Salma Mayorquin 0:00 example of like action AI where it was a tool to help you extract key points and then track them through time to try to identify like a motion or buy that into something. Like, for example, we took that base kind of modeling to then feed into another prototype project where we're trying to identify deep fakes. Terry Rodriguez 0:21 What we have done is in preparing for a lot of content that we want to share, especially for this kind of developer. And so there's platforms like Hackster, Dev post places where there's like, hackathons going, going on, and we'd like to tap into those communities. Max Matson 0:44 Hey there, everyone, welcome back to future of product. My guests today are sama may 14. And Terry Rodriguez, co founders at remix AI. Cemetery. It's so nice to meet you both. Would you mind introducing yourself and telling us a little bit more about your journeys with remix? Terry Rodriguez 0:59 Thanks, thanks for hosting us. I'm Terry, co founder of remix. We've been working together for a few years, and most recently, we've been building remix AI, it's a platform for enabling users to create custom machine learning models, especially computer vision ones right now, just using simple prompt and high level description. Salma Mayorquin 1:19 Yeah, thanks for hosting us. Again. I think it's lovely to be here and be able to chat with you. And my name is Salma. I'm also a co founder of REMAX AI, Terry and I have indeed been working about 10 years together, we've been very deep into the AI ml space. We started off hosting a blog called smells like ML, where we started open sourcing a bunch of our work over the years. And that kind of snowballed into all kinds of opportunities, working with all kinds of groups and robotics healthcare. I recently did a stint at data bricks, which is an organization that is involved with like, big data in AI infrastructure. So all those kinds of experiences have now led us into the path of starting up remix AI this year, which is we officially incorporated around March. So it's been a few months. So we have a long way to go yet, but we're really excited to kick it off. Max Matson 2:14 Awesome. Very cool. Well, to start off, let's talk a little bit more about smells like ML. First of all, great name, I do wonder what ml smells like? Salma Mayorquin 2:24 He has a great story about how that name came about? Oh, yeah. Terry Rodriguez 2:28 Sure. Well, I mean, it's a, you know, a reference to Smells Like Teen Spirit. But um, it's also our goal was to try to create ml content we're sharing that was aimed more at conveying the intuition. You know, when we started the blog, we thought there was a lot of kind of introductory content. And we wanted to do more with showcasing in melanin application and custom applications that we were building, and how we saw this tool fitting into the products we wanted to create. Salma Mayorquin 3:00 Kind of hitting on on that idea of extending it, we were really developing our intuition for how we would use the technology, how we might anticipate it would work and play a role in building like a larger project. So it's very much kind of like smell hinting to intuition. Terry Rodriguez 3:17 Yeah, since the people associate with intuition or kind of your your gut. Max Matson 3:25 Gotcha, gotcha. Well, first of all, I love nirvana. So great reference. What kind of motivated you guys to to start this right, because y'all have been in the AI space for some time now? Terry Rodriguez 3:38 Yeah. Yeah. For smells like in Mill. We were we were interested in putting out our projects. And we saw that as kind of an ideas lab, we had been exploring different ideas and thinking about where there might be an opportunity to build product. But as we were exploring more with the remix idea that we started, beginning of the year, we felt like there was a clear product need. We'd been working in helping other groups in their machine learning initiatives, as ICs are the single individual in a company trying to start off some initiative or consulting with groups that have bigger teams and is part of bigger engineering efforts. And there's a lot of problems that both experts and novices face something like data access can be a blocker for anybody. But we saw the chance to use data generators, image generators to help alleviate that problem. We saw a lot of people struggling through challenges converting models, like what works in research maybe requires a lot of modification before it can make it into product. And in product you might face a lot of the real world constraints for you know, is this model fast enough? Is it accurate? enough, and it takes a lot of expertise to find the right part of the trade off there. So bringing bringing something that works and academic benchmark data set pretty well into production, where you face different constraints and trade offs was a challenge. And we saw a chance to build up some of that machinery and some of these tools with generative AI, we're doing a lot to help resolve those gaps with data access. Salma Mayorquin 5:28 Yeah, so I think the idea is, through the years of experience that we've had working with various groups, you know, as Terry described the various levels of expertise, we found that there were still some recurring same blockers that a lot of these organizations were experiencing. So for example, as Terry said, the lack of data, you could be a large organization with plenty of data, but now you're stuck with the problem of processing it too, before handing it off to the machine learning engineers to actually produce a model out of it. Or you could be at the place where you don't even want to do data licensing, or you're not really interested in building up the infrastructure just to get a test or an experiment out. So we saw with the latest advancements, you know, in the last year or so with generative AI and LLM, be able to construct a product that could help bring down some of those barriers, and get more people through the finish line just getting started just getting something up. And as we continue developing the product be able to support an abstract, essentially, normal infrastructure, machine learned custom machine learning into a service that most people could use with minimal expertise required. Max Matson 6:30 I see kind of building the scaffolding for the future that we all see coming where this is a must have for every business. Right, exactly. Salma Mayorquin 6:36 I think the promise of AI has been around for many years now, prior to the the breakthroughs that we've had recently, but you know, we're still experiencing a lot of the problems a lot of the baggage in some way that you know, pressure is to just being able to produce like a machine learning model that then gets used in another context, Max Matson 6:58 I see makes a ton of sense. So just to rewind a little bit, could I hear a little bit more about both of your individual kind of paths to AI. And what led you to want to specialize in this in the first place, because like I mentioned earlier you to have been in the field for some time, long before all the hype kind of came around, right? Terry Rodriguez 7:17 I guess I got into machine learning as a career around 2014. Prior to that, I was doing a lot in math, and I was interested in analysis and optimization. But as I was looking for ways to apply that, that training, I became really interested in the machine learning problem in automation. I, I started in, in healthcare. And so as this was an experience that I had, as you know, the only person representing that, that initiative and a large organization that has you know, politics, and it's highly regulated, Data Access is difficult. And through that experience, I learned a lot about processing text and data and building cloud cloud infrastructure around that. But I became really excited by what I was seeing with computer vision, this is my 2016 2017 Deep Learning head was starting to take off maybe 2015 2016, it became easy with the tools that were out there. And I wanted to invest more in that it looked like deep learning was a was a powerful tool for approaching a lot of different problem types, different domains, different data types. So I really wanted to build out more experience with that became, we became very excited by what we were seeing with how people were able to go to market and the Internet economy. And it seemed like it was easier for a maker to reach an audience and make a real product. And I guess we started experimenting a lot with hardware products around that time and putting out a lot of our projects, prototyping, like aI enabled IoT projects to learn more and the constraints that you face building, building out those kinds of tools. I guess more recently, like, as we had been doing more with vision and video, and deep learning and kind of maybe more interested in software than hardware, it's easier to reach your users, we, we really became interested in trying to address these problems around building ml infrastructure, doing more automation using synthetic data, data generators and procedurally generated data to deal with these problems. And we've always been working as a small team with limited resources. So we've had to use transfer learning and be learned to be very efficient with limited data. So I think this experience kind of like put us in a mindset where we could we could help other people especially in that zero to one On phase when they're trying to bring all the resources required to stand up that that demo, get that first version out so that they can iterate. Max Matson 10:09 I see. Gotcha. Salma Mayorquin 10:11 And I think my journey into AI was definitely influenced by theories journey. I was still in college when Terry was starting out in that healthcare organization. And I was observing a lot of what he was into, generally what was happening in software overall. And I also have, I discovered my love for hardware, I started off kind of self teaching myself about computer science through working with Arduinos. So I was kind of tinkering around a lot in that area. And what excited me the most about that space was the idea of being able to one day have intelligent systems or software systems that can power things like hardware that they can actuate in the real world and interact with us, I thought that was super cool to get involved in. And I kind of just wanted to figure out if there was a way to intersect those two fields. And I think when Terry and I started smells like ML, and we were really looking into computer vision perception systems, we were experimenting a lot in that space of putting models on Raspberry Pi's, which at the time was not really supported. But it was very fun to make. And so through through that intersection we were experimenting on how do we bring machine learning systems to constrained worlds like, you know, small hardware devices, or, you know, limited scopes, or limited resources, like data resources that we had to bring together and curate to be able to train these models. And after some experience of kind of working on a consultancy basis and open sourcing these projects, I kind of wanted to expand to see how organizations that were a little bit more developed, or were looking to be more developed, how do they approach this problem? How did they set up this like, massive, vast infrastructure. And that led me to data bricks to help out with kind of their initiatives to help support VB like the ML expert in terms of advising folks in their infrastructure, to be able to be able to service and create these products. So through that experience, I was able to broaden my horizons from kind of the smaller prototype world to kind of like the larger enterprise or commercial, you know, startup kind of infrastructure that you might expect, or hope to get if you're trying to do these problems nowadays. And I think that through those experiences, still seeing the same patterns kind of emerged, you know, depending on what stage you were in, they were still there. And also just inspired by all the innovation in this last year or so, I think that sparked my interest, with Terry to come together and bring Remyx to fruition. Max Matson 12:44 Very cool. Very cool. So you both mentioned kind of playing around with hardware and you know, IoT type of things. What type of stuff were y'all building? Terry Rodriguez 12:52 Oh, well, we've made a lot of, I guess, ideas around smart cameras, right? Cameras are really cheap sensor. And there's so much information you can get from a camera feed. So for us prototyping different projects, with with like, little rovers or little like, like home automation tools, we use, like turtle bot or just something everywhere flying around, just trying to explore the different contexts that we could be be learning machine learning with. And I guess, I think I think that was important for us to kind of get a broad perspective on the problem, you know, especially lately, like, there's a lot of the narrative around how like, bigger and bigger models and and massive clusters is the way forward for AI. And that's true for some types of AI that's happening. But it's also true that we will expect to see a lot of products that have cameras on everywhere to enable them to move around in the world. So they'll need to perform perception, they'll need to perform scene understanding all of that processing. It's not realistically happening anytime soon. We're all just streaming, streaming, a continuous feed of video up to some server like this. And so the practice that people are using is bringing these models down to the edge. And it's getting cheaper to bring powerful processors closer to the camera. So you don't necessarily need a massive GPU to run these smaller models. But you can benefit from small hardware accelerators that are cheap enough to embed with the camera. And so it's a different paradigm for how you're deploying machine learning. And it comes with these additional challenges. So we're interested in helping people around kind of these really technical paths to bringing vision we're seeing a lot of breakthroughs in and being able to reason about text and, you know, go from few instances of text to doing doing, you know, spectacular feats and automation, which at GPT. But there's not some ability similar to that for vision yet where you can quickly adapt to some limited context and then have a, you know, a very performant solution for vision. Yeah. Max Matson 15:26 Did you have anything to add there? So, Salma Mayorquin 15:28 I think, kind of reflecting back on some of my favorite hardware projects that we've done. One of our earlier ones is called Yogi AI, what we did was use a Raspberry Pi, right, and we put it behind a kind of one way glass mirror that would allow you to practice yoga. So it would identify your pulses, and it would have like a kind of like a voice agent that would guide you through a flow. So it would be able to identify if you did it right, and then move you to the next step. Otherwise, it would keep you in, like, make sure you're hitting the pose. But I thought that was like a one really fun project. And also kind of like a great example of why you might want to consider, like, you know, small edge type of computer vision, you know, that might be an application where you might not be your customers might not be comfortable streaming, put yoga poses up to the cloud, necessarily. It could be done down on the edge, it still be functional, along with like other systems that are operating on that Raspberry Pi, something like super minimal. And we did this at a time was like 2018 19. So it was really early. And I'm sure like nowadays, there's even more opportunity to be able to bring our resources to kind of that level. Gotcha. Terry Rodriguez 16:45 Good project for us. Actually, like we got a lot of mileage out of just experimenting with body key point estimation on camera feeds. And one of the first things we learned is, you know, yoga is easier to do, because you're not moving that fast. But getting the inference time down so that we could work off of like faster motion and moving from like something like image classification to like video classification, being able to do things with activity recognition, just like kind of was the seed to maybe a half a dozen work projects that we were doing that are based around that idea of analyzing video feed, and looking at body key points to make inference about what a person might be doing. Max Matson 17:30 Very cool. Very cool. I just to kind of go on a tangent real quick. This is super interesting, right? So what are some of the end applications beyond like yoga that you guys see kind of coming out of this type of technology? Terry Rodriguez 17:44 Yeah, well, with the yoke AI and what became action AI or generalization of that idea, we were able to show lightweight activity recognition that can be deployed on something like a Raspberry Pi. And right now, the best Activity Recognition models, they don't really run in real time, they don't run at the edge. So it means that you can't really build an application that is going to be low latency and be able to infer a bunch of information about how people are interacting in a scene. And so you could ask yourself, where that might be useful. Like, if you are in some in some contexts, where you need to closely monitor how people are moving, some people are building products around helping helping to understand what's happening in an operating room, where some people are, you know, putting these this kind of sensing technology in some part of their warehouse. So there's a different different contexts where being able to decompose the scene into who's doing what, when how that's all that's all pretty important in action AI was a was a tool that we had built, the allowed people to really quickly customize this type of model to their activities that they might want to be monitoring specialized to the camera perspectives they might be facing. And all the other kind of practical things that you would deal with is you're trying to really build an application and you realize you can't run that research demo that's meant for like a GPU with 16 gigs, and you can't really build like, any kind of IoT prototype with that. Salma Mayorquin 19:26 Yeah, I think, you know, the example of like action AI was a tool to help you extract key points and then track them through time to try to identify like a motion or apply that into something. Like for example, we cook that base kind of modeling to then feed into another prototype project where we're trying to identify defects. So we took those same ideas about like, tracking key points on a human face or like you know, other body areas to try to identify whether a sequences do effect or if it was real. So we kind of anticipate that, you know, some of the components or the the foundational computer vision models that we're helping people create quickly, can then be used in all kinds of ways that we might not even imagine right now might tie into, right? Like, you might not necessarily connect the actual AI problem about identifying people dancing versus running or something to then identify defects. sure that Terry Rodriguez 20:27 the body key point estimation, you know, we can just generalize that to thinking about when do we need to take an image frame and make some really localized regression, like, we need to find a point. So its face key points in one context or body key points in another. People are tuned these models to identify key points on an animal. And it doesn't have to be a living thing, you know, you can find special points on any kind of object. And if it was moving around in a way that understanding the position of those key points in time was a strong signal, then you could use this same idea that we're talking about, to try to classify, classify the mode of activity for that thing. Max Matson 21:14 Very cool. Very cool. So it almost sounds like, you know, in sort of the way that LLM is are a lingual intelligence in a way, this is the visual side, right? So if you think about a human being having all of these different forms of intelligence that flow through each other and interact, this is kind of the building the eyes for the robots, if that makes sense. Terry Rodriguez 21:33 We want to do something like that. And to borrow that analogy with LLM, you know, there's a fantastic work, metal put out some work recently, around foundation models provision. And so you can kind of see things trending in that way, as you know, but there's a lot of these very specialized tasks. And vision is not something that is well handled by a unified model, or a single collection of models like Chet GPT, right, like when you look at vision, and where it's made the most progress. Now you're probably looking at autonomous vehicles, sort of, there's all these cameras, and some are like hyper specialized to seeing and the far the far distance, and others are for looking at things nearby. And it's like, very fit to the perspective and so forth. So we, you know, we imagine like to reach the best performance, we're talking about bringing spa models close to the camera device, so that it can tighten that iteration loop. But we're also talking about specializing to the spirit perspective of the camera. And so in robotics and autonomous vehicles, probably lots of IoT products, probably AR and VR, you're going to want to, you're going to want to recognize that there's like a mode. For that perspective, if it's a human point of view, if it's down by the wheel wheel well of a car, there's a there's things to take advantage of there so that you can get even faster models that are even more accurate, because they take advantage of all the priors, knowing that my perspective looks like this. So the morphology of the cars, the robots, that's all gonna be pretty important to bring vision to the hardware, in a way of vision is kind of more coupled with the hardware problem. Salma Mayorquin 23:19 Yeah, I would say the also borrowing off of the LLM analogy, obviously, that's it's been exploding in popularity, because it's so easy to use, right. And it can be so flexible in so many contexts. And I think taking inspiration from that, we kind of want to build the equivalent for vision, where essentially, you know, a user is able to just tell our engine or remix AI agent, what it needs to do, and then it will then customize the components division components under the hood, and put those together into a system that then just executes whatever the end user is looking to Max Matson 23:57 have. Gotcha. Okay. Very interesting. We'll come back to this. But I do want to bridge into kind of how does all of these pieces that you worked on, kind of in the lead up to remix? How did that inform your approach to actually, you know, go into market building the product, all of that? Terry Rodriguez 24:14 Will I guess, to take take the baton, from what you're saying here? I think we were really inspired by the ease of use of dealing with chat. I think when we first started building remix, our original conception was just like all in on automation eliminate any decision for the end user. But we actually found that people would prefer to have more transparency, they want to understand what decisions were being made. So it was important for us to earn the trust of the more advanced user who kind of knew what what they should expect, or to educate the novice who had no idea what to expect. It was important for us to kind of surface some of the information And behind like the execution plan for how we're training the model surface those options in case the advanced user wants to override a decision and wants to once it changed that. And so I think we really saw through the chat interface recently, LLM is with tools. And open as functions have made it possible to make this chat very interactive, very robust. And so we were really leaning in on bringing the LLM into the product that way, we're where we started was to bring auto ML to data generators. We understood that that LLM reasoning engine and the agent to interact with was a really valuable part of what we could do with remix. Salma Mayorquin 25:49 Yeah, I would say that, for me, my especially my latest experience, I was talking to a lot of customers every day, coming from all different backgrounds. And I think that all of the the various like hiccup, or blockers that a lot of these groups were experiencing, to me it kind of all distilled down to, if they had something that was very easy to use them, they will get past a lot of these hurdles. And so I think the latest developments in LLM and generative AI in the first idea of coupling generative AI with auto ml was kind of trying to come up with a solution that was hitting more on the nose on that ease of use. I think that over the last, you know, decade or so we've had lots of tools to enable more developers, but it's still kind of like a very technical developer, a very specialized developer that can use these tools effectively. And so I think with the latest waves, kind of like the latest iteration of AI, we think that we can make something a lot simpler for more people to use. So that's that's kind of the the goal. Max Matson 26:55 Gotcha, gotcha, that makes a lot of sense. So Who then is your guys's kind of ideal user? If you're making it more, you know, kind of adoptable to non specialized technical folks? Terry Rodriguez 27:06 I suppose we think that right now, an ideal user is somebody who's trying to build like IoT projects like we were and the the challenges of doing that for like a custom application are not trivial, as we're saying. But they, they that shouldn't necessarily be a blocker to their innovation. And so if we can provide that scaffolding that allow people to, to de risk the development of something or prototype something for a project, if it's for some early initiative and in their company, what they want to do with product, I think, we think that it's we think we think that tool is going to be really well well suited to knock down those those challenges for a lot of users and make it more accessible to a much broader audience. Salma Mayorquin 28:02 Yeah, I visualized that we start off with folks who are very interested in in building these applications, but maybe like, don't have the data or aren't really trying to get into the weeds of which architecture do I pull? And how do I put this together. And I see it overall, as we continue building out the platform and the features to support and just general software engineers. So you don't have to have any kind of context about machine learning, particularly to be able to produce something of that nature, and then use it in other context, and then maybe eventually even expand out to non technical engineering folks. So people who might be more kind of product development, they might have ideas, but maybe not necessarily practicing, they might be able to put together a concept or a demo that then can be relayed off to another team to build out more fully. So I think that's kind of the trajectory I see so far. Terry Rodriguez 28:58 And then we'll then just one more on that, you know, as we're talking about the tool, developing and becoming accessible to less and less technical personas, you know, ultimately, we see a space where agents need to see still. And so an agent might be tasked with processing a lot of video or going through a directory of images and finding something, it might be tasked with monitoring a camera feed. And all of these situations, we're talking about not just a one off, send an image up to the cloud and get your answer, but it's something more intensive and probably needs to be more customized. So bringing bringing vision to agents probably entails some some combination of what we're doing to quickly use transfer learning to quickly use some vaguely specified user intent, and map that to a recipe where you can get a high quality model to do the task. Max Matson 30:00 I see Gotcha. And you've mentioned a couple of times edge versus cloud. For those who aren't familiar, would you mind just kind of explaining that dichotomy? Salma Mayorquin 30:09 Yeah, sure, I would say cloud would be you're using one of these popular cloud services like AWS, Google Cloud Azure, and you're essentially using their services in large compute resources to stand up, for example, a server API that can then interface with your model and produce a result for you. Usually, this might entail, you know, larger infrastructure, many computers may be GPUs. So you have the ability to have you know, fast compute to then be able to execute and work on your model, or use your model down to the edge. I think when we say that term, we kind of mean more like you're the the machine learning model with the process of program is actually running on your local machine, whether it be your laptop, or whether it be like an NVIDIA Jetson or some device that you use locally, the browser that's so some area where there's much less resources, you're constrained. So if you, you know, you don't have the ability to now turn on a bunch of computers to help with being able to influence or run that model, Terry Rodriguez 31:18 there's gonna be, there will be situations where it makes sense to use the cloud. Like if you're running a model that needs a large GPU, that's going to be the best way to do the inference, most likely, because on robots, you have constraints around power, like how much battery can you pack around, so some types of workloads, it'll be just better to send it off. But there's that trade off of the latency. And depending on the kinds of decisions you're making, the kinds of data you have, it can be better to do more processing on the edge. Sometimes privacy is an issue, like, there's many reasons why you might prefer to not send the data around and tend to just get what you can off of some client device, Max Matson 32:04 it makes a lot of sense, makes a lot of sense. So just to pivot slightly, did you both always know that you were going to be entrepreneurs? Or was this something that kind of came up opportunity? Oh, Terry Rodriguez 32:16 I guess, you know, we've been, you know, pretty independent, with with our work and console consulting and entrepreneurial thinking about invention. And these trends we're talking about, like how you can build a brand and, you know, find find your audience on the internet. And I guess, like we'd gravitated towards that. I know, for me early, earlier on. Machine learning meant at one time meant, like you're just doing like click through rate prediction, those were the problems that mattered. And it wasn't interesting for me to work on those problems. And I That's why I was gravitating more towards using deep learning and understanding like content. And I was interested in kind of analyzing really complex data and getting structure out of it. And so I gravitated towards these problems, where were maybe the bottleneck in processing was something like, deep learning on a camera feed. And so I guess, to me, the problem space was more interesting. I thought the building for ourselves would be more interesting. Like, I don't have a click through rate prediction problem. But I imagine lots of exciting things that I would like to see like moving around doing stuff for me, I, I love the idea of kind of kicking back and knowing that there's a lot of automation happening, and I'm doing doing so many things because of the investments I'm making into those systems. And so I think lots of people want that it's just still a challenge for anybody, not everyone can build their own, but it is getting a lot easier for people who can build them to package them for others. And that's what we're really thinking a lot about now is like, how can we take the problems that we were really interested in, that have been difficult for others, and just try to package that to help others have a higher success rate. It's like building the model is just a small part of that product. You know, there's an awesome paper out there highly cited hidden technical debt in machine learning. And that's really like kind of the maybe the genesis of like ML ops, is understanding that the model training is just a small part of all the infrastructure that goes into maintaining and developing a system that can not succumb to the problems like data drifts and different things that you need to monitor for. So for us, getting into this space and seeing how complex it is, and understanding that if if we did what we're what we're seeing out there, which is to make a co pilot for these kinds of workflows, if we could if we could do that effect. Tivoli, there will be a lot more of these products that we're excited to see. And so I think like being able to being able to take something that we were so passionate about, we like kind of spent the last decade more or less, diving into it, taking that and making it more accessible and scalable, we think is like the most realistic approach to AI, making it in the products that we see. Like not everyone's going to be able to have have these big costly initiatives and like kind of research projects and everything. Salma Mayorquin 35:34 I think, for me, I definitely did not plan on entrepreneurship, I had a vague idea of what that meant. But when we started smells like a metal, I think that was kind of like one of my first taste of what that world was about it was, you know, I was very excited by building I love making things generally speaking, whether it be software, hardware, and other stuff. So that energizes me a lot. And then I also kind of through that endeavor, trying to put out a blog and promote it and connect with people and talk about that, I think kind of discovered other parts of entrepreneurship, even though I didn't really call it that or knew about it at the time. And so I gravitate towards just being able to wear a lot of hats and be able to put something together and then hopefully be able to service other people and have that help them in some capacity. I think that's very exciting and very fulfilling. In my case. Max Matson 36:31 That's awesome. So I'm gonna throw a big word out there. disambiguation. What is it? And why is it a problem? Terry Rodriguez 36:42 That is an interesting challenge that we were thinking a lot about, as we're describing, like, how can you allow a user to give you some vaguely specified intent, like the idea that I was keeping the example I was keeping in mind was a bass B A S S, it's a word that has many meanings, there's like something like nine different ways to interpret that word. And depending on the way it's pronounced, we might know that you mean fish. Or we might know that you mean meant bass for music, or an instrument, some quality of music. And so I guess, thinking about that example, you know, we were trying to figure out what ways to robustly parse the user's intent and make the right right decision. If, if you had additional context, like, you know, the other class that you needed to predict was trout or something like this, then it'd be it'd be the right assumption to make the bass was to fish. But maybe, maybe take that idea a step further. And what if the two words you had were bass and drum, it turns out that, you know, you can interpret those words both musically and as two types of fish. And so the important idea here is like getting as much context as you can from the user. Like when we have a user fill out through the text prompt, there's a lot of ways for them to describe what they're what they're doing with the problem. There's ways for additional context to help help tip the agent into reasoning about it one way or another, but also present the plan to the user so that they can kind of check off on things. And part of ways that we're generating the data is like, when we when we understand the context of, of these concepts that were introduced by the user, like I want to train a fish classifier, and my classes are trout and bass or whatever we can, we can, we can also use these powerful LLM to give us more information to help us with generating data or retrieving data. So when we know when we know more information about the context of how the application will be deployed, it can help us generate data that's closer to the distribution you expect to encounter. We're working on ways to extrapolate from a very, very small set of images and getting a lot of the information we can to to generate better data to Max Matson 39:24 gotcha, gotcha. Salma Mayorquin 39:27 When Terry came up with that example, and he was kind of like, working through this thinking logic on that I was like, Oh, wow, yeah, that's, that's definitely an area that we need to account for. And I think you're baking into the application further the use of LLM to in various aspects like to disambiguate what a user might mean to then also kind of guiding them through what context is necessary for us to then be able to produce an output from that has been an interesting you know, add to the to the product and I think helping us get further in To that ease of use for users. Max Matson 40:02 Gotcha, gotcha. And then I would imagine on the flip side of that is kind of the model being able to rationalize why it's doing what it's doing. Right. So that's one unique thing about remix that that you guys also provide that the rationale that kind of leads to the decisions. Have you guys found that that kind of increases user trust in the model? Well, we Terry Rodriguez 40:20 we did introduce that to address an issue that had come up, we were chatting with a redditor. And the comment was, you know about how intermediate and advanced users are going to want to understand what's under the hood there. And so this, this was a useful idea, we thought that would help earn the trust of these types of users, and also kind of educate the, the less opinionated or more novice users about the types of choices we're making. And you can imagine, like, adding more explanation to that you could imagine introducing more of the domain knowledge that you would surface there and make a chat agent that can really guide you through more of the how and why but behind these decisions in machine learning, for now, it's it's a pretty simple, it's like sticks to the plan we chose. And what we've done is we've used a lot of the best practices that have emerged in engineering and from research, we've combined a lot of that and kind of provided that context for the agent in your interaction so that you can benefit from the priors that we would have used. If we were talking with you right now about helping you build a new model for an application that you wanted to enable a vision for. Salma Mayorquin 41:38 Yeah, I think that kind of also ties into the transparency that we got feedback from, that would be a great add to the application. And then kind of using that concept or idea to then also move into more Explainable AI, thinking, something that kind of has been brewing in our minds is that in the future, there's going to be more of these agents kind of like independently executing logic. And we're going to want to have the software engineer right now that might have a heavy hand in creating that code. Now that it's becoming cheaper to generate and auto execute, have that engineer function more as an executive, so being able to check on what is occurring, or what the agent is going to plan to do, before it executes it. So kind of we're right now starting on the building blocks of being able to have that kind of executive summary that a user then gets presented and then signs off on. Max Matson 42:36 Gotcha. So actually, you know, what, that um, that kind of prompts an interesting question. So you guys are cofounders of remix and AI company, right? Machine learning more accurately? How do you approach kind of adding new headcount? Is that something that you think of through the lens of is this automatable? First? Terry Rodriguez 42:58 Well, for us, you know, the headcount question is not something we're thinking a lot about yet. It's still super early, it's just the two of us. But we do see, we do see these tools, enabling people to do a lot more. And I think a lot of people are saying, you know, your strategy and growing a team is gonna need to take this into account. And so for us, you know, we really value the fact that we're, we both own like, much of the decision making in here with the, with the, with remix. And I guess, like, we can operate as two people with high bandwidth, and a lot of kind of ability to bounce these ideas around and iterate and improve that. I guess, like, as, as the team grows, you know, we do think that there'll be plenty of things that we can sort of delegate and automation is going to play a big part of that we're investing a lot into kind of dog fooding around things that we think we'll be doing a lot of, you know, I we mentioned that we've built this automation around machine learning, but something that we've just been interested in, or finding ourselves doing a lot of as we're showing, showing our work is video production. So we've been developing tools around that, too. And so for us, I suppose, like headcount, you know, we think a lot of teams are going to be able to go far with fewer people building out more tools like this or using services delegating in different ways. And being able to kind of own the full on the means of production more fully in our teams. So we expect to enjoy that advantage since we've started as a small team, and we've been building these kinds of tools and products for ourselves a lot for years now. Salma Mayorquin 44:57 Yeah, I would say that obviously you Right now we're just getting started, we have to be scrappy with with all the resources we have. And so we're going to use automation to the best to our advantage. But obviously, as we hopefully hope to grow and expand, we're going to need to bring on more folks to help us essentially proliferate and develop this further. We haven't crossed that threshold yet. But when we do, I envision that I'd love to bring on folks from, you know, a wide range of experiences of wide range of thought of perspectives. I think that that has helped us a lot in terms of maybe not necessarily coming from traditional backgrounds, or approaching problems and novel ways that maybe, you know, pieces weren't tied together before. I think we found a lot of success in that. And I'd love to replicate that with like, as we expand the team, I think that'd be really awesome. And I'd love to well, as we bring on folks, for everyone to have, like a lot of sense of ownership. So just being able to have her share the the load across everyone, and we're all building something sharing our perspective or our spike into building, you know, a tool that it theoretically and hopefully will serve lots of people. So we're going to need to account for lots of people's experiences and requirements. Max Matson 46:12 Right on right on I so I would like to kind of get a sense of being, you know, co founders in the space, and being builders and tinkerers, what is kind of your guys's approach to, you know, GTM, all of those things that boxes that you need to check, because you guys have an interesting background having also done, you know, smells like ML and kind of flex that muscle to some degree. Terry Rodriguez 46:36 Well, we have found smells like canal has been a good way for us to get a bunch of signups, you know, the that that brand has caused like us to have the distribution. But we've also been trying to build this new brand from scratch. And so there's a challenge there and kind of converting an audience from one to the other actually. And we're still we're still trying to do that, I think we haven't taken as many opportunities like this to kind of put our ideas out and talk through more of the details. And so what we've have done, besides putting out blog content, where we do talk about some of this, and newsletters and things like that, you know, all the typical social media avenues to show our work, what we what we have done is in preparing for a lot of content that we want to share, especially for this kind of developer. And so there's platforms like Hackster dev post places where there's like hackathons going, going on. And we'd like to tap into those communities. When we were first building the engine early on, it was too early for us to reach this audience. But we saw a contest that snap and Mel was putting on like sponsoring contests on def post, there were 1000s of people like trying to build custom machine learning applique or trying to build machine learning applications in unsnap ml lab. And we would love to find that audience and say, Hey, use this tool to do something custom. Because what you find in practice is that there's lots of code reference examples. And they're often based around like maybe the same set of models or something like this. And when you want to new model, you might first go look to a model Zoo. But still many of those examples are based on having been pre trained on the same research data set. So it really is that last mile optimization, the customization steps that are required to go from, I have all the resources in theory to like now, this works very well, for my application. The way Salma Mayorquin 48:45 I think I see go to market, as Terry described is very developer centric to start. Obviously, we're a two person team. And we might be kind of like, you know, scrappy, you're kind of like an efficient resources type team. And I think, kind of going developer first might be a good strategy for us to start with. I'd love to win the hearts and minds of developers by enabling them. As Tara described, we were thinking about putting on contests, where folks are invited to come in and try out the engine to build something new. I think that that would you know, developers who are coming in to build maybe personal projects or an idea that they have on minimal resources kind of helps paint that theory or thesis that we have that you know, you can do something like this very minimal resources using a service like remix AI. And through being able to kind of take their feedback and see how those those first group of users use the tool and how they'd like to interface with it. I think that'll inform how we want to build out the ecosystem, the platform so that it could then service more startup commercial enterprise grade communities that think that you know, there's different motions, it could be top down or bottoms up in terms of you know how you might sell to an organization, and I'm a fan of bottoms up, I think that you know, being able to win the the minds and hearts of the users of the tool that are then building out through the tool, like actual products to the, you know, the stakeholders, I think that that works really well. Terry Rodriguez 50:18 On that note, you know, anyone can actually get a free model just by signing up. So if you want to try it out, make a model, it'll take about 30 minutes, probably. And, and I think just making it like really accessible for people to try it out, test it on your own examples, give us feedback, and we'll give you some more free free model runs, we're still trying to build it out and really find really find the, the customer base, this is going to be like really exciting for, but we think that there's not really anything like it out there for you, if you're just trying to find a model to infer if this thing is in the field of view or where it is, can you detect it, there's there's no faster way to get a custom model to try out. So you can do that. And we hope to Osama says like went over developers with a with a one of a kind kind of product, the product experience and machine learning. Max Matson 51:18 Perfect, you're gonna get a free model. So with that, I have just two questions left for you guys, because we run a little tight on time. This first one I'm not exactly sure what I even necessarily want to ask. But it's a hot topic kind of in I think computer vision, right is just to whatever extent you feel comfortable talking about it. What do you guys think about some of the concerns around, you know, security and surveillance when it comes to computer vision? Terry Rodriguez 51:47 Very, very serious issues, you know, as we're talking about cameras on everywhere, I think people are entitled to privacy. And I think, you know, we thought about privacy a lot. You know, we were thinking some time ago about solutions around that, like how could you automatically mask images to protect privacy and make it where you needed to unlock a video file to be be privileged to see like, what was happening there? Like, how can you how can you enable privacy by default. And so for us, we were excited that for now, like a lot of privacy's built in by design, because we're not taking user data at the moment, like we're gonna work on ways to work with very limited user data. But we want to allow the user to have you know, the total decision power on if if that data is persisted, and when it goes away. I think, as far as like surveillance, and in a lot of the kind of ethical issues in computer vision, it's it's another reason why we're excited about the tool. Having internet skill data and generators means that we don't have to compromise on providing the ability to help someone train a person detector. But we also don't want to necessarily be leaning into owning models of that type. You know, there's there are, you know, services or experts, people that are really committed to making models that are like very reliable and fair. And we don't think that that's the the strength of generative AI data sets right now. But where does shine is in getting new access to data that would be difficult to acquire, you know, internet skill data, you've seen many of the objects that you might realistically care about, you've seen on many times, and through enough variation, that these models are pretty good at giving you data that meets the requirements. And so being able to help someone build up that custom data set that doesn't exist yet is probably where the strength of generators and internet scale data like this is right now. But there's biases in that data set that won't work for every situation earlier, I was talking about a camera specialized to the perspective of the wheel hub of a car, but we don't have large scale datasets like that, you know, this is part of the investment that we'll need to make as more users are turning to those use cases. And more investments in procedurally generated data like going outside of what can be done with the generators that are have been open sourced specializing our own generators, as we get more users using the data and we're building out beyond just the model training as we're able to deal with model deployment. And as we're able to work out ways to sample that data will be able to special allows our models to better suit those those users. So we have a we have a starting point with this technology. But there's still a lot of work and especially as we're trying to like navigate the ethical issues around pretty cameras that can sense everything out there. Salma Mayorquin 55:17 Yeah, I think, you know, as we we start off, obviously, we're the very beginnings are kind of like starting to lay the foundation of some of those components. But I do think that as a space develops, as we develop as an organization that these these use cases are definitely going to be sought after many people. And I think that, you know, bringing it back to that Explainable AI angle, I think that's even going to be more and more important as we kind of get deeper into using these technologies, bringing more and more context to the end user about what the machine learning model has captured and how it operates, I think will be really important if people choose to use these technologies in that context. So I think I'm excited to see more of that happening so that we do have safer applications in those more sensitive contexts. Max Matson 56:07 No, right on, it's, it's definitely heartening to hear that, you know, the people building these types of things are thinking about it, though, right? Like, I think, yeah, 100%. All right. So last question. You two have had kind of long careers in this field, and built a name for yourselves, remix. I've talked to quite a few people who are very excited about you guys. So all that being said, From your vantage point, what advice would you give to budding entrepreneurs who are looking to get into the field? Terry Rodriguez 56:40 Well, I would definitely say like turn to the problems that you're interested in and try to copilot your own workflows, you'll be more productive, you'll be able to focus on the interesting, interesting parts. So much of these tasks are just kind of like a series of decisions that have to be made. Like there's some executive decision making it certain junctions, but then in between that is a bunch of rope work, things that a program or a tool is good at. And so try to find that and try to kind of abstract your workflow into parts or parts where you are focusing more on like those key decisions, and everything else will be kind of better, more reliable, faster, cheaper, I think trying to co pilot, the things that you're interested in, or you want to do better, is a great place for entrepreneurs to be looking to who are trying to get into AI. Salma Mayorquin 57:38 Yeah, I would say, Even so far, it was like our short journey already has taught me a lot of things. And I think one of one is stuck through me is don't let your current limitations limit you indefinitely. So I think that, you know, just because you don't know how to do a thing right now doesn't mean that you might have the tools accessible to you at this moment to go and learn it and we'll do it. One, one example that comes up to mind early in the year is that, you know, to build out the MVP for the first conception of the product, which is the auto ml plus the generative AI, we knew that probably a more complex infrastructure that I personally hadn't had experience building before would be required for the short term and hopefully in the long term, that would be kind of like the best decision to build into. But I definitely had to learn a lot of tools along the way to make that happen. And I think that you know, initially we thought it would take us a few other people few other folks to bring on me more money resources time to just get that MVP out. But by just kind of taking a step back and then seeing the ecosystem of tools, even things like GPT to help kind of guide you on how to do a thing. More of the things that you know you need to do and might not know how to do them now or actually closer in reach than you think they are. So kind of just go for. Max Matson 58:59 Yeah. 100% both great answers. No, I completely agree. I think it's it's more and more necessary, right? It's like I had a boss at one point he said never asked me something that you can google right. I think that's that's become relatively universal at this point. So guys, thank you so much for this was a fantastic conversation. Where can people find you follow you? Terry Rodriguez 59:21 Sure. Visit Remyx AI and follow us on LinkedIn, Twitter, Salma Mayorquin 59:27 YouTube, Roku, so we'll put more resources up there. The landing page, we mix that AI is the best place you'll find our most of the resources there. And you can always shoot us an email you can say Terry at REMAX AI or some of you may say if you need anything Max Matson 59:42 perfect, perfect ends R E M. Just so that everybody's clear. You got it. Awesome. Thank you so much, Terry & Salma. So um, I really appreciate it. Salma Mayorquin 59:50 Thank you so much. And And thanks again for hosting. I think it was a lovely conversation. Great. Oh, my Max Matson 59:55 pleasure. My pleasure.