A Product Market Fit Show | Startup Podcast for Founders
00:00:00 · Dylan, Pablo Srugo
[00:00:00] S0: 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. [00:00:21] S1: 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. [00:00:38] S0: Yeah, thanks for having me on. [00:00:40] S1: 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.
00:01:13 · Dylan, Pablo Srugo
[00:01:13] S0: I 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… [00:02:45] S1: It's terrible. Yeah, I don't understand what's going on.
00:02:45 · Pablo Srugo, Dylan
[00:02:45] S1: It's terrible. Yeah, I don't understand what's going on. [00:02:49] S0: I got really into this whole idea of natural language interfaces, especially over voice. Then I also totally saw where machine learning was headed. [00:02:59] S1: 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. [00:03:05] S0: 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. [00:03:34] S1: 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? [00:03:51] S0: 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.
00:03:51 · Dylan, Pablo Srugo
[00:03:51] S0: 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. [00:04:15] S1: It had some believers but wasn't necessarily a mainstream thing yet, but you were one of those, let's say, early believers. [00:04:21] S0: 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.
00:04:21 · Dylan, Pablo Srugo
[00:04:21] S0: 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. [00:05:19] S1: 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?
00:05:19 · Pablo Srugo, Dylan
[00:05:19] S1: 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? [00:05:34] S0: As a 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. [00:06:16] S1: Right, so it’s text to code, yeah.