the direction that the models were headed and then of course reading white papers as well. What's nice though is a lot of what we're doing and a lot of what's bred the innovation and drive towards generative AI has all been out of very direct necessity, which hasn't really been fun going through it, but with several near death experiences through Covid for the business, when you're bleeding tons and tons of money in a couple months because your entire product is obsolete, that just forces a gritty resourcefulness beyond anything else. And I think that's one thing that's been humbling for me as somebody who tends to overanalyze and think a lot about things. To see Sam and Joe navigate so well through is just having constant chaos thrown at them and then just the ability to Just not hold anything dear, but completely pivot on a dime and make really bold bets like WebRTC. And photos are the future generative AI and the ability to take any data, whether that's a photo or some text or whatever and turn that into a world, then that's going to continue to grow. As Joe alluded to. With video being a huge bet that we've had for I'd say four year, five years probably now, that's been a huge help. It's just, okay, we don't have an option. We have to figure this out and we have to address these, the clear customer needs and demands that we have. So I need to figure out a way to creatively bring that together through AI, through Joe's visual capacity. As somebody who studied photography in Europe and is just an incredible photographer. Like, how do we kind of take that vision and that aesthetic that he has, Sam's deep knowledge of the industry and kind of combine all that getter together into a product? So yeah, out of necessity is how I would answer your earlier question as well.
- [31:51]
A
You know what I love about that answer is this is actually what I'm speaking about next week, which is everybody in the world is feeling really overwhelmed by like, how fast is AI moving and what do I learn and what tools should I use and how do I keep up? And FOMO and I went from, in a 14 month period, I went from barely using AI to now I have multiple AI tools, I have an AI podcast, I have an AI blog. It's just what I do. I'm all in. And my talk is my story of I literally started with a specific customer need and I got curious, can AI help with this? And then the next week I ran into another problem and I got curious, can AI help with this? And I think this is maybe the thread people miss. Instead of chasing all the tools, instead of chasing all the technology, there's probably a need right in front of you. Which of those tools helps you with that need and then do it again. And this is. It just compounds over time. I look back over where I was 14 months ago and I'm blown away by where I've come. And I can imagine you have gone through a very similar thing with image processing and looking at your products. It's just really clear that you are now leading edge in this area. But you didn't have a background, you just started and that's the message. My message in my talk next week is just start.
- [33:04]
B
That's great.
- [33:05]
A
Follow the need.
- [33:07]
B
Yeah.
- [33:08]
D
One thing I'll Add on that as well is the people with the most experience in generative AI have arguably maybe four years of experience outside of a course, like the researchers. But like anybody that's applying the technology, the applied AI side of it, which is firmly where we fit. And what we found too is, as you're mentioning, it's not the technology that's really driving things, it's the customer need enabled by the technology. That's just one path that we see to fulfill the core problems that are driving us to figure out how to use this tech. And then what's also really exciting is more than anything else, I'm a technologist through and through. Absolutely love diving deep and experimenting and figuring out, especially with a new technology like Genai, where nobody knows what the bounds are. Yeah, this is novel research that we're doing that so many of these AI native companies are doing, like figuring out. I've all kinds of personal experiments I'm doing too, with an AI, second brain, et cetera, et cetera. But it's so cool to see what emergent capabilities are there. And then what's even more neat than that is being able to take again, Sam's experience in the industry for 14 years, Joe's creative expertise, and apply that sense of design that, that taste, if you will, a word that's used a lot with these models and they'll figure out how to truly make like new products and new creations just by bringing those in. So it's the technology is critical because it enables all of this, but it's really the experimentation and the creative thinking around it that has enabled these new products, if that makes sense.
- [34:34]
A
Yeah, this has definitely been a theme on the podcast. I think in the early days. We're still in the early days, but I think there was a period, and I think some still argue this, that like the foundation labs are going to eat the world, but I don't. I actually think it's the exact opposite. It's looking more and more like models are becoming commodities and where the real value is. All these companies that have deep subject domain expertise, have a deep understanding of the customer and are applying the technology in ways that work for the person who doesn't know what a transformer is. Right. There's just so much value creation at this applied level. And I think especially with open source models starting to catch up, this is going to become more and more true. I'm seeing products ship with small models and you don't even have to pay for tokens anymore. It's just infinite usage because the model's small enough and runs on your computer. This is amazing what is becoming possible. And I think my background's in discovery and teaching customers how to really understand their customer. And so to see that this is where a lot of the value layer is coming in AI products. I'm also a technologist, so I really love the intersection of both. And it's just. It's really delightful to me to see that the teams that are running ahead are the ones that are doing exactly what you're doing. You're combining industry, market, customer knowledge with some maybe photography domain expertise with the technical expertise and just running after it. And sure, maybe one day OpenAI or Anthropic is going to release some crummy version of an AI photo lab, but you'll be so far ahead, it won't matter.
- [36:10]
B
That's an interesting thing because when we first started Joe, we called him, I do a lot of sales and marketing. So as the sales team would be like, oh, our prompt engineers. And at the time it was like, ooh, like prompt engineers. What's that? Because not everyone was prompt engineering. And genuinely the kind of stuff that you had to do with stable diffusion 1.5 was really complicated compared to what it is today, where you just change some of the text you're writing and get a completely different output.
- [36:35]
D
I don't.
- [36:35]
B
Joe could speak to this, but there was all sorts of technical stuff. Literally like a page of.
- [36:39]
A
I don't know.
- [36:39]
B
Joe will have to speak to that. Whereas today anyone can go through and create their own image. And so we. When the lab started including image generation in the their. In their products, there was like this mixed feeling. I remember at the time of, is this good or bad for us? Like, it. We used to have this kind of, like, corner on the market where we would go to an event, turn people into, say, superhero or. It was usually more interesting than that, but let's just say that, for example, and people were like, this is so crazy. I've never seen anything like this. And then when they started being able to do it at home, we thought, does this ruin it for us because it's no longer special? But what actually ended up happening is that the requests just started getting more and more complex and interesting and creative from the clients because they'd start doing Inman's Generation at home on themselves and be like, can you guys do this? But also bring our logo and have our logo be in the background and also do this and somehow incorporate the person's name. And now we're Dealing with like reference imaging and prompt injection and all this other stuff. And that's what like led to some of our new products that you might have checked out. Like trading cards uses prompt injection for bringing in text to the image itself or AI and so on and so forth. So it's, they didn't actually hurt us at all. In fact if anything people doing this more at home only made them more creative and almost like elevated their standards a little bit. I think it would, it would be risky if we were doing. Yeah, I don't know. I think it's net.
- [38:04]
D
Net.
- [38:04]
B
It's helped us, it's helped us in every way.
- [38:06]
A
I have an analogy that keeps just jumping in the back of my mind that I'm going to share. As a kid I grew up in the arcade generation, right? Pinball, literally like the stand up fighting games, whatever. Yeah. Where you put in the quarter, you run out of tokens, you're done. Right. I'm also the generation where the first Nintendo came out and the first Nintendo did not put arcades out of business. What they, what it did was it forced arcades to have real guns and become rides and be more like dodge that bullet. They became more interactive, more real. And we still have this today. Now we have golf simulators that look like you're golfing. So I think there's this idea of, I think what happens is when it becomes available in the home, what that's doing is it's growing your market. They start to see the potential and then it lets you run way ahead and show what could never happen at home.
- [39:04]
B
Yeah, yeah, that's a great point. Good analogy.
- [39:08]
D
We also, we think a lot about this Sam Altman quote of when a new model is released. Are you thrilled because new capabilities are unlocked for your product or are you terrified because what your startup does just become a commodity built into the post training of the model or whatever's absorbed. And I do think these models will continue through mostly the post training in RLHF to absorb more and more of the harness and pieces of that infrastructure like around rag and some of the second brain stuff I'm experimenting with. But I just don't see a world in which, or at least in the near the medium term where it's going to touch much further into the application layer as you were mentioning. And for us it's our brands, they're not going to spend the time to build four years or the equivalent of prompting experience like Joe has and the tastes and just the ability to articul, articulate we want 35 millimeter film from this specific camera that Joe just happens to know about in order to get a very unique aesthetic for a style they want.
- [40:01]
A
Right?
- [40:02]
D
Yeah, exactly. So there's a lot that goes into. The models are very capable. But what is all that you're surrounding the models with? So these AI wrappers which again that was a bit of a bet as well. Are we going to become commoditized? But as we thought about it. Okay. Are B2B group context very different than a one off user using ChatGPT and if anything, the education of the market with that new capability built into ChatGPT that people are understanding or like Sam mentioned, are able to more competently communicate to us what aesthetic they're looking for because they can play around with the tech. That's amazing because it feeds right into enhancing the value proposition of our suite of products. So I agree with you. I think the application layer is in a very exciting place to be right now.
- [40:43]
A
This is why earlier I had made the comment, you're almost like an advertising agency. And I didn't mean that in the like adsense, I meant it in the like the high end branding advertising that's doing like the creative work. Right. Because you're bringing together this technology that opens up whole new avenues. But there's clearly this creative flair that comes with it that I think is a really cool intersection. Okay. I want to get a little bit into the tech stack so maybe what we could do. I know you have a lot of products and so if you want to pick one. What I think would be fun is just if you have an example of a customer campaign. We don't need to get into the details of the customer or anything like that if you're not able to share, but like something that we could talk through to make this really real. People understand the types of things people are doing with one of your products and then maybe get under the hood a little bit. What's AI, what's not? How does that work?
- [41:36]
B
So let me paint the. Because I think it'd be helpful for listeners the business picture and then I'll hand it over to Patrick who can speak into the tech stack stuff. Because your comment about we have a lot of different products. We do, but let's just focus in on like the core stuff.
- [41:52]
C
Right.
- [41:52]
B
We have, it's pretty new. We've done this kind of work for a long time. But what we're really trying to do in bringing all of our products together is build a cohesive platform that like will and we're pitching it as like an experiential marketing platform. Okay. Photo booths, photo experiences that kind of fit into that category. And we plan to expand beyond even photo video into even other mediums. So if you have this experiential marketing platform that allows you to create, spin up these activations, whether for a trade show or conference or whatever, you have a decision as an event planner or event producer, event marketer, who's my demographic? How many people are going to.
- [42:26]
C
I'm going to.
- [42:26]
B
Am I trying to reach. What exactly do I want from them? What do I want to. What do I want them to do for me? And you might choose. Okay, I want to turn everyone into an orange superhero because my brand color is orange and my product is related to superheroes. I don't know. Or you might say, I really want to put them into this particular environment. Right now is the World cup, just to make it real for everyone. So we have a lot of activations going on right now for the World cup, not necessarily directly for FIFA, but for all sorts of brands that are doing engagements around the world. Cup fan experiences. This is a perfect use case for an experiential marketing platform. And so many of the World cup experiences that we're doing are being hosted or used in real life. Like physically. People are like putting our. These micro sites we build onto iPads. They're allowing users to scan them via QR code at stadiums, in the ground, et cetera. And we've seen so much already, turning people into the soccer player, putting them, turning them into an animation, standing inside the stadium, wearing like a branded T shirt for some other company. Soccer is a huge thing. There's tons of soccer fans. And so the. They're using our platform to create these like marketing activations basically, right? Like it doesn't always have to be take the shape of a conference or a trade show, but it's like anytime you want to engage your audience from the business, the business values that you're usually getting some type of lead data or an email. And then you're also helping them generate content and fingers crossed, they post it to social media. And now your brand amplification begins to happen. Right? This is generally the way they view our business from like a, is this worth it or why would you do this? So a lot of the products work together inside the platform. So AI photo booths, digital photo booths, AI video, AI stories. These are all separate experiences, but they're under the umbrella of our experiential marketing platform. And this is a lot of that architecture is, Patrick, kind of thinking through how that will work.
- [44:22]
A
Okay. And then I want to add one layer to that, because I think one of the things that stood out to me. I hope this is true. This is what I understood on your site. Maybe it was like your Mad Lib product. Maybe it was the trading cards. I don't remember. But you do have one element where. So I'm a soccer fan. I'm at a game, I see a QR code, I'm taking a picture from what you described. You can turn me into a soccer player. You can put me in my country's colors, whatever. But I believe there's also a product where the fan can put in something like the brand could say, when you take the picture, what does it mean to be a Netherlands fan to you? And they can type something in and then that gets integrated into the picture. Is that correct?
- [45:02]
C
100%.
- [45:03]
B
Yeah, 100%.
- [45:04]
A
This is the part that lit me up. Because now it's participatory.
- [45:07]
B
Correct.
- [45:08]
C
Right.
- [45:08]
A
And now you get. It's not just the brand's view. Yeah. The brand's gonna put you. The country's gonna put you in the country colors. And you're at a soccer game and the brand is deciding that. But you get to add your individual element that makes your photo unique. And I think that's where. Think about World Cup. I don't remember which country it was. It was just like marching through the streets of New York being insane. Like, probably the same. I think the Scots are in Boston. But you could imagine, like, how much camaraderie that creates within that community, all under the guise of your brand, which is really cool.
- [45:42]
B
Yes. Yeah, 100%.
- [45:44]
A
Okay, let's. Okay, so let's talk about how does this work? Like, I. I see a QR code, QR code, I take a picture of it. What happens on my phone.
- [45:53]
D
So this was one of our early bets, which is going cross platform because a lot of people are focused on apps and then like, you know, mobile apps and then app clips as well. But at the end of the day, you want to be able to reach everybody, no matter what device they're on. And by far the best way to do that is using web based, like web apps. And then it's specifically for us WebRTC to capture the image. Given that every manufacturer is dead focused on building out the best camera, the best photo and video tech that they can. And then the. Of course, the WebRTC side has come a long ways, especially through Covid, given the forcing function of all of that streaming.
- [46:29]
A
So I'm not familiar with WebRTC. Do you want to give the. Just a quick Overview?
- [46:33]
D
Yeah. So WebRTC is a set of different APIs within the web browser that allow you as a web developer to utilize different resources on a person's device.
- [46:46]
A
Okay.
- [46:46]
D
So in our context, WebRTC specifically is typically like the camera sensor and audio sensor as well. So you can stream like right now as we're recording this, or if you use platforms like Google Meet or Riverside or Zoom or anything else that is going through the WebRTC protocol using that web API on the browser in order to basically stream those packets of photo and audio information. And then there's a few niceties like being able to capture stills with the high res sensor, snapshot and a few other things that have been built on top of that. But at the end of the day, it's just a way of. You can think of it as like the SDK for building iOS apps that allow you to access the phone camera, but like the web version of that.
- [47:26]
A
Okay, yeah, that's helpful.
- [47:29]
D
Yeah. So we basically we built off of that. So when you take your phone and take a screenshot, or click on a QR code, or click on a link, we open up a web application that then brings you into an experience. And this is highly customizable within our product. So different brands want to organize this differently. Or as Sam mentioned, this is where the idea of all of these different products lives. A brand can have the trading card experience or a number of different experiences depending on what they're going for. So we can configure that to be presented, however is best for the end user. And then we will walk them through a few pieces. Sometimes they're entering some information, as you mentioned. Why are you a fan of this brand? Or why do you like this flavor of this product or some deeper. We had some really cool concepts where we were writing a poem and having the user give different input and that poem also turned into an AI image. So there's all kinds of data or like where do you live in the world that we can collect there, not just from a lead capture perspective, but also from a churning that data along with optionally a user's image into some sort of really creative video image, audio, video based thing, or even just like a text output. So we have this Capture series using WebRTC, using user inputted data, using any integrations with brands as well, in order to automatically pull in data from an id, for example, and then we transform that into the end product, going through our AI native generation pipeline and then have a series of different displays in order to show this. So you can imagine if you're at in Expo, how do you engage people, how do you show them the outputs that people are generating at the event? That's where the gallery live. Mosaic. A few of our other product offerings show as well. So we've got the capture and data input side, we've got the AI native transforming of all of that input data piece and then we've got the display piece which is where these assets go. There's a lot more detail there. Joe's got some really interesting insights too on the prompt engineering. But one thing I want to make sure to mention as well is the software development lifecycle, which is totally separate from the actual product itself. This is more of the the way that we build a semi dark software factory suite of tools.
- [49:47]
A
Yeah, explain that last comment a little bit.
- [49:50]
D
Yeah, so one of the most exciting things for us, and this is almost a lucky factor of just us being in on the generative AI side of things so early because of stable diffusion again out of a forcing function of oh my goodness, this is an amazing way to deliver on these customer needs. We got really AI pilled really early, especially with ChatGPT off of 3GPT, 3.5 and beyond. So the software lifecycle that we're currently running, or as a lot of folks are saying that like dark Software Factory, which we're not fully at, is being able to use AI models throughout the entire process, from figuring out and kind of generating roadmap or product ideas all the way through the development and review and then the verification side of pushing all of our new product features and software into production.
- [50:40]
A
I'm sorry, I got you. So not just you clearly use AI in your image processing, but you're also using AI to help you build your entire platform.
- [50:52]
D
One of the features that I'm really excited that we're just pushing hard into right now is the ability to allow our customers and our users to essentially vibe code within the confines of our platform. So this is taking the building blocks like WebRTC and all the years and years of development and stress testing that we put that through along with different elements of our products and we can combine all that together in an intelligent way using things like the Cloud Agent SD or the Cloud Agent SDK and other tooling in order to bring a level of that done for you service through Genai, but also having the security confines and frankly just the lack of overwhelm that tends to come with using something like a bolt or level ball and trying to build something from scratch, which we're very excited about that. Cause I think there's a quite a sweet spot of giving direction and taste and discernment around what that end result should be from a product perspective, from a prompting perspective, et cetera. But then also allowing people to have the creativity as Sam was alluding to, with our clients coming to us with concepts for different visuals that they want to fully express themselves within the platform. So that's another area outside of just the image generation pipeline and the codegen. From a software development perspective, there's that middle ground as well that we're. It's just been a lot of fun for us to explore.
- [52:12]
A
This is also becoming a theme on this podcast, which is surprising because I'm seeing it across a wide variety of domains. But what ends up happening is a company builds tools to help them service a customer and then with time they realize those tools are getting easier and easier to use, but they can just push them to the customer and then that frees them up to kind of then look at the next layer beyond that. And I see this a lot. Like we. I've interviewed probably three different teams working on customer support agents. And like what's hard about customer support agents is the evals and who's looking at the data and doing the error analysis. And like at first they were a lot of the teams are doing it on behalf of their customers. Then they pushed it to like an admin interface and their customers were doing the error analysis and then the company would do evals for them and then they built agent that would help them walk through the evals themselves. And then I was just talking to another company where they actually do fine tuning and they built a fine tuning platform. We ingest this data, here's what happens in the training, here's what happens on the outside. And then they pushed all of that to their customers that are not technical people. But it was like so componentized that they could just start to push it to their customers. I think this goes back to our earlier conversation about the value at the application level. I don't think we could build a fine tuning pipeline for all use cases for any types of data and hand it to a non technical person. But if you get a really specific domain and you understand the challenges in the domain and you understand the boundaries and you can put safeguards in place and you know the type of data that you're going to get, suddenly you can build A fine tuning pipeline that works for a very specific type of data that is safe enough to push to a customer. That's pretty fantastic. So it's funny, I've been doing this podcast. I think I started it last fall. So coming up on a year and early on, the themes were like everybody was learning context engineering and like maybe how to define tools for their agents. But it's been really interesting to see as more and more teams mature, how the themes and patterns just keep maturing. It's a lot of fun. Anyway, I'm thrilled to hear that even in your like video, AI, image generation and video production stuff, you're also getting to the point where your customer can describe the world they want to create and you're able to do that for them, which I imagine is just going to free you up to think about what's the next layer.
- [54:44]
D
It's a really exciting time. And as you were mentioning too, even with evals, like for us, like we know a lot of the gotchas or I would almost argue all of the likely to happen gotchas we for sure know, especially around image models, which of course we're extra safe with. And our brands really care about that safety side. So for us, being able to run through a bunch of those use cases and our evals are nothing too exciting, but just to be able to like through a spreadsheet quickly check a bunch of the core use cases or. Joe has such a amazing eye for what to look out for, both from the safety perspective, but especially from the creative and aesthetic perspective as well. He can get in there and just like quickly tune the system prompts in order to anytime a new model comes out. But that expertise is as you're mentioning, where do you even begin to describe that to somebody who all they care about is bringing a couple thousand satisfied people through their event in like with that contact info, they don't care at all. Or should they care about having to figure out the engineering behind that? So it's this really fun middle ground of exposing just enough creativity but then having those safeguards in place.
- [55:47]
A
Yeah, this is reminding me of one of the things I wanted to ask your team is, okay, so you've got a brand giving some input, they're creating a world. You have a user who is giving some input, but ultimately AI is generating an image or generating at least parts of the image. What are you doing to make sure that I can imagine there's already the image generation models have some safety around, maybe violent images or nude images or whatever. And so Maybe you get some out of the box. But I also know with branding, specific companies have brand guidelines and principles that like, what may not look like a problem to us is absolutely a problem to them. Or even things like that maybe are associated with a competitor. Maybe it's not even a competitor brand, but it's a concept that's really like polar bears. Pepsi doesn't want a polar bear in their images ever, no matter what the user types in. Right. What do you do in that case? Like, how are you making sure that these images are going to work for the brand?
- [56:51]
C
A lot of it has to do with the system prompts or the prompting that goes into these. And it was really exciting actually when the reasoning models, reasoning image models came out like a nanobananapro or GPT image 2 because previously there is like for multiple models you'd have your main prompt and then you'd have negative prompts as well. So you do just as much negative prompting as you would the regular prompt. And then these reasoning models came out and now it's just all one prompt. And so a lot of the time it'll look like paragraphs and paragraphs of text that most people are used to like to just type in a sentence or two to create an image. These reasonable models can ingest so much information and so it'll look like a huge document of some of these prompts that we're doing it and it'll include what not to have and it's all specific to what's actually going. So if it's like soccer themed, okay, I don't want these logos to be in there that are typically on these jerseys or whatever. And I only want this and that. Or if you don't want the polar bear. And then you go and you're basically, you're creating the entire scene not only with the regular prompting, but also what not to include with some of those safeguards and everything. So there's a lot that goes into. But the reasoning models have changed how we used to do it. And so now it's just these huge, typically long form prompts that are massive and to most people would seem overwhelming. But these models can ingest all that data and then send it out. So it's fascinating how it's actually progressed. So it has changed for us over the years.
- [58:29]
A
And is that something that you're creating on behalf of your customer? Are there ways that they provide inputs and you're generating that prompt? How does that context get defined?
- [58:39]
C
Or usually creating that master prompt and then anything that they're usually injecting is a piece of that prompt within our con, like our sort of rails in that sense. So it's not like they're coming up with it totally on their own. The way it will look, the brains that they don't want or the all this kind of stuff, we can do that for them. And then whatever they want to be doing or in a location or whatever comes after the fact. So it's a mix of both.
- [59:06]
A
Okay, yeah, I can imagine. I'm also curious. Okay, so we talked about like this world building. I could imagine you could design a prompt that is the visual style of the world, the types of objects in the world, the elements. And then am I wrong in thinking that if I take it, if I am a customer and generate an image, and then Patrick is a customer and generates an image, we're not in front of the same. It doesn't look like a static background, it's more dynamic. Like my image is going to look different than Patrick's. And so there is some like, randomness that you're getting from the GPT.
- [59:42]
C
That's the cool part, is that it's all unique. So everything that's being generated because a lot of these people display them, whether on huge slideshows or mosaics or everything, everything's unique. Now in certain cases, they do want it all the same so that they can. Certain brands are extremely. They're like, we can't have anything changed in this. We just really want people's faces and bodies to be changed. We have that option too. So it's flexible, but most people have it to where it's just unique every time.
- [60:09]
A
The dynamic piece is the piece that feels really compelling, but it also feels like that's where you have risk. Right. I understand this concept of like, we can give the reasoning agent a prop that sort of gives it rules, but it also feels like the universe is infinite here. Are there. I guess I'm just curious, am I making up a problem that doesn't exist or is there like work that has to be done to keep it within the boundaries?
- [60:32]
D
No, your judgment is spot on there. We've invested a lot in a pretty sophisticated meta prompting or meta analysis pipeline. So we do some pre processing before the images get sent to the. The final render, if you will, in order to look out for specific things that we again, through lots of trial and error, through lots of eval testing, have determined our. The core areas that we really want to hone in on. And there's some kind of unintuitive things too. For example, if somebody has a disability that might not be a hundred percent obvious from the video or from the photo, but there's a couple tells then we want to make sure that they're correctly represented in the final output so it's truly their likeness and not just some generic AI generation. So there's a couple of different pieces that we really look out for in order to make sure that all of that information is correctly compiled. And this also too is where it's really exciting to allow the users to to have more creative control over the output. Joe's working on some really amazing concepts around this that bring it to another level. But even at a basic point of through what somebody inputs in the form field, being able to ensure that helps impact the end result. And again really almost any uncanny like I can't believe that image represents that person way some of those subtle details that we can automatically capture from the pre processing and also from the user input really helps sell that final image, if that makes sense.
- [61:53]
A
There's so much you just said that I really want to dig into. So the non obvious disability piece really resonates with me. I have really poor vision and I have an eye that wanders. And when I first started doing video online people used to complain about it all the time. And I just feel like this is who I am. I can't do it. I can't do a thing about it. It's just the reality. You're going to get used to it. And if you were here with me in person, you'd see the exact same thing. It's just what it is. And it made me immediately think about skin tone emojis and how it seems like such a small little thing, but it's not a small little thing to the people that aren't represented by the default skin tone emoji. And so I love that this is something that's already come up and that you're already considering and you're already doing stuff with and the next piece that you brought up that I just got really excited about the potential is my understanding was like your little mad lib thing. You can answer a question. I thought you get emailed a picture but it sounds like you're working on or maybe even already have an even more interactive piece where the user can customize their image even further. And I think that just drills in even further on this. Like I think opportunities for people to participate with a brand where it's co creation is really powerful. I just feel like it unlocks I Personally hate most advertising. It feels to me like you're screaming at me. But I have brands that I love and to create these moments where you get to engage with them and create with them and participate with them both in person, in a physical world. Like your World cup example is so compelling because what a great example of brands people love, right? Their soccer team. I can just see so much potential in this. So it's such a fun problem space to be exploring.
- [63:31]
B
It is. I. I just sent my Rev ops director this afternoon, like, two screenshots. I was on LinkedIn and I was like, looking up the term experiential to see what pops up. Because this experiential marketing is like the business that we're in. And both Mattel and Hoka were both hiring directors of experiential. Literally within three hours of me looking, there were three posts of, like, job searches for this term. And there's experiential directors at canva. Cannes is happening soon, right? The film festival that's like experiential central. Like, what they do at Cannes Film Festival is looked at by so many other people. Why? Because brands are. There are investing an insane amount of resources into somehow getting you to feel something, do something, versus just being, like, shown something.
- [64:19]
A
Right?
- [64:19]
B
And that's the difference between, like, traditional marketing, the way we would define it, and experiential marketing is like the difference between a brand saying something to you and then inviting you into some type of experience. Even if it's like a minimal interaction, it's better, it's going to be more impactful than you just like, seeing the upteenth ad that you've seen today across whatever platform you're on. There's ads, like, very rarely do you get to interact with a brand in that sense, right? Other than, like, the tools we use all day.
- [64:48]
A
I feel like it's such a great way to create brand evangelists. Like the other idea that came to me when I saw your site for the first time. I'm an author. I have. When I go to this conference next week, people are going to come up to me all day telling me about my book. Like, how fun would it be to let them take a picture with their book and enter a concept that resonated with them and be able to have a page of all these fans and what resonated with them about the book? It's. It lets them participate and be a part of it. Not me. Just be like, yeah, I'll sign your book. Like, I. I can just see so many Ways that an audience, a community, a group, and then they can connect with each other. If you have this place where people can go see, they can be like, oh, I want to go meet that other person that also is doing this habit the same way I do it. It's such a cool. I can see so much potential.
- [65:38]
B
You're reading into the future of what? We were having some discussions yesterday as a team.
- [65:42]
A
I'm going to be following along and I think it's only a matter of time before I run a campaign with you. And I'm not a BSer. I have genuine enthusiasm for what you're doing. We did not get to dive super deep in the technical parts. We are running out of time. Let me ask you this, is there anything that you really wanted to dig into that we didn't get a chance to.
- [66:06]
B
The technical guy is Patrick, so he would be the one to ask some last questions to if you really want to go. If we want to nerd out at
- [66:13]
A
the very end, I'm actually really happy with where we've been and how things have turned out. But I just want to give you. If there's something you really want to get into, I want to give you the opportunity.
- [66:22]
D
There's one piece I could talk about which is the orchestration layer that we follow, which is like how we get really good performance out of the LLMs from a coding perspective, if that would be interesting, I can do like a succinct one minute version.
- [66:32]
B
Okay.
- [66:32]
D
Yeah, yeah. One thing that has been like crazy exciting for us and a huge unlock. So we talked a lot about the AI technology that's in our product and that's that we're exposing more and more for our users to actually interact with to help build and create these experiences. Then of course, like the AI generation image and video piece. But what's actually been more of a focus for me over the past year especially has been the code gen side. So just to give some perspective, we've been heavy users of cloud code since February 25, the day it came out of 2025, and then Codex roughly I think a month and a half after that. So we've just been in the deep end in terms of agent orchestration and all of this, and I've spoken quite a bit on those topics through those talks. One of the frameworks that I just keep coming back to, that's I think the best way to articulate a lot of what we've discovered that helps us build products incredibly quickly with a very small bootstrap team Is the orchestration layer of allowing your agents to live in these. The four things that I'll lay out here to give some quick context though, I think it's so helpful to think of these cogent agents as a really smart college student. So you can imagine like a PhD that is just brilliant at with spiky intelligence. So they're brilliant at a couple of different things and. But they have zero context about your company, about the project that you're working on, about your customers, about like basically anything specific to you. They're great in the abstract about engineering and a bunch of other disciplines, but they need to understand like you would be onboarding a new employee. Okay, what do I do? Like why are we doing this? What are the resources? Who do I go talk to figure out this? Why are we doing it this way? Like why is the architecture built like this, where are servers, et cetera, et cetera. So in order to answer those questions for these AI agents, whether it's cloud code or Codex or anything else that you're using, we found that context. So being able to is really important. In addition to tools, verification and then workflows as well. So to quickly explain that context is a lot of what I was just asking or mentioning. So for example, I scraped our entire marketing site, converted it onto markdown and then created a directory. I've got a bunch of these, but this is just one example to where the agent knows via declaration in our cloud MD and agents MD files that it can go there to figure out. From a marketing perspective, how are we looked at? Like how are people perceiving us? How are customers? Like how are we speaking to our customers? And then we've got our ICPs. So who are our customers? What are they looking for? Going down the entire line, just like I would onboarding a new engineer and then we've got tools. So can these agents go. And this one's kind of crazy, but read all of our production software. Full access to Terraform from a read perspective, not from a write perspective. Full access to GCloud. A lot of our inferences on GCP, but we use basically all the providers. Can it go and gather all the context that it needs? Is all our telemetry and logging and alert infrastructure set up to where it can go in Query Sentry or Datadog and get all the information that it needs, et cetera, et cetera. So a lot of tools, mostly CLIs to be honest, but MCPS as well, in order to gather that and then verification is a huge thing that I'm always hampering on whenever I'm talking to people, which is, can these LLMs iterate against something to know if the job's done well or not? Just like a performance review or any sort of acceptance criteria for an engineer. Like how do they know whether what they're building is good and how did they know to keep iterating in a loop or however you have things structured until they get to that end result. So an obvious example.
- [70:12]
A
Yeah, I can see you're really passionate about this. I will share, I share this passion with you. I started using Claude code, I think in June of 2025. So a few months after it came out. But I dove deep. There's four Claude instances running right now as we talk building stuff for me. And everybody on the planet is talking about loops now. And I'm like, we've been using loops forever. Like, why is this a new trend? Of course you have to have it check its work, otherwise it's going to do terrible things. I am personally struggling. We are recording this on June 18th. I had three days of using Fable and it was like the most magical three days of my life. Here's how I describe Used to be when I first started using Claude code, I followed Andrew Karpoth, Andre Kaparthes 4 line diff guidance. It would write code, I'd make sure it looked good, it would write some code and make sure it looked good. Last November, that changed. Opus 4 came out, or 4 or 5, I can't remember which, 4 or 5. And I just started writing really detailed plans. And when the plan was good, I just never looked at the code again. I had a plan reviewer, I had a code reviewer, but it was all Opus, checking its own work. And I just looked at the plans. With Fable, what changed was I didn't even have to look at the plans. Like we would have a conversation, it would write the plan, I would look at it and I had zero changes. And after that happened six times, I was like, okay, we'll talk about the plan. And then you go, and I'm going to move on with the rest of my life. And my funny story about Fable is on. I live on the West Coast. So Friday at 3pm I upgraded from a hundred dollar max to two hundred max because I was out of Fable. I literally upgraded two hours before it was turned off. I check the news like every hour. Is it back? Is it back? Yeah, this is the area. I started building my first AI product a year ago March. And I'm not really an engineer. I'm more a product person. And I just, I jumped so far down the rabbit hole. All my product people are like, Teresa, what happened to you? You just talk about engineering these days, but I just love it and it's just so fun. And I think, I think this is everybody's future. Like, I am an engineer, but I don't write code. And I know longtime engineers that don't write code anymore. So one of the things that's really exciting to me is we all have the ability to build this way. And so I've been nerding out on the like personal productivity coding side because I feel like this is what's going to unlock it for non engineers. This is also what's driving my mom and tech mom and pop tech stuff. Like we're getting to a point. We had no code tools for a long time. They were kind of good. But now real people can build real software, which is just mind blowingly awesome. Patrick, I am responding to your passion with passion. We are way over time. I want to thank all three of you. I am genuinely a fan of what you're doing. I will continue to follow along. Like I said, as soon as I get some days to plan a fun campaign, I will be working with you at some point, I am sure. Okay, so for folks who want to learn more, snapbar.com, super fun website and then Patrick, I understand you are a YouTuber. So you took all your experience learning from YouTube and you're now giving back. Where can we Find you on YouTube?
- [73:28]
D
Just searching my name, Patrick Ellis. It's the best way to find me.
- [73:32]
A
I'm going to do that immediately after this episode and then Sam and Joe, anything else listeners should know.
- [73:39]
B
Now, if people want to talk experiential marketing. I don't know if that's your audience space, but I'm on LinkedIn. That's where people can find me. I'm not a YouTuber, just Samites on LinkedIn. All right, you have to go track him down in person because he's basically a ghost on the Internet.
- [73:54]
A
I love that too. Some days I wish I was a ghost on the Internet. I'm a little too public on the Internet. Excellent. This has been absolutely delightful and thank you all three of you for taking the time.
- [74:07]
D
Appreciate it. Thanks for.
- [74:08]
A
If you enjoyed this conversation, please subscribe in your favorite podcast app and give us a rating as it helps others find the show. Thanks, I appreciate it.