This solo founder bet on AI 7 years ago. Now he has 5,000 customers & $115M raised. | Dylan, Founder of Assembly AI

Dylan 0:00 This is actually maybe a good tip for startup founders that are fundraising. I actually write up a big document with all of the objections I anticipate to get, and I write out lengthy answers to all of them. I don't share that with anyone. It's for me to make sure I'm really, really tight and articulate on all of the competitive and market questions that I anticipate to get. Pablo 0:21 Welcome to the Product Market Fit Show brought to you by Mistral, a seed-stage firm based in Canada. I'm Pablo. I'm a founder turned VC. My goal is to help early-stage founders like you find product market fit. Well, Dylan, welcome to the show. Dylan 0:38 Yeah, thanks for having me on. Pablo 0:40 I'm really excited to have you here. You've had a pretty tremendous journey. I mean, you've been doing this now since looks like 2017. Over the last three years, at least from a fundraising perspective, things just blew up. You raised multiple series A, series B, series C, all $50 million plus rounds. Curious to see what you did in the early days that got you there today. Let's start at the beginning. I mean, just walk us through just the context in 2017 when you started Assembly AI and just what led to the idea in the first place. Dylan 1:13 I The Start Of Assembly AI Dylan 1:14 knew that I loved startups and working on startups because I'd actually started a company when I was in college. It was a terrible company, terrible idea. That's where I learned how I learned how to program because I built everything for that startup. After we shut the startup down and we were working on in college, I just knew that I loved software development. I loved building things. Creating a startup is the ultimate task of building something because you're just constantly building. In 2015, ‘16, I moved out to San Francisco and I took a job as a machine learning engineer at Cisco. I was working on natural language processing and understanding systems. It was around that time that the Amazon Echo came out. I remember that just felt like such a futuristic experience. I still remember the commercial. My immediate reaction was no way that actually works. These people are just sitting on their couch and asking for this song to play and some speakers hidden somewhere. I bought an Echo and I just remember feeling so blown away by how well it worked, how futuristic it felt. I would constantly test it. I would be rooms away with the shower on and the sink on and the TV on and see if I could still get the Alexa to play this song or answer my question. It worked and it was just so crazy and so cool what a good experience felt like, because prior to that, I mean still today, which is crazy, you talk into your phone sometimes… Pablo 2:45 It's terrible. Yeah, I don't understand what's going on. Dylan 2:49 I got really into this whole idea of natural language interfaces, especially over voice. Then I also totally saw where machine learning was headed. Pablo 2:59 You saw this Gen AI, I mean, maybe you didn't, let's say, predicted, but you already felt things were going to go there. Dylan 3:05 Totally, because classical machine learning, you really had to build these task-specific models, do this handcrafted feature engineering for vision, for speech, for natural language were performing pretty well. They were very much the v0 and you just saw, okay, if the ceiling's nowhere in sight and these models are going to get bigger, they're going to train on more data, compute is getting better. Pablo 3:34 Was that generally accepted? Because that's still pretty early. I mean, obviously AI's been a buzzword for a while and it's gone through its own hype cycle. I'm just trying to remember back to like 2015, 2016, deep learning obviously was a thing, but would you say it was generally accepted back then that that was going to be the thing or was that a bit of an insight on your part? Dylan 3:51 It was definitely… I don't know if mainstream is the right word, because for example, I remember going to the first TensorFlow meetup down at Google's headquarters in Mountain View or wherever Google's headquarters are. That was in 2016, 2015. There was already this ecosystem around TensorFlow, which was the deep learning library of choice at the time. Pablo 4:15 It had some believers but wasn't necessarily a mainstream thing yet, but you were one of those, let's say, early believers. Dylan 4:21 The models, they had a ton of potential and it was still so early, but the actual accuracy of these models was still not that good, right? Even speech to text models that were deep learning based in 2016, 2017, the error rate of those models compared to the error rate now, I mean, it has dramatically improved. The first model I ever trained for speech to text was on 10,000 hours of audio data. Then now the model that we're going to release soon, that's trained on 12.5 million hours of audio data. It's on hundreds of TPUs. Just the scale has increased so much, and as a result, the accuracy and the capability and the robustness of the models that you can create now has improved so much. Now it is mainstream. Everyone's working on it. Every company is trying to implement AI or think about AI as part of their strategy, but back then, it was only the early adopters that were actually putting this stuff into production. Pablo 5:19 Did you, back then, was it always your idea to be this layer that others would build on top of or did you think through, because let's say, arguably if you have the best speech to text, maybe you need to just build an app to do transcription and you offer transcription. Did you think through that? Dylan 5:34 As a Product vs Platform Dylan 5:35 developer, I was always so inspired by the iconic developer companies like Twilio, like Stripe. There was a company at the time in 2015, ‘16, when all this was formulating called wit.ai, I don’t know if you've heard of them. They were acquired by Facebook. They're a small company, but they built this developer platform for building these text-based natural language interfaces. You could train it to understand, hey, I want to set an alarm. AI would take that prompt and would turn it into this structured JSON that your app could operate on. If someone said, “Hey, I want to set my alarm for 7 p.m.” they would spit out a JSON it was like action. Pablo 6:16 Right, so it’s text to code, yeah. Dylan 6:18 Exactly, and the community around it was so cool and people were building so many cool things with it. I say that because for me, building a developer platform was always something, as a developer, I was really passionate about. The goal really when we started the company was let's build really advanced new models for speech tasks, speech to text, speaker diarization, speech understanding, and let's make them available through a developer platform that is super easy to use, that has great docs, that you can get started on for free, and that can scale with you if you deploy it at really large scale, even if you're a big enterprise organization, but let's build that platform that anyone can just pick up and use, whether you're a college student or a developer at a Fortune 500. Democratizing a really powerful and cool tech, whether it's the ability to process a credit card or send a text message or do something with an AI model, as a developer, I feel like unleashes the potential of this tech to so many people. There's entire businesses that have been built on Twilio as a result, but it just shows when you make this technology just really accessible and easy, people will unleash their creativity. I wanted to work in that space. Pablo 7:32 At this time, you're in California, you're working at Cisco as a brand name company, you're a machine learning developer. You're, I'm assuming, making a pretty good salary. When do you decide to take that leap of faith? How far along do you get on Assembly AI before you quit that full-time job and just go all in? Dylan 7:49 It Taking the Leap of Faith Dylan 7:50 is funny because I worked on a startup in college and then I just did software development as a contractor for two years. Then I took this job at Cisco and then I started Assembly. I feel like the time at Cisco was almost this paid vacation, so to speak. It was active, maybe, no, that's not the right word. It's active rest mode. You know when you're working out you don't want to sprint but… Pablo 8:11 Yeah, you do stuff, not too much struggling, just enough that it gives you a little bit of… Dylan 8:16 I went to the gym every day at 5 p.m. I was in amazing shape. I was reading. I was learning things in my spare time. I just felt bored because nothing I was working on just felt like it had impact. I wanted to feel like what I was doing was having impact. Pablo 8:37 Was that your mindset? Through it, through that time at Cisco, you were like, this is my in-between. I'm going to find a startup. I'm going to find a problem I really care about and work on it. Was that set in your mind that you found something again or not really? Dylan 8:49 It just happened. I mean, I think it's always hard to predict out what your personal future will look like. I knew I wanted to go to San Francisco. I knew I wanted to get closer to the startup ecosystem. I knew that I would eventually want to start another company. Pablo 9:04 Where is Assembly AI when you take that… quitting is a big decision, going in, actually giving your two weeks. Dylan 9:09 There is nothing. It was just me and I was like, okay, I just want to work on this, and so I quit. I didn't want to work on it while I was at Cisco. I quit and I started working on it, forget this specific time, but it was near the YC summer batch application deadline. I wanted to try to get my thoughts out for what I was working on. I submitted an application to YC for their summer batch. This is summer of 2017 and it was 30 days past the deadline. I was single founder, past the deadline, there's no way I'm going to get in. I was planning to just work on it, recruit a co-founder, apply to YC later once there was more traction and we had more built and I had a team, but long story short, ended up getting into YC. Was in Europe at the time I found out I got… Pablo 9:57 What did they see? What did YC see back then? Dylan 9:59 Our group partner is this really brilliant guy. His name's Daniel Gross. He does a ton of AI investing. He had worked at Apple and just saw that there was an opportunity here because he had interfaced with companies building this type of tech. He saw that it wasn't that good. He saw that there was no real easy way for developers or companies to get access to this tech at the time. He was a believer and he was on the interview panel and then he became my group partner and he's now a major investor in the company too. He really believed in the idea, in large part I think is probably why we got accepted to YC. It was the fact that he had, I think, firsthand experience to see that this was possible because a lot of people at the time were like, oh, aren't the big tech companies just going to make this stuff and it's going to be the best? Pablo 10:45 What was your answer to that? Because yeah, that's the obvious maybe first-level question is like, oh, speech to text API, Google will do this or whatever. Dylan 10:55 I believed at the time and still believe that to create the best product is not just a function of resources. You don't just put a budget and people in a bottle and shake and then out comes this amazing product. You can definitely increase the likelihood of that the more resources and people you have to a point, because at a certain point, more resources and people actually becomes a hindrance. There's that software development book The Mythical Man-Month. I don't know if you've ever heard of it, but it's like the more people you add on a project, sometimes the slower it actually goes. I always felt like there was an opportunity here. Going head-to-head with a Google on a core area of their product, like search, that's a much harder task. For something that they're not focused on and they weren't at the time, the YC partner actually I saw recently wrote something about this on Twitter or LinkedIn, which was like, you're really not competing with Google. You're competing with a PM at Google that’s trying to have a successful product under their belt and that's navigating all these internal politics. Of course, if it becomes the main focus of a company, again, like search or now with Gemini trying to compete with GPT, it's a different story, but at the time and still now where we're focusing on creating our product is not within the core focus area of these larger companies. The opportunity space is around working much more closely with developers and with customers in this emerging market to understand what they need and to build the best products for them. Pablo 12:26 Let's go back to that. You get into YC, which oftentimes is a huge moment, inflection point, right or wrong, I think, for startups. Is that what happened to you? Did you find there's a before and after YC and things just took off? Getting into YC and Raising Funds Dylan 12:40 It definitely became more real when we got into YC. I remember going down to YC for the first day and I was like, I don't know if I can curse on the show, but I was like, oh, shit, this is real now because I had started the company a month prior and then in YC and YC is basically a sprint to demo day. It's really just 90 days, right? It's three months. You need to have progress made and I had nothing when I started. When I was building, you can't just quickly pull together a web app. You have to train these models. You have to get data. It takes a while. I remember at the time each model iteration took a week to train or something, two weeks to train. If you think about that, if you're just consecutively training models, you have six model runs essentially, because there's two a month and YC’s three months. Startups are all about iteration, especially in the early stage and still now. My ability to iterate was pretty slow back then because I was bounded by the speed at which I could train models and iterate on the models to get them to be good enough so that people would start using them. We didn't make a ton of progress in terms of go-to market traction during YC, but we did make progress and we just got started, were able to raise a seed round after YC, which gave us the runway to hire some people. Pablo 13:58 First of all, you keep saying we, but I remember it was just you. Did you recruit some early people through that YC stage? Dylan 14:05 Yeah, yeah. I hired some people through that early stage. I mean, when I got into YC, it was like, okay, you have 90 days. I can either use those 90 days or a significant portion of those 90 days to recruit a co-founder or I can try to just get really far myself and then raise capital at the end of YC and then hire people as, essentially, late co-founders. That's what I chose to do because I felt that was the better use of time. For the first month, two months, it was really just me. Then leveraged a mix of people in my network that I hired as contractors within the first couple months of YC just to get stuff going. Then after demo day when we raised, I think we raised a million dollars. Pablo 14:50 Was that hard? What was that raise like? I mean, because you have no real customer traction. I assume the model still doesn't really work all that well. I mean, it's really just a bet on you and deep learning in a high-level thesis. Dylan 15:02 Yeah, so no institutional funds invested. It was only angels that invested because a lot of the angels that invested, they saw that either through their own personal experience or their own career or other investments that there's totally an opportunity for startups to build great developer products in spaces where there's large incumbents. You've seen this, Stripe, PayPal is a great example and there's a lot of others, like Heroku. We had a lot of angel investors that invested in the first trial we did after YC that saw that there was an opportunity for us to create this company. Pablo 15:36 What was the VC path like? I assume that you spoke to them. What was their reaction? Why did they pass? Dylan 15:42 It was just the, how will you compete with Google question. That was pretty much it. I was like, okay, I know the questions they're going to ask. Now, actually, every time I fund-raise, and this is actually maybe a good tip for startup founders that are fundraising, I actually write up a big document with all of the objections I anticipate to get. I write out lengthy answers to all of them. I don't share that with anyone. It's for me to make sure I'm really, really tight and articulate on all of the competitive and market questions that I anticipate to get. That has been super helpful now. I had never raised money before, right? Never raised money for a startup. The very first meeting I took at the end of YC was with Sequoia. Pablo 16:25 Oh, no, okay. Dylan 16:28 No experience fundraising, no real progress, don't really know what I'm doing, and just start with the most experienced investors possible, probably not the right strategy. Pablo 16:44 Yeah, I mean, you went to the all-star game right off the bench, yeah. Dylan 16:48 In hindsight, probably stupid. I didn't have a name, right? It'd be different if I was this PhD or this very seasoned entrepreneur and I could show that I had a resume, I knew what I was doing, but I was unknown. We didn't really have a lot of traction. Even now still you're biased to think it will be hard for Assembly to compete, but there are thankfully really brave and ambitious angel investors out there. Pablo 17:13 That's right. That's why they’re called angels for a reason. Dylan 17:16 Yeah, exactly, and still investors, our investors now, I mean, Steve and Sarah at Excel and Rebecca at Insight, and then the more recent investors from Smith Point, like Keith Block, they all believe too. I think when you're raising capital, the biggest thing is just finding people that believe in the opportunity and that are excited about it. It's a bit of a numbers game, so you have to just meet as many people as possible to find that. It's like dating, right? I mean, you're not going to find your soulmate on your first date. Pablo 17:47 I think that's totally true, and especially for the early rounds. In later rounds, you have numbers, you have traction. You start fitting in specific boxes. There's still a numbers game element to it. I think in the early days, you're totally right. It's really about just finding believers. If somebody doesn't get deep learning, doesn't care about building platforms, doesn't really matter how great your story is, they're probably not going to be the right fit for you. That's really what you're looking for. Okay, so moving along, you raised that round. You have, I mean, even a million dollars, it's not a lot of money. I guess you hire a few handful of people or five people or so at that point? Dylan 18:22 We were three people for a long time. I think in hindsight still today, it's difficult building a research-based product because you're never done in software development, but there's a spectrum to the quality of everything you build, right? There's accuracy metrics and there's issues even with our current models that are really good and industry leading. It's very hard to know what is the threshold that you have to pass on certain metrics for the products that you're building. Pablo 18:51 Back then, what was… were you like we need to be 95% accurate or what was even the goal? You have three people, you have a bit of money, a bit of runway, what are you shooting for? Dylan 19:00 Because even if you take a look at GPT3.5 and GPT3 existed for a long time prior to ChatGPT. It was the RLHF and the dialogue component of ChatGPT that triggered the takeoff. That was the capability threshold. It was at that time that this capability threshold passed where now it is LMs are priority for every organization. Pablo 19:23 It's crazy. It's this good enough or not good enough, right? It's binary in a sense. At some point, you don't work in the eyes of people and there's some point where you pass it, well, obviously you still can get better but you go from not working to working. At least that's what it feels like from the outside. Dylan 19:38 Exactly, and I would argue GPT2 or GPT3 or I forget the specific version numbers, but it worked pretty well prior to ChatGPT. I think my point with this is, in the early days, I think we could have actually gone a lot faster if we had said, okay, let's do nothing but pick accuracy metrics that we feel we want to surpass. Once we surpass those metrics, then let's go and take this thing to market. Pablo 20:06 The thing is, didn't you just not know at what point the market would accept? Dylan 20:10 You didn't know back then, but I also think that we were trying to make short-term progress because we're in YC. It was hard to break out of that mindset post-YC. You have these weekly meetings in YC with your partners and it's like, okay, what progresses have you made? What progress have you made? Pablo 20:25 Well, the mantra is quick iteration, right? It's lean startup methodology. Dylan 20:30 I think there's some you should be growing 10% week over week during YC because if you're building a consumer or even B2B app and you can quickly acquire users and iterate over the weekend or at night, ship new features and iterate super quickly, you have a higher chance of finding that traction quickly within a compressed timeframe if you're working super hard, like you do in YC, but for us, because we were building models, it was like, all right, let's train this model. Let's see if we can get people to use it. They can't so let's train another one, see if we can get people to use it. Because we were trying to make progress really quickly because we didn't have a ton of capital. Back in YC, when we did it, you received a $125,000 investment and I knew that we would need to raise more funding at the end of YC. We had to try to show some progress, some traction, and we had a little bit, to be honest, which helped the round, it wasn't like we had nothing, but I think it was hard to break out of that mindset post-YC. For the first year, we were still, all right, let's iterate, try to get customers, iterate, try to get customers, whereas if we took a longer-term approach earlier on and said for the next year we know there's market demand, right? It's not like we're building a productivity tool where we don't know is there going to be product market fit for this mode of note taking that we are pioneering. We know there's market demand for this tech. There are industry established accuracy metrics that we can leverage. Let's set goals for these metrics and let's surpass them and then let's go to market after. We could have been much more explicit about that. I think learning lesson, for me, if I were to start another company again is really when starting, as much as possible, try to set really, really clear long-term goals and then work towards them and be very disciplined. You might have a customer that will want to use your early MVP, but if that customer pulls you in the wrong direction, they're going to have a ton of influence because they're your first customer and then you have to build for them, and for those early days, I think we could have gone faster if we were more focused and disciplined. Pablo 22:33 I think that makes sense. I think in your case, I mean, the way I abstracted out is, in the early days, you're trying to de-risk and you're trying to find what's the biggest source of risk for this to work to get to the next step and let's de-risk that. I think in most cases that risk is really demand. It's like you're doing something new, you don't even know if people are going to buy it. That's why you get into this iterative customer, hey, is this good enough, whatever. In your case, I think you're right, demand wasn't the biggest risk. The biggest risk was just can you actually with this deep learning infrastructure, get the technology to a place where it's better than everything else or whatever, and so it would've made sense. Dylan 23:08 If you're building a productivity tool and you don't show customers and you build for a year, it might be like, I don't care, this thing's stupid. Pablo 23:16 Correct, and there's no risk on building. You will definitely be able to build the tool. It's more just will they buy. Maybe walk me through this. You've got a million dollars, you got whatever, three-ish people, you're doing this cycle, you said, for over a year. I mean, cash is running out. I don't think you can get profitable. What's going through your mind in terms of where you need to get to to raise that next round of funding? Dylan 23:37 We Landing First Customers Dylan 23:38 started to get our first customers in, I think, late 2019. The first two years were just building. Honestly, our models were pretty good at the time from an industry perspective, but the industry even best in class back then was just not that good. You couldn't power that many use cases back then because the models were not very robust. They were not very good. For the first two years, we were really just building, wandering in the forest. Then it was around 2019 that we started to get our first couple customers. I think that was the time that we passed this initial threshold of like, okay, it's good enough now to power certain use cases and applications and customers started switching to it from whatever they were trying to do before or they were able to now build something that they had been wanting to for a long time, but it just wasn't good enough anywhere yet. Then that's continued to happen, right? As we've made models that have gotten better, as other tech within the ecosystem has gotten better, like vector databases and text to speech models and large language models, the amount of things you can do now with this tech is increasing and it's increasing every day. We always are seeing more and more and newer and newer projects and products and features being built because the tech continues to get better, the ecosystem continues to get more mature, adoption continues to mature, and it's still super early from an adoption perspective when you look at the market. Enterprises are still really figuring out what they're doing with AI, what the use cases are, but it's maturing rapidly. Pablo 25:12 Do you remember some of the first use cases, some of those first customers, what their use cases were or maybe even one of the first things you remember where you were like, wow, we're powering this application. This was what we've been building towards. Dylan 25:25 Yeah, so one of our first customers was building this product where they were analyzing TV and radio stations 24/7 and then they were looking for certain brand mentions that were spoken and alerting those brands when their names were mentioned. That was one of the first customers we had and they actually found out about us on Hacker News. The CTO reached out and they checked out our API and they liked it and they built this product with it. Then there were a few others. Some of the initial customers were in the call center space, so analyzing customer support calls to create insights. There were some voice agent, like voice bot applications in the early days too. You're seeing a lot more of those now, but primarily in the early days it was media contact center type use cases. Pablo 26:17 Based on Crunchbase, what I see is you raise this YC thing, you raise this million-dollar round, and then the next thing that's listed at least there is 2020 raised $50 million. What happened between? Was there a smaller, a C+ or a series A or something that bridged that gap? Dylan 26:32 Yeah, so we raised a million dollars after YC in 2017. Then I think it was right after COVID, it was the summer of 2020, we raised a $5 million round and that was also from angels. It was Daniel Gross, Nat Friedman. Pablo 26:49 What was traction like at that point? Dylan 26:51 We had passed a million dollars in ARR. We had decent traction at that time. We were growing pretty quickly and we're still a super small team. I think we were sub10 people. Then it was a year and a half later that we raised the A and then a couple months later our B and then most recently our C. The 2020 is really when things started to compound and the trajectory really changed. Pablo 27:16 What do you attribute that to? Was that just basically the stack getting good enough, getting over that bar where all of a sudden people could just build all these applications on top of? Dylan 27:24 Yeah, it was a combination. It was macro, market adoption, matured in part because the tech was getting better, right? As the tech gets better and early adopters demonstrate its capability out in market that pulls the market forward and then more people adopt. That's just been happening. Then of course 2022, this whole… Pablo 27:46 Gen AI wave. Dylan 27:47 All this innovation has just taken that market adoption and accelerated it more than I think anyone could have imagined even 18 months, two years ago. It's hard to remember beginning of 2023, there's one mainstream LM, and now there's thousands and there's open-source models and there's tons of companies in this space. This acceleration we're in the middle of is very new, but I think it will continue to be honest. Pablo 28:14 Perfect, well, let's stop it there. Let me just end with the two questions that we always end on. The first question is, when did you first feel like you had true product market fit? Finding True PMF Dylan 28:23 It's interesting, right, because for us, I knew that there was market demand even before we started building the thing, whereas with a lot of companies you don't know, hey, are people going to want this service or is there going to be product market fit for this type of product or service? For us, similar to self-driving cars, you know that Level 5 self-driving cars will have product market fit even though they don't exist yet. Now, of course, if they cost a million dollars per vehicle, the addressable market is probably much smaller versus if they cost $10,000 or if there were just robotaxis everywhere. There is a question of economics, but there's not a question of market demand, unit economics, but not market demand. For us, I always had conviction that there was market demand and product market fit for what we were building. That being said, I think that, we serve many different use cases, so we have people building so many different types of things with our API. There's still opportunity for us to improve product market fit for what people are predominantly choosing to build. That's really the work right now. I would say it never stops because the market's changing what customers want to do and users want to do is always evolving, but I always had conviction that there was product market fit for what we were building. Pablo 29:33 Final question, if you could go back to when you were just starting Assembly AI back in 2016, 2017 with one piece of advice for yourself, what would that be? Dylan 29:42 I One Piece of Advice Dylan 29:43 would have advised myself to take a longer-term view in the first year, especially coming out of YC. Slow down. Now that you're done with YC, what do you want to accomplish over the next two years? Where do you want to be in two years and then just work towards that. I think, like I was mentioning earlier, coming out of YC, you're trained within those 90 days to just be like, the clock is ticking, the clock is ticking, the clock is ticking. You're always thinking that when you're running a startup because there are macro factors, right? Other people are competing in the same space. You can’t just operate on your own timeline. I would have given myself, urged myself to take more permission to think longer term about what I want to have accomplished within a year or two years and then break that back down into quarterly goals and use that as a foundation of hiring and prioritization versus trying to just make quick progress as quickly as possible. Pablo 30:40 Perfect, well, thanks a lot, Dylan. I mean, you're now one of the leading companies in what is probably the hottest sector. Really appreciate you taking us through the early days and what it took for you to build what's now having so much success. Appreciate you jumping on the show. Dylan 30:55 Yeah, thanks for having me on. Appreciate it. Pablo 30:58 I just gave you content that you liked so much, you actually listened to the end. Guess what? You didn't pay a single dollar. Not only that, I didn't even put any ads in your face so you just got a bunch of content for free. Now that I've delivered that value, I'm asking for something in return. Open your app, open Apple Podcasts, open Spotify, open whatever app you use to listen to this and hit that follow button. It's actually going to help you because it's going to help you make sure you don't miss out on the next episode, which you like so much that you listen to the whole thing.