Transcript | How to build an ethical AI product | Marko Jak (Co-founder @ Secta Labs)
The transcript from my podcast with Marko Jak from Secta Labs
Marko Jak 0:00
You never know what's going to happen with AI next. I mean, I don't sleep at night sometimes because we don't know what's going to happen. So there's always new techniques that are coming out which require less and less data, less and less training, less and less of everything.
Thanks for reading Future of Product! Subscribe for free to receive new posts and support my work.
Max Matson 0:17
Welcome to the future of product podcast, where I max Matson interview founders and product leaders at the most exciting AI startups to give you an exclusive glimpse into the workflows, philosophies and product trends that are shaping the current and future AI landscape. This week, I sit down with Marco Jacques, co founder at the AI company that turns any image of you into a professional headshot, set the labs learn more about how he sees the current AI landscape, how we achieve viral success, and how he builds an AI product with responsibility at the forefront. With all that said, let's dive right in. Hey there, everybody. Welcome to another episode of feature product. today. I've got a very exciting guest, his name is Marco check. And he is the founder at sector labs. Very exciting AI startup. Marco, would you mind telling us a little bit more about sector and about yourself?
Marko Jak 1:03
Hey, Max, thanks for having me. Yeah. So I really enjoy talking about what we're doing, because it's kind of the simplest product to explain, which is you upload a bunch of pictures that we give you a lot of professional headshots powered by AI. So a little bit like about myself, I guess I've been building products for like, as long as I can recall. Many of them have failed, some has. I succeeded. And basically, we started the company, my co founder and I towards the end of last year, we'll give you a bit more about the backstory, I think, as we go forward. And yeah, we've we actually launched in the middle of January, or so I think we were the first ones to be doing it. And yeah, since then we've did we've done about 3.5 million headshots, and over like 12,000 users. So and that's all through basically word of mouth. So that's kind of just like holding on for dear life right now.
Max Matson 2:03
Yeah. Well, it's all 1000. That's your, you're doing all right, I'd say. Yeah. That's awesome. Would you mind getting a little bit into that background with sector? You know, what motivated you to actually solve this problem?
Marko Jak 2:18
Yeah, so actually, it's sort of evolved from experimentation. You know, different companies have different starting stories. Our one actually my co founder, my co founder, Alex, him, even I, at the end of last year, I sort of pitched him let's do the fun, like Avatar thing that everyone else is doing. And just like, have some fun with it. We thought we could do it in a unique way. So we did like a whatsapp bot. So you upload your photos, or like a whatsapp bot. And then the bot comes back to you with all kinds of pictures of yourself as like a superhero or anime character. And, of course, the downside of it is like, what do people do with this life? Cool. I look like a superhero what I do with it. And so instead of doing that, I showed the images, actually, to my wife, and she was her reaction was, what do I do with this? And we actually had like a fight at the one point and she wasn't speaking to me that day. But she actually sent me a text. She said, not that I'm speaking to you. But can you do this for a professional headshots? And I was like, yes, like exclamation marks. So it sort of clicked on the application being something much more kind of valuable to people because they can actually use it.
Yeah, so that's a bit of like the backstory. I mean, I've been just the back backstories. I've been building in AI since 2014. Actually, prior to the whole art transforming architecture. First startup was a text processing startup. And I recall, one of the first things I did in 2014 or so was make an article with like, a Shakespeare in this in the style of Shakespeare was really bad. But even then I saw, hey, there's something here. And towards the middle of last year, I think, beginning of last year, when open AIS, Dali came out, right, it was a closed source model. I was like, this is pretty cool. Right? You can do some things. Yeah. And so I think just by staying in this space, by playing in the space, this idea, like, evolved out of that, so it's something I would kind of recommend to anyone, if you're interested in this space, just start doing something in that space.
Max Matson 4:26
Yeah. 100% I love that. I mean, it's like, yeah, it sounds like you got in very early relative to you know, definitely, I'd say the vast majority of people today. But kind of having that background and being able to see the trends right. So I love that kind of anecdote that you give when it comes to Sector of having this thing that's, you know, fun and nice to have, but how do you make it actually solve a problem that people are willing to pay for, you know, with? regularity, right. Yeah. Kind of on that that train of thoughts. How do you see you know, a tool like sector A shaping the future of you know, not just generative AI, but the way that people actually work and go about their professional lives.
Marko Jak 5:10
Yeah, I mean, just to add on to the other point about playing with things, so like, I just want to add something that it's pretty cool. Like, when you start when you see something happening, it's generally there's this great thing by Chris Dixon from Andreessen Horowitz. Right, he says, the next big thing will start off, start off kind of looking like a toy. And that's really like, so if you ever, you know, if you're sort of exposed to what's happening in tech, and you see something like, this is pretty bad, it can't do what it can do. But if you see something and you say, well, it can do a little bit more than I expected it to be able to do, it's able to make a picture, it's a really bad picture. But it's able to do something and I can sort of see your hand, I can sort of see something you can like imagine with the evolution of things where it could go. And that's really how you get early in on these things. So how do we fit into the landscape of generative AI? Well, I would say and how does it kind of shape things? That's what you want to know. Right? So you see, like, what you'll see straightaway, I guess, is the landscape is sort of split right now into the different types of content, I would say.
So yeah, audio, video, image, it's all falling under generative AI. And each of these again have like a cascade of applications. So even during of AI for the images, which is the space that you could say we are in as it is now, right? The applications, there could be in a professional sense, like what we're doing now. But there's applications in the law, generating digital art, or, you know, I just actually advised a startup that is doing the same thing as we are doing but for 10 babies, so see your baby in like a Darth Vader outfit, and then, you know, pretty cool. Again, that's it, it will shape all of these spaces, if it adds value to the customer. How do we see it kind of shaping work? I think right now with the product as it is, I would say the first thing that we're doing is we're trying to get you a professional image of yourself where it looks like you've been on like a six month holiday you've been eating? Well, you've been exercising, like, that's what we are aiming for with the tech is like, authentically you. And so if you abstract that you're saying what we're trying to do is we're trying to help people portray them in themselves, sorry, in the most accurate and aesthetically accurate way in the most accurate and aesthetically pleasing way.
Max Matson 7:49
Yeah, the best
Marko Jak 7:51
Yeah, I don't know if I answered your question.
Max Matson 7:54
No, yeah, absolutely. Yeah. So I would like to just dive in a little bit more. So one thing that, and honestly, part of the reason I was very interested in talking to you is because this is a problem that I myself have have dealt with, right? Where you're usually when you're searching for a job, maybe it's just me, but at least in the last few years with market conditions, it can be you know, very last minute, you're not prepared. And the last thing that you're thinking about is going to get some professional headshots taken. Right. Exactly. Yeah. Yeah. So I've definitely been in that position where it's like, how do I figure out how to make this selfie look, you know, not not like hot garbage. And so I was just thinking, you know, with your technology, with tools like sector, I wonder if it's going to democratize in a lot of ways, the kind of lens of professionalism, right, like our ability to project that we're professional members of the business community. What do you what do you think about kind of that thought?
Marko Jak 8:52
Yeah, that's, I mean, that's really, really good. I mean, you could also add on top of that, and say, under tools that are helping you rewrite your profile, or even like the aims, right, give me a better kind of introduction about myself, have a look at my work experience and do that. So I would say it definitely contributes there. One of the things that I didn't think about that much at the beginning, but I've seen from people just coming back to us is like, what's in Thank you for seeing me as I kind of see myself, and I never expected that. So there's this ability. Our tool has the ability to give you a boost of self confidence, or to make you feel really bad about yourself. uploads are bad and the quality is bad.
You're like, Oh, my God, this is so bad. Of course, most people don't feel that bad. They just go Oh, my word, this thing screwed up. And yeah, so I think it's not just that it allows you to present that image which to your point, all of these tools are going to be doing that helping you present a more professional image of yourself. But in our case, we actually have a sort of big responsibility and a privilege to show you something that helps you feel good like about yourself.
And if you sort of you sort of touched on actually layoffs, I just want to emphasize something they will we, if you have been laid off, come to us, we'll give you like a free headshot of what we've done. For 200. People, I just feel it's so cool that we can actually help someone in some small way. I didn't think I could do that, like in someone in some small way, that maybe it increases the odds of you getting that call by like, 5% 20%. I don't know, you know, and if we can do that, that's actually great. So just putting that out there. Like 5000 People coming tomorrow, but if someone's ought to hear everything good, yeah.
Max Matson 10:46
No, absolutely. No, I, for everybody in the audience. Yeah. If you've been impacted by layoffs, definitely let them know. Yeah, that's awesome. Yeah. Awesome. So kind of moving on from there. Would you mind just touching briefly on how the technology works? You know, I think my audience is decently familiar with generative AI, on a broad level, right. But you, could you tell me kind of a little bit of what's behind the magic?
Marko Jak 11:15
Yeah. How, how technical Do you want me to go? Maybe, maybe I'll start off on like a broad explanation, anyone who isn't. And then if you feel one of the areas want to dig more into, we can get into that. So I could probably best explain how it works in kinda what's in three parts. So the first part is, you need to upload some pictures. So we need what we call the kind of training data. That's a data acquisition step, as we call it, the intended it basically just means upload your pictures. And the key part here, I can go into more detail, of course, is just like you basically trying to teach the AI what your face looks like. So a bunch of photos with you wearing sunglasses, that I can't guess what your eyes look like.
So there's, there's good data, bad data, and there's some uploads, which sort of fall in the middle ground where we do train on them. But we have assumptions internally that these aren't that good. So that's an important step, because you're giving the general model and so I'll go into that part as well. Now, the key part of the tech is the second part, which is the actual fine tuning. That's what that's called. So if anyone wants to Google that the fine tuning of LVM, that's what we're doing. And really, I'll try and explain this. While I was thinking about this, like, if I get asked this question, how do I explain it in a way that anyone can understand this, I'm gonna give it a go. And you tell me if it actually makes sense. So the diffusion model is an open source model. We'll talk a little bit about that. But it's basically a repository of many, many pictures that this model has been kind of trained on right now with us with our the process, what we're doing, you can think of the model as a natural painter. And that painter has been trained to paint all kinds of landscapes, right, or all kinds of pictures. And he's been trained to do that based on looking at 1000s of pictures.
And now we want the painter to paint only the Golden Gate Bridge, or only the Grand Canyon, right, so a very specific kind of landscape. And this is where a fine tuning comes in. Right. So we're adjusting the painters technique based on the few pictures that we have of the Golden Gate Bridge, or like the Grand Canyon. And we say we're trying to get the painter to adjust the process. And that's where the technical terminology comes in which I can get that into, but close trying to match what you upload. So it can resemble that. In our case, of course, we're not painting a bridge with painting you. And so that's the second part of the tech. And the third part is really once we generate the picture. So once we've taught the model, how to paint you, then it will will sort of prompt the model in extremely similar way to how people are prompting open AI, you know, you can download all the prompts, guides all of that stuff, right?
We prompted give me a picture of a man with this background. And you know that and so once we get those pictures, we have a unique process. That's very different to anyone else. We spent some time a lot of time, a lot of credits, and some money muncher basically complicated building what we call a aesthetic score. So we score the images which come out of it. So at the top of your gallery, the first image that you get won't be a hand coming out of your head because the score it says that's not a good picture. We're getting better at that. And that's actually a very key part of the kind of process. from a product perspective. It sort of makes the machine learning more robust Almost, if that makes sense, because we can afford to make mistakes in other areas, or it's not as good. But we Robusta phi, and it's actually some old tech in there. So that part has a bit of AI now, but when we started it, it was all about old tech. Okay. And yeah, so does that give you a broad overview of it?
Max Matson 15:20
Yeah, absolutely. Yeah. No. So it sounds like the fusion is is a big core of it. Right. And that's kind of the shot. Yeah, the first part, and then you've got kind of this proprietary tech, that's, that's really doing the training. It sounds like, yeah. Okay, perfect.
Marko Jak 15:34
Yeah. So so the, the fine tuning or the kind of training is proprietary, as well, the initial technology that we used was on an open source, stable diffusion model, but we've trained our own model on top of that, got it. So now, then we fine tune that model on your data. So there's basically like, models all the way down kind of thing, you know.
Max Matson 15:59
Okay, gotcha. Ai inside of AI. Yeah. Awesome. I think that was very well explained that, that I think they that people should understand that. Awesome. So
Marko Jak 16:11
cool. Thanks. I hope so.
Max Matson 16:14
So, moving forward, how do you guys use AI, you know, in your own internal processes, it's something that I like to talk about quite a bit, because I, you know, I would consider myself a bit of an early adopter, when it comes to AI tooling, just in terms of especially like day to day tasks, right? Automating the small stuff is a big thing that we focus on here. How do you get?
Marko Jak 16:39
Well, I'd say that there's like, three things we're doing that probably some people are doing a similar kind of versions of, but the third thing is actually more than I think we're others are doing. So like, let's start with the the first thing we're doing is we are kind of using it on all of our code reviews. So as we do code, I can talk a bit about that. The second thing we're doing is just, I mean, I use it to write SQL queries, or kind of regex just prompted like, here's all my columns, like, here's my table, give me a query that like returns that it's excellent at that. And the what's the third application is I actually use it to do a whole bunch of market research. And in doing that, I actually run an auto kind of GPT. So I don't do a single prompt, I give it a task, and I actually allow it to run through all the tasks. I can talk about any one of those but yeah, that's kind Yeah.
Max Matson 17:33
Yeah, I mean, let's break that down a little bit more. I, I'm especially interested in kind of the market research aspect. You know, kind of like, from your position as a founder, what does that look like?
Marko Jak 17:45
Yeah, so it'll be something like, one of the tasks I gave it at the one point was identify the top five markets when it comes to portrait photography, by size, and come back to me with the with the kind of sentiment in each of these kinds of markets who are the leaders. So it's like a whole task that you would probably give an intern and spend like five days on it. And there's a different there's different tools out there. So it's a friend of mine has a tool called prognosis, which is available artists, I'm giving him like a shot, and you should give me a, like an affiliate link with somebody. They trying to build it out as a product, which you can buy. So I would say if you want to try that sort of thing, you can just go there. But there's a whole bunch of open source code out there, which you can do it with as well. I'm running it in my own, you know, open source tool, but that tends to break and tends to get all kinds of problems. So I might just end up with paying for something.
Max Matson 18:46
Gotcha, gotcha. So so just to
Marko Jak 18:49
unpack that a bit more. So if it's unclear, like the task, what actually happens is just like you would prompt, open a eyes, what's a GPT to do something, the task, what it does, it generates its own prompts. So it generates let's say 5100 prompts, and it recreates prompts as it actually goes. So it's actually feeding itself prompts in a recursive manner. So behind the auto process is sort of like what are the prompts I need to generate in order to answer this prompt or this task. So it's prompt all the way down. But instead of you prompting it every time you just give it a task, it does like the five minute prompts and then he carries on for one hour. And then it gives you the the output at the end of that. I think that kind of AI you know, the agent is going to be big. That's going to be a really big thing from like a lawyer to an intern to all kinds of tasks that you can do like a first pass on.
Max Matson 19:48
You said earlier that sector is all word of mouth.
Marko Jak 19:52
Yeah. We actually that has changed in the last week and a half we did a influencer campaign as well with Um, a place called Putin job seeking is hard. And that's like a online community basically helping people get jobs, or how to prepare for jobs job interview, it's got a whole bunch of great kind of content on there. So that was our first we've done a bit of paid just like experimentation. Yeah, so we haven't done anything paid. It's just all people telling other people. And the great part about that is I also get emails now. And again, like, I've tried everything else out there. And now we're coming to you, and you are the best. So that's actually nice to hear that we should probably do a better job of speaking about that. But we just really focused on the product and taken sometimes doing not as good a job on the kind of marketing side, like, you know, we're not we want to build the technology more than we want to make more money.
Max Matson 20:53
That's what you want to hear. Right? I think that's the the core ethos of product lead.
Marko Jak 20:56
Yeah. Honestly, I feel like a memorable experience is the most difficult thing to create. And that speaks for itself a lot over time. And yeah,
Max Matson 21:07
no, absolutely. So just to kind of, you know, go on that topic a little bit more. Is there anything kind of inherent to the way that you built the product that has led to this viral success that you've seen?
Marko Jak 21:24
Yeah, so I would say the first part is that we, when you get your gallery, we do, we, the first goal that we have is that you get your gallery, and you go, Oh, my God, this is great, right, that's like the first goal that we need to get to. And then we've unlocked a few other things. So we have a thing where you can create a new kind of custom shoot, and you can choose, I want a smiling photo with me in like a suit. In this background, we have all kinds of back control over it, we actually going to be kind of shipping a v2 of that, where you can retouch an image. So you can say actually do this on this image. And, and so to answer the actual question that you're asking to get more access to that actual feature, we ask people to kind of share. And that's just to ensure that when they continue to generate hundreds of images, we sort of covered the cost by asking them to kind of share some stuff we do on referrals and discounts as well.
So we you know, people have great experience, just doing it kind of manually now, like you got a great experience. He has some codes like, hey, but it is very much I would admit in your question was a good one. Because you said is there anything we doing to sort of amplify that, but a big part of anything AI right now is driven by the remarkable aspect of it like, oh, wow, I didn't know I got like, uploaded 25 pictures, and I got these hundreds of pictures back. And so we I'm sort of very conscious of that, like, that might be great a bit over time. But I see, I'm not playing the kind of short game here for generate some pictures, and then lo Off I go somewhere else. For us. This is like a long game.
Max Matson 23:13
Right? I love that. It's, you know, first and foremost providing that excellent experience. And then really just, it sounds like I mean, you and a lot of the successful product lead companies that I've talked to, it's just doubling down from there, right? It's expecting that if somebody truly got value from you that they'll be willing to share. And it's amazing to see how often that bears out.
Marko Jak 23:36
Yeah, yeah, actually, one of the things we sort of guide on the I wanted to the sort of principles I try to adhere to on the pm side is, if I asked you to do something, I should give you something Why should explain why I'm asking you like, I think the mature move sometimes is a little bit like you expect people that they will do things without thinking like what are they actually they have motivations for doing that. So sort of a level of skepticism is needed whenever you think of any kind of feature, because what you got to ask yourself question, why would they do this? And if you don't know the mind of the customer, then you will make up a story. No, it's you know, it's easy to click on or something and it is just that isn't going to happen?
Max Matson 24:24
Totally. How did you know what was some of the early work that you all did to really get to know your customer to be able to proactively solve that need?
Marko Jak 24:33
I have to give my co founder Alex a whole bunch of credit. Yeah, because he sort of he sort of, I wouldn't say he forced me but he was like, okay, you know, we were growing so fast. We couldn't keep up. We had no actual products. So I mean, we had the MO but we didn't have a product kind of workflow. We were actually when we went live we had upload form. Like you could just upload your images to like a note Code form, we didn't have anything. And so at the beginning, and this was the best thing I did is I reviewed every single order that came in, I looked at every upload, I looked at, I took out the bad images, I tried to keep good images.
And I mean, people knew that we were doing this because I was like, sorry, please wait, this will take about two days to come back to you. Because we have to review all of them. And that was the best thing I did. Because I saw what people are giving us. I saw what they thought, you know, the asking for I heard about their problems. And I think it's helped me and I still do it. So I still email customers every day. And I don't come back to all of them sometimes. But sometimes, because they email back and some people like actually, I need a support query, then I fought with him that on there, but always help them out. So that's actually something I think everyone should always be doing. Like you don't have to have like a mental model in your head of the needs of the customer. Otherwise, you're not really going to build anything. That's good.
Max Matson 25:59
No, that's exactly right. That's exactly right. I think that's the point that kind of gets lost in like the hype cycle when it comes to AI. Right, is that? Yeah, so that you have to be able to do the process, understand the process, understand all of the pain points associated with the process in order to build something that's going to automate it. Right. If you just jump straight into that part, you're gonna miss something.
Marko Jak 26:19
Yeah. So what have you seen with that, as well also, like people who are successful at doing that, that kind of approach?
Max Matson 26:25
Yeah, no, absolutely. I mean, it's something that I actually talked recently to our head of product, Matt about is something that I really admire in him is that he's willing to do something that sucks just long enough to be able to figure out how to automate it. Right?
Marko Jak 26:39
There you go. That's it. Yeah, exactly. That's a very good way to actually say it, right, you have to be willing to do something that actually sucks so that you know what's actually going on there. It's very easy as a co founder to get caught up in the more like glamorous stuff, like, I want to go to a conference, and they want me to speak there and all of that, but then you can sort of lose contact with who you actually are trying to build for. I've actually seen that happen a couple of times. So yeah. I'm trying to be cautious that I don't end up there.
Max Matson 27:09
Yeah, absolutely. Well, sounds like you guys are on the right track. So, you know, kind of going alongside that, what are some of the, you know, unique challenges that you have faced, you know, kind of building an AI product during this time? Getting it to market?
Marko Jak 27:25
I think, you know, I think we had we were we had at the right kind of timing, right? So we would like to first to come out with this unique the positioning of the AI avatars within this kind of space. The AI avatars we've done since I think September last year. And and and you know, so like, we in that way, we got the, you know, kind of lucky, I would say although we saw the opportunity, and we grabbed it right? So what's the opportunity? And actually, like, you know, those two things, but unique kind of challenges? Definitely, I would say, we sort of had challenges on the ML, we had challenges on the product.
And we had challenges on the kind of marketing side I'm my guess is you want to hear more about the Machine Learning and the kind of what separates product side of things. Yeah. So the main challenge on the product that we had, and we still have it's an ongoing challenge is, how do we help you users upload the right images. Now, you can take an arrogant approach and just tell people, This is what you should do. And if you don't do it, we don't give you back your money kind of thing. Our approach is more like if someone is not uploading the correct images, we failed in onboarding them. And that's not always the case. Of course, we don't some people just don't read at all and just upload anything. But more and more.
It's like how do we nudge you at the right moment, give you the right context, give you the feedback on the picture. So if you do upload images now where like you were in sunglasses, we tell you Sorry, we can't accept this image. So that's, that's an ongoing kind of challenge. They the hygiene is an ongoing challenge, because we can manually review every order which comes in I mean, at this point, it's just way too much more than any team could even actually review. And the other part, I think this is a challenge that maybe others don't know much about, which if you might want to dig into a bit more because it's really it's not emphasized in the space, but there is a bias problem in the underlying data that our technology is actually built on. So stable. Diffusion 1.5 2.1. I mean, it's all based on like the lie on five b two B datasets. These are open source. You know, they open source pictures, and but there's an inherent ethnic bias in there. And what that means is that if you train on those actual datasets or on the models, which are built on top of them, if you're, let's say North Indian, South Indian, or you know, the Asian or Hispanic, Asian, of course, it's very broad, you know, but you may have to, it will be more difficult to get a good outcome. And that's something that I'll be honest, I've never had to deal with that.
Because when you bought products, please do kind of stop me from going off here. But I care about this part quite a quite a quite a bit. Like, I bought products in the past, right. And one of the things that you need to learn as a pm or building products is like, who isn't this for? Who is this for? And who isn't this for right? And it's okay to say, this is for stay at home moms, and I'm not building for anyone else. And that's fine, because you have to kind of focus, but it's not okay to say this is for people of a certain skin color. I mean, that's, that's the definition of actual racism, right? So it's like good for you, but not good for you. So that's actually, I would say, like rank order the challenges, that's actually the first challenge. And that's why we are the ones who are collecting ethnic data. So when you upload your lead information, we ask you, you can skip the question. But that's to improve your outcome and to improve the outcome for everyone else.
So that's like, one of my top priorities is how do we adjust the base models so that anyone of any skin colour any ethnic background can have as good an experience? And that's something, don't they? I mean, there's a bias question which people are talking about now, right? Like, I don't know, if you've seen those things where like, if you prompt the models for an image of a CEO, it's always a man or something, because everything's biased. But this is bad bias. And that's one of the challenges that I'm very, very active about. And then the third challenge, I've sort of gotten into like a bit. The third challenge is the aesthetic scoring of images, what is a good image? And what isn't a good image balance? sort of let me stop here. And if you want to dig into any one of these, or we can move on?
Max Matson 32:08
How do you kind of I understand like, the the self reporting aspect kind of helps you be able to provide a better experience. But is there a fundamental answer, you know, to a problem like this, when it comes to what the models are actually trained on?
Marko Jak 32:22
Yeah, so the most kind loads of fundamental answer is making the data set that the initial models are trained on more diverse. So by more like diverse, I mean, you know, there is no one, you know, black person or one white person, as you know, it's all spectrum, right? So someone from Kenya may have a different appearance in general, or they have different kind of what's in facial characteristics to someone from South Africa as I grew up, so I'm just like, giving names now. Right. And so having more of those images trained in would already be like a big help. But the problem is that whoever's working on open source models, they're working on general, all purpose models. So they want art, and they want landscapes, and they want people and they want all kinds of things. As I would say, we're the only ones that I know of, please keep it like a secret. And we're working on this sort of we pilots of fine tuning our own models, so that it can be can represent a group of people. And and I think that's quite key. I mean, especially to developers tech.
Yeah. But yeah, I would just kind of let all of us to say like, you never know what's going to happen with AI next. I mean, I don't sleep at night sometimes, because we don't know what's gonna happen. So there's always new techniques that are coming out, which require less and less data, less and less training, less and less of everything. And so, yeah, it could all be. So for example, there was a product called, there was a technique called control net that came out, which helps you control condition the images, so put the person here in their face at the angle and does that prior to control it there were companies I know who were spending 1000s of dollars every day trying to create their own, like approximation of that. And then our control net came out it just deprecated all of that stuff.
Max Matson 34:35
Right, yeah, things change so quickly. I mean, you're already hearing about kind of the the leagues of magnitude smaller data that you're able to train models on. So that could potentially be one answer, right. And I didn't mean to cut you off earlier. I know that you were kind of getting to another challenge or ethical consideration that you guys face. Would you mind going into that a little deeper? I think what we were just talking about kind of Naturally dovetails into it. And then we can kind of get into a talk about, like the broader, bigger picture for AI.
Marko Jak 35:07
Yeah. So what do you want to know about the ethical part? Because that's like such a broad, totally kind of topic, maybe let's try to like narrow that down a bit.
Max Matson 35:15
Yeah. So what I'm thinking is, you know, a user data, right, I think that's kind of a hot button issue be the fact that, you know, it's it's people's faces and images of them. Yeah. I think we could just start there. Probably, there's, there's a lot to unpack there, I would imagine.
Marko Jak 35:34
Yeah. So I mean, you wonder, but it's, it's a really good thing to bring up. Because it's also like to unpack to add to your question, there's all these questions about now, like, you have a Moloch, engineering a picture of me. And like, possibly Am I seen that picture of me, right? Like, in the past, if someone took a photo of you, you were a participant in some way, because you were there, right. And so, I mean, this has been like top of mind for us, when we started off and it at every point, we sort of like, informing people as to what's happening with all of their data. And I mean, just to be actually clear, is that people can always ask and say, take out my data. So we actually have I just had an email that I saw as part of the kind of what are the so you know, daily, we have a deletion sheet, these are the accounts that are going to be deleted, and that only happens every 24 hours, like a batch run. And so that's always an option. Just to be clear, right? People can request all their data to be gone. The other thing is that we see the AI model as being owned by the customer. So like, I see it as like, we don't own it, you can generate offered, we help you create images out of it.
But I can't create an image out of it to use for look forward to something else. So So why our gallery on our site sucks, because I need to ask everyone, Hey, can I use this picture of you? Can you sign it off? So we should probably like automate that process like he has a $50 battery. If we can use all your pictures, please, please that kind of sign here. So anyone who's had their headshots in this hearing this please just email me and telling me I can use your pictures would be great. But we Yeah, so we need to ask as well, the way I see it, like we don't have any claim over any of your pictures. We're enabling you to generate, like audio into the model. So I guess that's one key thing I would say.
And this actually brings up another ethical question, which is for teams. The way I think that other companies are doing any team kind of headshots, because companies also have this problem, right? I have 20 people on my team that based all over the world, how do I coordinators and the other startups, I think that might be doing this for teams, they just sort of allow you to get all the pictures. To me that feels weird. Like, it doesn't feel right, that the company can just get all your pictures like to me it shouldn't be a process of the company is paying for it. But by default, they don't see any of your pictures, you have to choose which ones that company can actually see. So there's an opt in or opt out process. And that that is what we do for all of our teams. So you can invite people on your team, they upload their pictures, but only when they save the ones in there on lob on branch shoot to those pictures appear on the company dashboard. Got it? So that's a bit of an ethical one. We can always unpack some more. Yeah,
Max Matson 38:31
yeah, no, I honestly, I'm very impressed by your approach here. Right? Both with the, you know, honestly, I mean, I've talked to a lot of people and a lot of people in the space have kind of given me their opinions on ways to make AI, you know, more equitable and more ethical. And it sounds like over at sector, you guys are really doing the work to make sure that that happens.
Marko Jak 38:55
Yeah, yeah. Do you think that I'm actually saying I can give you a kind of demo? Yeah.
Max Matson 39:03
You have the proof. But I'd love to actually get your opinions on some of the bigger ethical considerations in the industry at large, because I like the way that you look at these problems. So you know, when it comes to generative AI and ownership, right, I think that the way that you guys do it is really smart, which is to say that the end user is effectively the owner, right. What about these models that are trained on publicly available data? What are your thoughts there? Do Is there any obligation to kind of the content that, you know, was used in the training? What are your thoughts?
Marko Jak 39:41
Yeah, I mean, I think that's a really difficult one, right? I don't know anyone who can answer that in an authentic way without making some sort of a political answer to that kind of question. I'd say it's a, it's a key question. And when you start asking that question, if you actually abstract into sort of answering your asking the question of what is a derivative kind of work, right? Like you say, if I make something that's similar to what you've made? Is that, is that yours? Or is it mine? Right? And so, YouTubers figure that out a bit, I think with like, you know, you can make a video of someone else's video, like I react to someone's video, right? You can do those ones, right? And that seems to be fine, because but you can play back parts of a movie, but you can't play back the entire movie, right? So you can react to that.
So I think it's like, we'll see that evolving in this space where certain outputs are very luck, recognizable, or you can see it's actually trained on this kind of data. But the we're all doing a derivative works. If, when it whatever we do, right, every you know, they say great artists, with him steal, right, that's like everyone's copying from everyone else. AI is accelerating that, of course, and it's making it easier for you to actually copy. So I'm giving you a very hand wavy answer to some extent. But I would say that the way I see it is that there is a question there. The the line between what is copyright, what is owned by WHO, and what is a derivative work of that is going to have to be made more clear.
That's the people who are building these huge the open source models, I guess they're going to be challenged more and more on that. Our perspective on it. And this is a bit revealing where we go in but our perspective, is we actually working with our photographers, I haven't said too much about it. And I'm not going to say too much about it now as to what we're doing. But I believe. So I genuinely believe this having both companies in the past were sort of the bullet companies here, sprinkled technology, and we don't need anyone, I just take, like, I find like, if there's a way we can take photographers along on the kind of journey with us, I would like to do it. At the beginning. I didn't have the answer to that. So if you asked me at the beginning, I would have said I don't know. But now we do. And so call me again in like a month and you'll see.
Max Matson 42:08
Ya know, we'll get an update out. I'm excited. I'm excited to see what's to come. Yeah.
Marko Jak 42:13
Yeah, I'll send you a patient update.
Max Matson 42:15
Yep, please do. Very cool. So kind of shifting back to two sector. It sounds like you guys have got some stuff in the works. I've been kind of amazed by the pace at which things have been changing and happening and growing. And if you look at the tech giants here, it's pretty nuts. The you know how quickly it's insane. I mean, you got, you know, a new announcement every day it feels like so how do you all kind of stay ahead of the curve when it comes to, you know, as a small business, staying competitive and truly making your product the one that that users want?
Marko Jak 42:52
Yeah, I just don't sleep on that. I actually, the truth is I go into Twitter every night at about 11 o'clock depends when I'm done with all the things I want to do. And then I just scroll like I have different AI, a Twitter list, right where you add people and I've built an entire bunch. And so I just scrolled through those and do scrolling. And sometimes I see something like Oh, my God, and then I don't see for four hours trying to implement it in the code or something. I try to stay away from doing that. Because my experience is that the academic papers that are published are not as good as they think, as they say they are, it's never as good. So that being said, so how do we stay ahead? I think the we look at everything through the lens of how will it actually help the customer. So you have to work your way backwards from there always.
And sometimes you're not sure, but it seems cool. And that's also fine, right? Like, sometimes you should allow yourself that play. And so we have a prompt playground, as we call it internally, which is like a really bad, ugly UI backend interface that you can just take 500 lock boxes, and you can try stuff on. But it's really key to be able to do that. Because you might discover something, you go wow, that actually improves the quality. So that's the first thing. And the second thing is I'd say be careful about things that are just like an academic paper. Because if there's no code to back it up, you know, might be kind of promising it could be it could be like a rabbit hole. So sort of in the rank ordered list of for our spaces like academic paper only. Okay, code and academic paper. Now that's interesting. And then maybe implemented within like automatic 1111 or something else like that's even more interesting because you can play with it some more. So you just got to find the right points of entry depending on what your customer wants. So again, always coming, coming back to them.
Max Matson 45:07
So you mentioned the Doom scrolling, which I find really, really funny because I mean, I find myself doing the exact same thing. But yeah, when you got into this industry, did you say 2014? I believe?
Marko Jak 45:20
Yeah. 2014 is when I decided to start a startup, and we'll do some AI 2014 will come, let's say, end of 2014.
Max Matson 45:28
Could you have ever seen any of this, you know transpiring? And the amount of time that it has, from that vantage point?
Marko Jak 45:36
That's a That's a good one. I think anyone who says, Yeah, I saw it all coming, like he's a liar. I would say that there was definitely when backprop came out 2020 12 I think it was, like, wow, was it pretty interesting what it can do. But I think no one including me thought that by continuing to build on a certain, like more data more compute would get you as far as we are now. And it was sort of everyone had these ideas that you have to build a brain or consciousness or like imagination, how do we train like imagination into an AI?
And so that I think caught everyone, including some of the greatest minds in AI, and a Facebook and other companies and a way of well, like, just smaller parameters, more tokens, more computers, and we are where we are now. And then the the, the secondary nature of that now is people are scared that while just more of that means that AGI is coming. And so that doesn't mean that that's the case, of course. But like it is scary to think that because while we didn't think it would get this far, maybe I can still get even further with just more data and more of what we're doing.
Max Matson 46:54
Yeah, absolutely. i That's the up at nine part. Right. I I definitely. It's an interesting spot to be in to to be, you know, at the head of an AI company at this time, because I do feel like everybody that I talked to, you think that we'd be complete optimists. Right. But I do think that everybody in this industry, for the most part is pretty level headed and pretty realistic about both the positives and the negatives.
Marko Jak 47:23
Yeah, yeah, I think so. I mean, I haven't actually met everyone. And there's some people who are doing some weird stuff with alliums. But ultimately, like, everyone knows that it's their skin in the game to some extent. So I'm actually lucky in that regenerating pictures. It's not like I'm trying to teach the AI How to open a bank account, right? So so I'm lucky in that way, in that it's, we won't have that sort of problem.
But the same technology that is used for our VMs is used for alliums, like the same kind of fine tuning methods, of course, with a whole bunch of variation, but it's basically matrices all the way down kind of thing, you know, so. So definitely very conscious of it that as well. And yeah, to your point, like, you can scroll and then you're like, oh, Google's kind of midpoint is now beating all the doctors out there. And this is beating this. So it does sort of bring up the question of like, well, what is what's gonna happen to earnings jobs? What's gonna happen there? Right
Max Matson 48:22
now? Absolutely. It's a question that I think we're all hoping has a good answer, but we shall see. Right? Yeah. Okay, so kind of, lastly, I'd love to get your kind of advice for product managers, for entrepreneurs who, you know, are seeing everything happening in the AI space and know that it's crucial to build into their products. What advice would you have for them, when it comes to actually building it in?
Marko Jak 48:51
Yeah, that's a good one. So if you already have an existing product, that means and you have some users who is paying you, that's already a very good start, right? Then you've got a feel of what problem you're trying to solve. And so again, start with the problem, don't get carried away with like, the shiny tech. We've all been there. I've done both companies based on that. And so start with the problem and think about, well, what are the gaps in the what the customer is trying to achieve?
Like, what outcome? Are they trying to get to? Why are they motivated to do it? And what are the gaps? And of course, you sort of like doing two things at the same time, you have some knowledge of what the AI can do, and you have this gap and you sort of moving these two planes, and sort of like, you know, 3d space of these two planes, and each time you move one, the other one needs to be moved as well, almost. And so I would say that would be of course, the first place to start.
But if you find that there's an application and you say I want to implement AI then Definitely, if you already have an existing product, you have a team, then I would say, definitely look for an API, like, look for something out there that someone's already built. And even if you pay like $5, a call or something, you know, just get an API that you can plug and play. And you can try it out. Because the first thing that you'll find, again, this advice may not be applicable to all companies. But the first thing that you'll find is that it doesn't work. In all cases, it doesn't work for all customers. Even our what's a fine tuning process, we have to personalize for every single user, because it's not the same parameters. And so definitely see it well. These are the problems but let me try to solve them with the you know that.
So that would be a great place to start to see if you can solve the actual problem without caring about how to implement it, or how to productionize it or put the ML infrastructure up into many people jump to we should be doing something with AI. And then they buy a whole bunch of service maybe and then they try and clone some code and try implemented like, really? Yeah. So that would be the first place to start. And then yeah, and then of course, I guess the kind of rest of the answer follows from there, right? If you see how much value how much uplift that brings the actual customer, you can decide how much you want to invest in that feature being a bolt verse by right, we just want to keep on actually buying it, do we want to build it. And you want to create as much value for the customer. However, all of that being said, there's also the argument that customers, depending on where you are there might be asking for AI.
So you know, people might I see companies like, we have a brand new feature, it's got AI in it, but like the features not doing anything. So they want to show many companies want to show that they're implementing AI because of the kind of hype curve that everyone's on. And I think you should just be careful with that. Like, it's okay to say, It's okay to say we should explore it. But don't over invest in implementing your product was not solving a problem. And maybe it doesn't have to solve a problem now, but you have to have a good hypothesis about what it actually enables your customer to do. So I can understand the sort of pressure to just do it. But that's inevitably isn't gonna work if no one clicks on it. Not so close it. Yeah.
Max Matson 52:33
Yeah, it's kind of like what we talked about earlier, you have to have an actual use case. Right? And you have to have that use case proven out before you, you know, jump to automating it.
Marko Jak 52:41
Yeah, I mean, it is. What's complicated, because also like, I would also encourage at the same time, right? That's why they say product management is actually hard. You should listen to customers, except when you don't know you should build new features, except when you don't. Like it's all yes. And no at the same time. So if you have the capacity in your team, and you have even a weak hypothesis, start to play with it. Don't implement it into your product. So you can say, Let's take our customer data, and we will build like a an AI panda with him what's a playground and internal tool, and you all have it and you can basically play with it. And then stuff can come out of that. So So actually, you could also do that and say, Look, that could be something yet we're going to allocate and let everyone just explore it. That's really where our idea came from. Rod, the company that we currently have is because of just being kind of willing to put something out there. That is really bad. Right? And then it turned out to be good. Yeah,
Max Matson 53:46
yeah, getting it out there and perfecting it along the way. Right. Yeah, that's awesome. Well, Marco, this has been a fantastic conversation. It's been very enlightening. And it's been great to learn a little bit more about yourself about sector and your your view on on AI. Is there anything that you want to plug to the future of product listeners before before we're out? Ah,
Marko Jak 54:09
no, haven't prepared any kind of plug, maybe go get your headshots with us if you can, if you really want to. And if you have a good experience, please tell everyone else. Maybe we can do a promo for like your listeners, if you are keen. I would say yeah, I just want to say whatever. Thank you. For this. It's, I think you'd what you're doing is quite a key topic in this space. And especially as we sort of unpack, like, what does it take to build product in this space? Like you look at stuff like generative kind of generative UI that's coming out now. So you ask a prompt and it actually returns the entire interface for you to like, what kind of headphones should I buy generates the entire like, interface you want to choose those choose those those kinds of noise cancelling Bluetooth? So there's building products in the space He's going to change. It's going to be quite a big change, but we don't yet know. So having these kinds of talks is really, really good. I think you're going to blow up you're going to be big. One day I'll show up to like your studio and you know, there Yeah.
Max Matson 55:16
No, I can't wait for that. Thank you, Marco. I greatly appreciate that. It's been a ton of fun. I mean, getting to talk to people like you see where the space is moving. It's been so exciting.
Marko Jak 55:27
Awesome, Max. Appreciate the opportunity.
Max Matson 55:32
Thanks for listening to another episode feature of product podcast, and a special thanks to my amazing guests. Marco. You enjoyed this episode. Want to learn more about what to do over player zero? You find is the player zero.ai You're looking to go even deeper on the subjects we talked about in the pod. Subscribe to future product and substack Be sure not to miss this Thursday's newsletter. Can't wait to see you there.
Thanks for reading Future of Product! Subscribe for free to receive new posts and support my work.