🎧🍌 Giving an AI a Computer with Rahul Sonwalker, Founder and CEO of Julius AI
Inside the "Ligma Johnson" prank, acquiring users with ChatGPT plugins, why Rahul gives his number to customers, AI in classroom, the power of small teams, and why Julius uses so many models
The episode is brought to you by Attio
Attio is the next generation of CRM.
It’s powerful, flexible, and easily configurable.
Join OpenAI, Replicate, ElevenLabs, and more and try it today
To inquire about sponsoring future episodes, click here.
Rahul Sonwalkar is the founder and CEO of Julius, an AI data scientist with over 600k users.
Rahul takes us inside his epic “Ligma Johnson” prank where he pretended to be fired from Twitter the day Elon acquired the company. He then goes inside his journey of building Julius, sharing lessons learned along the way and his vision for the product.
Highlights include using ChatGPT plugins to acquire their first 10k users, what happens when you give an AI a computer, the importance of speed of execution, how to avoid analysis paralysis, why talking to customers is so valuable, why they love when you ship new features, what he thinks happens to all the AI startups, and using Twitter to grow his network.
Timestamps to jump in:
04:16 The "Ligma Johnson" prank
09:02 Meeting Elon
13:10 Hackathons in college
14:30 Scraping emails from Hacker News to get internships
16:25 Using Twitter to learn and meet people
22:26 Lessons from his first failed startup
26:19 Taking too long to quit Big Tech + his trucking startup
32:12 Convincing Guillermo Rauch to invest with speed of execution
34:48 How to avoid analysis paralysis
36:15 Building Julius, the AI data scientist
39:25 How the COO of a hot tub company uses Julius
40:40 Professor embracing AI, using Julius to teach his class
42:47 Iterating on early versions of the product
44:42 PMF is as much about the market as the product
45:08 Building dozens of ChatGPT plugins to acquire Julius’ first users
45:41 Using dev API keys and missing the first paying customers
49:22 Talking to hundreds of early customers
50:10 Why customers love when you ship new features every week
52:14 The power of Julius’ small team
54:38 Why Rahul gives his number to customers
57:47 How to avoid idea backlogs
59:27 Why Julius tests so many models
1:01:44 Why it feels great when people love your product
1:03:43 AI will write more code than humans
1:06:33 Giving an AI a computer
1:11:05 What happens to all the AI startups?
1:12:36 Why you have to Ride the Tiger
1:16:43 How NVIDIA beat 89 other graphics card startups
1:20:14 Building a moat as a startup
1:22:50 Rahul’s favorite AI companies
1:25:11 Why Julius’ changes UI components based on the use case
1:27:42 Benefits of lifting
1:29:54 Why Rahul loves SF
1:31:38 The early days of Microsoft
Referenced:
Find Rahul on Twitter and LinkedIn.
Transcript
Find transcripts of all prior episodes here.
Turner Novak:
Rahul, how's it going? Welcome to the show.
Rahul Sonwalker:
Thanks, Turner, for having me.
Turner Novak:
So I think most people know you as the guy who faked getting fired from Twitter, right after Elon took over. And maybe some people are not familiar with the story, but can you just take us inside that day, tell us what happened? I think that'd be fun to talk about first.
Rahul Sonwalker:
Totally. Well, I guess for context, I live in San Francisco. I hang out on Twitter quite a lot, as a user. I've been on Twitter for a few years now, and of course I've been a big Elon fanboy since forever. He's doing really cool stuff with rockets and with cars. And in San Francisco, I happen to live walking distance from the Twitter headquarters, or what used to be the Twitter headquarters. And my gym is actually in the Twitter building.
So every morning I wake up, I walk to the gym, I work out, and then I head to the office. I believe a year and a half ago, on my way to the gym, there were just a bunch of cameras outside, a bunch of TV reporters waiting for something to happen. And I just found that to be really hilarious, that this was happening. And I walk in, I'm sure these reporters, they definitely saw me walk in, into the gym. I work out.
And then towards the end of my workout, I just texted my friend group chat saying, "Hey, it would be really cool if we prank these guys. If we just walk out with a box and see if they put us on TV."
And I first texted my roommate, he's not down to come through with a box. I text my other friend, Daniel, who is... you know that meme of your unemployed friend at 1:00 P.M. on a Wednesday? He had that vibe back then. He was just hanging out, having fun. So I text him on a Wednesday, or a Friday, to come to my gym with a box, and it would be really funny if we just walk out with it.
And he shows up with two boxes and we walk out, and the cameras, they fall for it. They ask us a bunch of questions. We don't prep any lines, we don't prep what to say. We just wing it. And I said, "My name is Rahul Ligma," and he said his name is Daniel Johnson. And it just, out of coincidence, became Ligma Johnson.
Turner Novak:
Oh, you didn't plan that? I thought you guys planned that.
Rahul Sonwalker:
No, it was all coincidence. And so, I get followed for the next few blocks as I'm walking home, and then they eventually let go of us. I get back to my apartment and I realize that the thing blew up, literally 10 minutes after it happened.
Turner Novak:
Wow, okay. That was fast.
Rahul Sonwalker:
Yeah, and honestly, I didn't expect it to blow up, because I was just out there having fun with my friends, and I thought a few of my friends would find it funny, but it became really big within the next 30 minutes or so. My phone's blowing up, somebody gave my phone number to the reporters, they're calling to ask me if I actually work at Twitter, or I don't.
I was going through YC back then and I walk into my office hours, literally an hour after the thing happened, and my partner asks me, "Hey, are you working at Twitter while you're doing the startup?" And I said, "No, I have to tell you, this is not real. This is a prank. I thought this would be funny if I did it, and my friends would find it funny." Really, that was the whole reason to just have fun with my friends.
And then it really blew up after Elon tweeted about it, and the day became really crazy.
Turner Novak:
At what point was that? Was this, like, two, three, four hours later, or pretty quickly or-
Rahul Sonwalker:
It was picking up within the first 30 minutes, but two hours in, it went from just Twitter to actual TV news.
Turner Novak:
Like real life.
Rahul Sonwalker:
Yeah. Like Daily Mail in the UK. And I started getting recognized on the street. So I'm walking over to get lunch, I get recognized. I walk home in the evening, I get recognized, and it kind of blew up from there.
Turner Novak:
Yeah, I remember when it was all going down, 'cause I think we were kind of just Twitter friends, basically. I don't think we'd ever met in person, or even really talked before. Same with Daniel too, that you were with. I was just kind of Twitter friends with him. I didn't actually know what he looked like, actually.
And I remember when this whole thing was happening, I think I DM'd him and was like, "Is this you? Are you this Daniel Johnson guy?" And he's like, "Yeah, don't... like, it's not public yet. There's no connection, so please keep it quiet." I remember that.
I remember just, it was probably one of my top favorite just trolls ever. It was so funny how big it got. The videos, all the memes from it were great. Just took over the internet for that day.
Rahul Sonwalker:
Yeah, thank you. We actually got a chance to meet Elon a few times after, and meeting him was super cool. He is super high energy in person, even more high energy in person than he is in his videos or on Twitter. And the story of how we met him was also really interesting. We just showed up to the office and we said, "Hey, can we meet Elon?" And they made us wait 20 minutes, and then they took us upstairs and they said, "All right, here's Elon." So yeah, he's pretty cool.
Turner Novak:
What you guys talk about?
Rahul Sonwalker:
We talked about two minutes about the prank and then the remaining 40 minutes about what's going on on Twitter. Back then he was really concerned about how many lines of scholar code there is on Twitter, in the Twitter code base. It was really cool as a founder to know that he has that level of detail and insight in a company that he just bought two days ago.
Turner Novak:
Oh, this was two days... This was two days after the acquisition closed?
Rahul Sonwalker:
The thing happened on the day the acquisition closed; the morning of.
Turner Novak:
Okay, so these reporters were ready for just a fire sale, whatever.
Rahul Sonwalker:
I guess so.
Turner Novak:
Yeah, fascinating.
Rahul Sonwalker:
And he knew that level of detail, so that was really, really cool.
Turner Novak:
Yeah, I remember the whole thing about it. It kind of became a meme of printing off the code, like, "Show me how much code you've shipped," or something. Which, if you just think, printing off code, what a stupid exercise, but it does show a decent amount. Is it a big stack of paper? Is there one piece of paper? It shows how big it is and also what you've contributed, too.
Rahul Sonwalker:
Yeah.
Turner Novak:
And then I think the thing that I thought was pretty cool was... We'll get into Julius the company that you're kind of building, but you kinda used that whole event to just kind jumpstart the initial seeds of distribution for everything that you're doing now. Is that a fair way to think about it?
Rahul Sonwalker:
I think it definitely helped me meet a lot of really cool people that I got a chance to work with, that give me good feedback and definitely a platform where I can share my ideas.
Turner Novak:
So going kind of all the way back to when you first moved to the US, you were a teenager. How old were you?
Rahul Sonwalker:
I was 18.
Turner Novak:
And you came here for school?
Rahul Sonwalker:
Yeah.
Turner Novak:
I think what stands out to me, just from understanding and learning your story, you were super determined and focused on accomplishing your goals that you had at such a young age.
How did you know that you wanted to get a job at a start-up, work in tech, move to San Francisco? Was there a line of thinking that solidified that for you early on?
Rahul Sonwalker:
Yeah, totally. Well, I watched The Social Network, and this was before I was old enough to even use Facebook legally.
Turner Novak:
Okay.
Rahul Sonwalker:
But I watched the movie, and just the idea of building something with code and giving it to millions and millions of people, and have those people use it and be known for it, is really cool. And that kind of took me down on this path of, "What is startups? How does one start a startup? How do you build a website? What is code?"
And I just went down that rabbit hole throughout high school and I found a way to get to the US by taking the SATs and the ACTs, and I took the SAT when my English comprehension skills weren't as good. So I got a 1260 on the SAT.
Turner Novak:
Is that good?
Rahul Sonwalker:
That's out of 1600, so that's not good, if you want to go to a good school. But I got an 800 on math and then 460 on critical reading. So math was pretty easy. Critical reading was pretty hard, but somebody on the internet told me I should do the ACT because it has science, it has grammar, which is so more logical. And I was able to get a much better score, and find a school to basically give me a full tuition scholarship, to move to the US.
Turner Novak:
And then I think even when you were in college, you were doing hackathons. I think you won at least one. That's what I saw.
Rahul Sonwalker:
Yeah.
Turner Novak:
Did you do a lot of hackathons?
Rahul Sonwalker:
I did. That's the only way I learned. So I found that, in school, most of the learning was not super applicable to building cool things. You learn things that are... In computer science, you might end up learning things that are completely outdated and you don't need to know them.
But I basically found that everyone around me was not at the same level of ambition or ideas that I was at. So people just wanted to graduate and go work at State Farm for the next 20 years, or go work at Allstate or Raytheon. They weren't into building startups or building cool things, cool projects.
So I started going to hackathons, and attending hackathons on the weekends. And that's where I met people who were more like me, who wanted to build cool things, work at cool companies, start their own startup, and I got those people to refer me to the companies they had interned at or refer me to startups where I could apply, interview, show my skills and get an internship.
Turner Novak:
And I think you did get, at least according to LinkedIn, you got an internship at Facebook. Was that the first kind of big one that you landed?
Rahul Sonwalker:
There's a couple that are no longer on my LinkedIn, just to keep it a little organized.
Turner Novak:
Keep it clean, okay. Well, I think you were telling me before you had a really, in my opinion, pretty smart strategy that you used to outreach people. Can you talk through that?
Rahul Sonwalker:
Totally. So, I believe, freshman year of college, I cold emailed Ivan of Notion.
Turner Novak:
Who is that, just for people who don't know.
Rahul Sonwalker:
Yeah. Notion is this really cool productivity app that lets you take notes, create databases, share the notes with people that you work with, or people you like. It's really popular among knowledge workers.
Turner Novak:
And you said he responded?
Rahul Sonwalker:
He responded. He said, "Hey, thanks for reaching out. Appreciate the gumption, but you don't have the experience we're looking for."
So I basically found smaller startups to give me an internship. And eventually, by junior year, I was scraping Hacker News for all their hiring posts from back in 2010 and getting all the emails from those hiring posts, and then mass emailing a bunch of companies and recruiters and founders asking for internships. And I found somebody to refer me to Facebook.
I interned at Facebook, doing misinformation stuff. So I was an engineer on the team that stops the spread of misinformation. Which is kind of ironic because just a few years after that I was on social media, or on national TV spreading misinformation, so.
Turner Novak:
You dissected the playbook of how to do it.
Rahul Sonwalker:
Yeah, exactly. But I did Facebook, then I did an internship at Uber after that.
Turner Novak:
Nice. And then you graduated and then you moved to San Francisco, full-time.
Rahul Sonwalker:
Exactly. I graduated a semester early. I put all my stuff in a car, drove cross-country to San Francisco, moved into an apartment, and then two weeks after, COVID happened, and everything shut down. And that was a very interesting time, because that period introduced me to Twitter. I was on Twitter, replying to cool people, building cool things, cool founders that you read about in the news, and they will sometimes Like your tweet and reply to you and you can DM these people and people respond.
Turner Novak:
So this is like what you were already doing, but you just unlock the superpower of, instead of having to scrape them from Hacker News, scrape the emails. You could just slide in their DMs.
Rahul Sonwalker:
Exactly. And it also compounds, because the more you post, the more you show up in people's feeds and people think you're interesting and you have interesting ideas, and they follow you back and then there's more social proof, and you sort of compound that over time.
Turner Novak:
Yeah, I remember that's kind of how we met. I don't remember the exact time, but just through that, kind of got to know you and then-
Rahul Sonwalker:
Well, you were an early believer, 'cause we knew each other before the pop-off.
Turner Novak:
Yeah, before the pop-off.
Rahul Sonwalker:
So you have a great eye for early talent. I guess.
Turner Novak:
That's why I'm a VC.
Rahul Sonwalker:
I love that.
Turner Novak:
Can't build anything. So that's why I'm not actually a founder.
Rahul Sonwalker:
You're in the early stage. You're the pre, pre-seed.
Turner Novak:
Yeah. So then you've definitely benefited from using Twitter. Can you kind of explain, if somebody's never used Twitter or they're on the fence, is that even worth it? What does it do for you? Just really leaning into it and growing it and using it to your advantage? Can you just talk us through maybe the benefits and/or how you do it? If you were to do it again, how would you make a name for yourself on Twitter?
Rahul Sonwalker:
Totally. Well, I guess I can share what works for me and why I use Twitter. I think Twitter is a really great place to find a subgroup of people that you connect with, and they could be across the world, building really cool things or doing the same thing as you. And an example of this is, I watch football a lot.
Turner Novak:
Is this soccer? American football?
Rahul Sonwalker:
American, yeah, I'm a big Cowboys fan.
Turner Novak:
That's cool. American football. Okay. Oh, you're a Cowboys fan. Oh, man, I'm sorry.
Rahul Sonwalker:
Are you Lions or-
Turner Novak:
I mean, I just know it's hard to be a Cowboys fan.
Rahul Sonwalker:
It is. It's not easy. It builds character though.
Turner Novak:
I saw this one amazing meme on Instagram. It was like, there's been more solar eclipses in the last two years than Cowboys playoff wins or something.
Rahul Sonwalker:
No way.
Turner Novak:
It was something like that.
Rahul Sonwalker:
That doesn't surprise me. The last time the Cowboys won a Super Bowl, I wasn't even born.
Turner Novak:
Yeah. So you use Twitter to follow football? Is that where you were going?
Rahul Sonwalker:
Well, so I basically... On Twitter you can see football fans and football is really big in America. But there's these small subgroups of fans in Canada, in Japan, who also follow football, and they reply to these tweets.
But it's sort of like a way for you to meet people who share the same interests as you and build that community, and connect with those people. For me, it was partly football, but also partly startups, and people doing cool stuff with startups and building cool things.
So I basically used that to follow really good founders that I admired, and thinkers, and builders, and reply to them, but also keep it lighthearted. You don't want to be super serious at all times. You always want to be having fun, because if you're super serious at all the times, then somebody who is just having fun will beat you.
I think Turner, you're a great example of this, 'cause you're in Michigan, I believe, and you're just the funniest VC on all of Twitter and you have your own fund now. There's so many people out in Michigan, or some areas of the world that are not tech centers, and they just keep lurking and kind of wish that they were part of the action. But then there's Turner who is just out there, connecting with people who are in San Francisco, and connecting them on a level like your friend in San Francisco would. You know the jokes, you know the gotchas, the insights. You're just out here sharing all that. So I was something similar, but I was already in San Francisco. It was just COVID lockdown.
Turner Novak:
And then what benefits has it brought you, just in terms of you started a company, I guess, around the same time after... What all have you gotten from being a presence on Twitter?
Rahul Sonwalker:
I've definitely met some of the smartest people I know today, off of Twitter. I've also met most of the Julius team, either on Twitter, on Clubhouse, or at Hackathons. In fact, that's the whole Julius team. I met some of them on Twitter. I met some of them... One of them on Clubhouse, and one of them at a Hackathon. So yeah, I think it's a great way to meet people.
Turner Novak:
Wow, that's not an app, or a word you hear much lately, Clubhouse.
Rahul Sonwalker:
Yeah, I know. Yeah, it was back in 2021 when we met.
Turner Novak:
Okay. Oh, you were using it in 2021, too.
Rahul Sonwalker:
I was. I was one of the suckers who got in very late.
Turner Novak:
I mean, it was still... Elon did his big thing in, I think, it was January of 2021, if I'm remembering. So it was actually probably like the peak. Peak Clubhouse was 2021.
Rahul Sonwalker:
Yeah, I think we met February.
Turner Novak:
Yeah, those are fun times.
And then I think, maybe you mentioned earlier, maybe it was while we were talking beforehand, you did YC, and this was before the big Twitter prank. What was the whole process with YC? Did you know what you were going to build? Were you like, "I want to start a company," you're working at Uber beforehand. What was that whole process like of leaving, starting something, and then doing YC?
Rahul Sonwalker:
Yeah, so I basically spent all of COVID trying to meet cool people, you know, online, learn from their ideas and build ideas of my own. So one of my first ideas was, I was listening to a ton of podcasts, watching a ton of YouTube videos.
And what I found was that an interesting part of watching podcasts on YouTube was, you can actually scroll down to the comments and get interesting questions, talk to other people listening to the same podcast. People will post jokes, or they will expand on certain things discussed in the podcast and you can reply to them. And there was this additional content after the podcast, in the comments, and that was missing on the Apple podcast app and the Spotify app, where there was no comment section.
So my first idea was, can you build a comment section for podcasts? And it didn't take off. I built it. It was called PodcastComments.com. It didn't take off, it didn't go anywhere. But regardless, it was really fun to build it because that's something I wanted, and I wanted to use, and I learned a lot.
I learned for the first time, how to make a mobile app. I had never made a mobile app before. So how to make a mobile app, how much it sucks to work with iOS. And it was very useful.
Turner Novak:
And you used it, and did other people use it?
Rahul Sonwalker:
Well, the thing is that only I used it.
Turner Novak:
Okay, so you were commenting.
Rahul Sonwalker:
Yeah, usually, you want to... One thing I realized was that there's way more consumers than creators. For every one person who is creating content, there are 20 or 50 other people who are just consuming. And that's true for most social platforms. Twitter, there's one to 20 ratio, probably, of people posting tweets to just people lurking and reading tweets. Same with YouTube, maybe even more. One to a thousand, perhaps.
Turner Novak:
Yeah. YouTube is tough, 'cause, I mean, Twitter, it's a keyboard. You type it in, it takes two seconds. YouTube, it's like, you got to make a video, you got to edit a video, upload it, make a thumbnail. No one knows, making the thumbnail on YouTube is half of the process, basically.
Rahul Sonwalker:
Yeah. Wow. Yeah, totally. I've tried to make some joyous tutorials.
Rahul Sonwalker:
Wow. Yeah, totally. I've tried to make some joyous tutorials and I use Canva for it, but they're just nowhere as good as Mr. Beast's thumbnails.
Turner Novak:
Yeah, I need to up my thumbnail game. Probably like a third of the people who listen to the podcast listen on YouTube and I mean, my YouTube thumbnails suck. I got to up my game. If you're listening to this and you're like, "Turner, I agree, your thumbnail suck and I think I can do a better job." Let me know because I need help.
Rahul Sonwalker:
Totally.
Turner Novak:
One thing I thought was really interesting, I kind of talked to a couple of people before this. I got some feedback on good questions to ask you.
So one of the people I talked to was Guillermo Rauch, founder of Vercel. So he's a huge fan. I think he actually was one of your angel investors when you raised a little bit of angel money, he said that a big part of your pitch was the speed of execution. He just said, you guys move really, really fast. And he thought that was really impressive.
How do you show that in a pitch? You said it was actually integrated in your pitch was how quick you moved and it kind of convinced him that he needed to back you.
Rahul Sonwalker:
So I guess I can maybe give a little bit of context about what I was building back when we pitched Guillermo, and this was right after I had quit Uber in 2022. And I honestly took way too long to quit because I was building these things on the side and sort of launching them and they weren't taking off.
And my initial plan was, okay, I'm going to continue working at Uber on machine learning stuff, and when I have something that's sort of taking off, I'm going to quit my job and go full-time on it. And honestly, that didn't happen.
So after about a year of just nights and weekends, I quit because I just got tired of half assing it. And my first idea after I quit was, you know how you book an Uber, you can see on your app where the car is, you can see where it's moving, what's the ETA, when it's going to pick you up, when it's going to drop you off.
All of that doesn't exist for logistics. And there's thousands and thousands of trucks that are moving every day today in the US that are moving produce, moving goods, moving packages that the people, the brokers and shippers who book these trucks, have no idea where they are.
The only way they figure out is by picking up the phone and calling truck drivers and asking truck drivers, "Hey, are you going to make it at 4:00 PM or can I put you on the 8:00 PM slot?" And this creates a ton of inefficiency. Imagine if you couldn't know where your Uber is on the Uber app, that would be really hard.
Turner Novak:
You could call the driver every five minutes. "Hey, are you close? Where are you?"
Rahul Sonwalker:
Exactly. Or where are you picking me up? What street are you on? That would just get so complicated. So the first idea was, well, the government mandates that all these trucks have these things called ELDs. It mandates that-
Turner Novak:
Like tracking devices, right?
Rahul Sonwalker:
Exactly. Tracking devices. And the idea was, "Hey, can we take all that data, aggregate it and make an experience as smooth as Uber and give it to shippers and brokers who book these trucks to let them see on a site where a truck is?"
Turner Novak:
Yeah, it seems like a cool idea.
Rahul Sonwalker:
It seems like a cool idea in theory. We built that in a few weeks. We just hacked through it and got it working.
And the way we did it was that there was this long tail of ELD providers. So what we would do is every time we would get a new truck load, we will figure out how to reverse engineer that ELDs API on the fly because there's hundreds of these ELD providers and we can't build all the API integrations upfront.
So the plan was, "Hey, let's start taking truck loads. And as we get new drivers and new hardware, we can figure out the API for it."
Turner Novak:
I mean, you have to do that instantly to make them show up on the map?
Rahul Sonwalker:
So luckily for trucks, they move over days. So if we have six-hour heads up, a truck is moving from Nashville to Los Angeles or Long Beach.
Turner Novak:
So you'd have six hours, the course of the freight trip to build the integration?
Rahul Sonwalker:
Yeah. Well, the thing was that we had nothing else going on. It's not like we had thousands of loads.
So we did that for the first 10, 12 loads and we sort of tried to keep scaling that up, but soon we realized was that people, we were solving the right problem, the people really wanted it because we're giving them so much efficiency that solves scheduling, planning, delivery, all these different things. But what was really hard was to convince the truck drivers who were 60-year-old guys in Kentucky to share their live location at all times. That became a real challenge and we had to basically call, we had to beg them. We tried all different things to get these drivers to accept.
Turner Novak:
And so your customers were people that needed to track the freight. Those are the ones who were probably paying you, but the truck drivers needed to just participate in this?
Rahul Sonwalker:
Exactly. So for some context, you have the Amazons or the Walmarts of the world who own the truck and who are the shipper, and they're the receivers too. So they own the truck. The driver is an employee or contractor. And so in that case, the driver, they know where the truck is, but there's this whole other side of logistics where, for example, Flexport, they help you move freight across the ocean. The way they do it is they find a ship that's moving from point A to point B, get your load on that ship and bring it to you.
So there's a lot of quarterbacking happening and sometimes these trucks are moving loads for multiple shippers to the same receiver, sometimes multiple receivers, but the same shipper, all kinds of complexity.
And these drivers, they just do jobs, they pick up a job and then they move the load and then go to the next job. And it became a real challenge to convince these drivers to share their location with us. And we tried all sorts of workarounds.
But to answer your original question, how did Guillermo know we were moving fast or shipping fast is I actually met him at an event in San Francisco and everyone was just hovering him, asking him about Vercel, and we were using Versel back then.
I'm a huge fan of Vercel. It just makes building front end fast, performing front end so easy for someone like me who completely sucks at front end. If Alex is watching this podcast right now, he knows how much I suck at front end. He's the one who builds most of the front end for Julius. He's our founding engineer.
Well, so Guillermo, I met him at the event. I asked him for his contact information and then I sent him a link of a live truck load moving on live trucks. "Hey, you can follow this load going from point A to point B, and it's a real truck moving right now." And we build this whole thing in six weeks or seven weeks. And then we told him how much sales we have closed, and he was super impressed and he supported us.
Turner Novak:
That's amazing. Why trucking? It's kind of a random, was it just one of many ideas that you worked on and then moving on from it when it didn't quite work?
Rahul Sonwalker:
I left Uber and I knew how, to give you a smooth Uber ride, there's so much that happens in the background from dispatching to pricing to taxes. How do you collect taxes on an Uber ride? All that is really complex. But the experience of the driver and the rider is really seamless. And I was looking to apply that to a space that didn't have that. And I looked at trucking and it seemed like it didn't have any of that, which I think I was right about. But I vastly underestimated how hard it would be to convince truck drivers to share their location.
Turner Novak:
So maybe, I'm trying to think of what would be a good lesson there for other founders. It's just you got to think through will people actually use the product or will all the touch points, does it solve problems for them? Is it urgent enough?
Rahul Sonwalker:
I think the incentives need to be aligned for everyone. Because we had so many different pieces we had to convince in many of these loads, there were brokers who would coordinate the load between a driver and truck and a shipper, and then there was a driver who had very little incentive to share his location because they have driven a truck for 40 years a certain way, and they have a certain pattern where they'll stop at a Walmart the night before and sleep overnight. So we didn't consider all the pieces.
It was also, I think it was also something we needed to do to, no, it would've been really hard if I think of 2022 me, to come up with all the possible ways this could go wrong and then plan for that ahead of time as opposed to just doing it, spending six months of my time trying to make it work and then eventually move, go to something else.
Turner Novak:
So sometimes it helps to just do the wrong thing just to know why it's wrong or why it won't work.
Rahul Sonwalker:
I think I see a lot of my founder friends get stuck in this analysis paralysis of trying to come up with the idea that is completely bulletproof, can never go wrong, the execution plan is completely de-risked. The team is the right team to build what you are trying to build. And there's a lot of money that you can raise on that idea. You want all of that.
But to be honest, there's no idea like that. The best way to get validation is to do something. Action produces information, and then take the information and then act on it.
Turner Novak:
Yeah, I think the only thing you can really control as an early stage founder is just yourself, how quick you move, how you fast execute on things.
To the point earlier, it's always good to see as someone outside is like, "Wow, these guys are moving super fast." I mean, I've invested in, I don't know, 70, 80 startups, literally not a single one has gone exactly how anyone thought it would. So it happens. It's not an easy thing.
Sometimes it happens faster, sometimes it happens slower. It's never what you thought, half the time it's totally different business or maybe it's similar customer, but you solve their problem in an entirely different way and your business model, your product looks different.
So to that point, how did you go from software to track trucking trucks to you're building Julius, it's like an AI native data analysis tool. I mean, I'll let you explain it, but what was that journey like of going from point A to point B? Or maybe it was point C and D and E in many different points?
Rahul Sonwalker:
Yeah, totally. I will first give a two sentence description of what Julius is so that people can orient their mind around what it does. And then I'll walk you guys through the path and how we got there. But today, Julius is an AI data scientist that helps you analyze spreadsheets. So if you have a spreadsheet, you can feed it to Julius and have the AI analyze it for you, give you insights, create data visualizations, all with just simple English.
Turner Novak:
So then how do you take, what's the benefit then if I would normally be a Tableau user or use spreadsheets in Excel? What gets me excited when I hear or use Julius I'm like, "Holy shit, this is cool." What's usually the thing?
Rahul Sonwalker:
Well, if you're a Tableau or Excel user, you can do all the things you can do in Tableau or Excel just much faster and with just simple English commands. So instead of grappling your way in Excel or Tableau for an hour, you can do it in Julius within 5 or 10 minutes in just plain English.
But then there's way more people out there who don't know Tableau or Excel or Python or Jupyter Notebooks. They can bring their spreadsheets and have the AI give them insights about that spreadsheet. So some of our users are people who work heavily with spreadsheets and Excel files, but lack the technical expertise to dive into that data and get the insights out of it. And they use Julius as their data scientist in their pocket to do that.
Turner Novak:
So what's kind of a prime example of use case? What might my role be and what might I be uploading and getting it out of it and using it for just to get people in the mindset of how it kind of works.
Rahul Sonwalker:
So there's a few different use cases for Julius. The most popular ones are finance.
So if you have financial data in a bunch of Excel files, you can upload all of those to Julius and the AI will crunch the numbers, give you the insights you're looking for, create the data visualizations within minutes.
If you work in academia or research, you can upload massive data sets that you've collected about your scientific experiments and the AI will run regression, ANOVA, all these different kinds of complex analysis that you normally would need a data scientist for, and AI will just do all of that for you within minutes.
Lastly, if you work in sales and marketing, you might have ad campaign data or marketing campaign data. You can use Julius to do complicated things like clustering and cohort analysis and campaign performance, all of this without needing to know the technical details of how to implement that or the details of how to learn those techniques, you can just have the AI act as your data scientist and do all that for you.
Turner Novak:
What's your favorite use case or favorite user persona that you've come across that someone's used it for?
Rahul Sonwalker:
Well, I'll give you two. So one of our users, he is the COO of a hot tub company in middle America.
Turner Novak:
You said hot tub company?
Rahul Sonwalker:
Hot tub company, yeah.
Turner Novak:
Okay.
Rahul Sonwalker:
And he uses Julius to analyze and forecast his sales. And I got on a call with him one day and he literally has a single Google sheet with each tab for each week's sales going back almost a decade. It's crazy.
And he just feeds that into Julius and has the AI do forecasting or some of the interesting things was that it was out of the box able to pick up things like seasonality, right? People don't buy hot tubs in the summer, but they buy hot tubs late in the fall.
Turner Novak:
Yeah, that's true.
Rahul Sonwalker:
It was able to pick up that seasonality just from that decade old data. And it blew his mind. So this was really cool. We thought there will be people in San Francisco who will use it, but then there's these users who found Julius and they really love Julius.
But there's also a group of academics who have really embraced Julius. And generally academics have been somewhat anti AI because it tends to seem like cheating in a classroom, in homework. But we have academics who found Julius on the internet and they're using Julius to teach a whole class.
There's a professor at Harvard Business School who's teaching who intro to data science to his business school students using Julius. And all these people will want to know the useful things you can do with data science, but they lack the expertise to do that. And they're learning all of that with Julius.
There's a professor at Rice University who is teaching AI financial analysis using Julius, and I was able to fly to Rice, be in the class, and just watch a classroom full of people learn Julius and financial analysis.
Turner Novak:
So his class is, he teaches the class and they use the product throughout the course to complete assignments and things?
Rahul Sonwalker:
Exactly, assignments. So up until this semester, he used Excel. And starting this semester he's been using Julius.
And if you look at his slides, it's like "Ask Julius to load this file and then ask Julius to do forecasting or ask Julius to do regression." And he has some prompting tips, but pretty much his whole slides are how to use Julius to do the things that he's teaching.
Turner Novak:
Yeah, I guess when I think about how you'd usually do a class like that, you probably have to learn this data analysis software. I can't even remember the one had to use once in one of my finance classes. I think it was called MATLAB.
Rahul Sonwalker:
Yeah.
Turner Novak:
Yeah. And I don't even think I did the assignment. I don't even think I did the project because I didn't feel like learning the software. I mean, it was my last semester. I was kind of already graduated, but...
Rahul Sonwalker:
I love that.
Turner Novak:
We're going to cut this part. I'm just kidding. We're not cutting that.
It brings up an interesting question. I think you mentioned you have about 600,000 people that use it, correct?
Rahul Sonwalker:
Yeah.
Turner Novak:
How'd you get people to start using this thing?
Rahul Sonwalker:
Totally. Well, so we went through a few different iterations of taking the data aggregation thing that we were building into Julius, a bunch of learnings along the way.
Our next thing that we tried, and one of the things that I was really passionate about was just AI, because I was spending my nights and weekends tinkering with language models, trying to get them to do interesting things, and trying to find ways on how we can incorporate that into what we're building.
So our first idea was can we take massive datasets and make it really easy to query those with SQL using just English? So you ask in English what you want out of the data, and the AI will write SQL for you and give you the insights.
Soon when we build that, we realize that these models are not as good at SQL as they are at Python. Python is just a way more powerful language with an ecosystem of modules that let you do amazing things like data visualizations, machine learning, regression, statistical analysis, all that out of the box.
And we sort of found our way into Julius about 38 weeks ago. I actually count the number of weeks since we launched Julius. So we are 38 weeks into Julius after multiple tries trying to build something people want. And we built Julius, launched it 38 weeks ago, and since then crossed over half a million users.
And how we got those users, well, initially we were launching on social media, trying to get people to use what we were building.
Turner Novak:
You had probably 20,000 followers or something. I don't know. Maybe I should have looked, but you had a little bit of a base to work with.
Rahul Sonwalker:
I did, yes. But one thing interesting I learned about PMF is that PMF is as much about the market as it is about the product. It's PMF, it's not just product fit, you have to find a product that meets the market. And what we found was that the users, the market that really wanted Julius was not on Twitter back then.
Turner Novak:
Really?
Rahul Sonwalker:
We actually hacked the Chat GPT plugin store to get our users. So we basically made a plugin on the Chat GPT plugin store that was called AI Data Analyst or something, or Chat with Excel, I believe. And we built, and we launched that, and for the first week or two weeks, we saw that people were using it.
It was pretty clear from the first or second day, but for the first two weeks, we realized that people were actually using it. And then we realized that people are trying to pay for it because we added a Stripe integration.
But it took us a week to realize that because we weren't getting the payments. Turns out the API key that we were using were Dev API keys, not production API keys. So people were trying to pay us and we just weren't getting the payment, which was really funny.
But in the Chat GPT plugin store, we found this initial group of highly intentful users who knew how to work with AI products, who understood prompting, who in that moment had a need, which was to analyze data, and they were looking up on the plugin store, how to analyze data, or how do I work with the spreadsheet in Chat GPT.
So we had a plugin for that, and then pretty soon we realized that the more plugins we have, the more users we get. So we took the same plugin and just made over a dozen plugins that did the exact same thing. We just called it a different thing. It was a plugin called Data Interpreter, which would rank higher than code interpreter on the plugin store. If you try to look up code interpreter on the plugin store, the Julius's plugin would show up first.
And that got us about the first 10 to 20,000 users until OpenAI said, "All right, guys, you're being a little too sneaky. No more playing with the Plugin store."
Turner Novak:
Yeah. Did you have the most plugins of any company of 10+ plugins?
Rahul Sonwalker:
I think so. So that was a scary point in a way that we realized that our top source of users was going to disappear overnight.
Turner Novak:
So was it like Julius was a separate thing, you had your plugins that lived in Chat GPT, were those completely separate products?
Rahul Sonwalker:
So the plugin is how users discover Julius. So when you get the plugin, you sign up on Julius.
So the plugin would trigger you to go to Julius, make an account, and then upload your files to Julius. And then in fact, most users would just stay on Julius and they would, because we had a better interface, we had built out the features that users needed to work with spreadsheets and actually do really convoluted analysis with spreadsheets. So some of them would still use us via the plugin, but many users would just stay on Julius because the plugin sent them there.
Turner Novak:
Yeah, I can see why that wouldn't be the greatest thing. I wouldn't be happy about that if I was Chat GPT.
Rahul Sonwalker:
I guess so, yeah, I'm really grateful for OpenAI for making really great models that we use, and they're doing amazing work and we're grateful for them, but I totally see where they're coming from.
But a couple of months in that was, we were iterating because we had this strong influx of users that were actually using what we were building because the things we built before, users would use them on the first day and then not come back.
Turner Novak:
Oh, so this was when you were doing Julius, but you weren't on the plugin store yet?
Rahul Sonwalker:
This was before we built Julius. There were a couple smaller things we built with SQL and people would play with it for a day and then just not come back the next day.
Turner Novak:
So there's just no retention.
Rahul Sonwalker:
Yeah, retention was zero, but the retention, there was at least some retention with Julius early on and more than we had ever seen before. So it felt like the greatest dopamine hit ever.
Turner Novak:
Yeah, I mean, it's awesome. People want what you're building or what you're selling. That's pretty cool. So that was really the first time it kind of started to work?
Rahul Sonwalker:
Yeah, exactly. And then we just talked to these users. We tried to get as many users as we can on a Zoom call and just ask them what they're trying to do and how can we build a better product for them and just figure it out from there. And within a month of constantly shipping a ton of code every day, iterating, fixing bugs, we knew we had a better product.
But then our main source of users was going to disappear overnight. And this was about a month or two into the life of Julius, about eight or nine weeks in. And I think we had 10,000 users back then.
And then we had to just figure out how to work with the users we have and then figure out how do we do marketing, reengaging these users, sending them emails, getting the word of mouth going, and then eventually building connection between the users and our social media, and talking about Julius and doing launches and getting more users from the launches
And just every week we launch a new feature, which first of all shows the user that this is a live product. We're constantly building new things and adding them to the product. And second of all, it reminds them that Julius exists and they should go use it.
So it helped us get that initial source of users and then we work our way from there to now over half a million users.
Turner Novak:
That's amazing. So you went from 10K to half a million. You skipped over this like it was no big deal. What were some of the monumental moments between those two times? Any big feature or product launches or marketing experience or experiments? Anything that really stands out as being a pretty big moment in that kind of journey?
Rahul Sonwalker:
When ChatGPT launched the GPT store, we tried the same strategy as plugins. It just didn't fit the same format, so we didn't get as many users from the GPT store as we did from the plugin store. But we basically kept trying a bunch of things.
And one thing about startups is that it's never like this. You don't go from 10 to 500, like this. It's more like this, and then a little bit of this, and then a little bit of this. If you zoom out, 38 weeks, it looks like this. From week to week, it's a little different.
A big, big part of it was talking to users and figuring out what they're trying to do and building a superior product experience for them and leveraging all the best models that are out there to do useful things.
We haven't invested in marketing yet a lot, but that's the goal. And hopefully, that's how we go from half a million to 5 million.
Turner Novak:
Yeah, that's fair. So one thing I think that you've done a really good job at not spending on marketing, you've been really capital efficient. What's been the secret to that? I mean, you have a super small team, a ton of people using it ...
Rahul Sonwalker:
I think very little marketing spend. Really, the only spend is salaries.
Today, we are four full-time people, including me. And we all write code every day. Even if we are talking to users or doing something that is not correlated, try to get a commit in. So that is very efficient. It helps everyone get a better sense of what's feasible in the product today and what could be feasible in the future.
If you are very far away from the code, you tend to propose ideas that are not feasible all the time, or you tend to think of things in a very wishy-washy way.
An example, this is like, what if we just add a button here? But then it turns out to add a button there, you have to go refactor this whole view and react and then rewrite some of the rendering and the padding and the margins or whatever. It's not easy as it looks, but it also helps the velocity going, helps the velocity of the product.
So we are pretty efficient in a way that we are really small, four full-time. All of us are in a room all day writing code and talking to users.
And the second thing is we let the users fund us and fund our innovation by paying for the product. So we have a limited free tier where users can come in and try Julius, get insights out of it. And if they find it useful, we ask them to pay us so that we can use the funds to build a better product for them.
Turner Novak:
So it's free to use, there's no official paywall, but you ask them to upgrade. Can you explain that? How does that work?
Rahul Sonwalker:
Yeah. Well, you get a certain amount of free usage every month. So if you're a very casual user, that's good enough for you. But if you're a power user, we paywall after a certain amount of usage and we ask you to pay. And when you pay, we're able to use those funds to build a better product for you.
Turner Novak:
Yeah, and I think one thing of your investors, Sandy Kory, told me to ask about this, the way you do your pricing packaging is super interesting, specifically the most expensive tier… you give your phone number! You're just like, "Call me, text me." Why did you do that?
Rahul Sonwalker:
It's a great way to talk to users, and it's a great way to filter out the feedback that you should prioritize. You can look at any product today and give a ton of unsolicited feedback.
If the PM of Slack reached out to me and asked me, "Hey, what's your feedback," and I could just fire off 50 pieces of feedback that I think would be cool to have, but maybe other users don't want it.
So talking to users is really important, and then being able to prioritize the feedback you get from users is really important.
So even today, users text me, call me all the time. It's a very candid way to share feedback with me that I can take it back to the team. You can text me at 10:00 PM on a Friday, and if I'm looking at my phone at that moment, I will reply to you. And my replies would be less formal. It wouldn't be, "Hello, thank you for reaching out." It would be, "Hey, Jake, let me see what I can do here," or, "Let me just limit for you," or, "I think you should try this prompt to get it to work." It's very casual, and turns out that the users like it. It removes the barrier of formality to reach out to you. They get to talk to the CEO, ask him questions.
Many times, users just reach out and tell us the cool things that they're doing with Julius. They just pay the highest tier plan to just tell me the cool things that they're doing.
So an example of this is there's a professor who teaches an undergrad class with Julius. His whole class uses Julius. And this spring semester, he did an open book test. And not only were the students allowed to look at the books, they were allowed to use Julius during the test to answer the questions.
Turner Novak:
Wow, that's a bold use of AI in the classroom right there.
Rahul Sonwalker:
He is really bold, yeah. And he texted us that experience and shared all of that. Not just open book, it's open Julius.
Turner Novak:
Yeah, that's awesome. I definitely had a class like that, where it was usually the capstone, highest level class in a track or whatever, like my capstone finance class. The teacher was like, "You can use your books, you can use your notes. I'll give you all the questions ahead of time. You can cheat however you want. You can use whatever method and come to the test and take the test."
I mean, it's kind of like how the real world is, right? You're not going to show up at work and be like, "Okay, you can't use the internet to code this." Why would you not use tools?
Rahul Sonwalker:
Yeah, 100%.
Turner Novak:
So Mike Knoop, he's one of the founders of Zapier, he said, "You guys will be brainstorming and then 12 hours later, you'll come back with a demo of the feature, the product." So do you think the way you've structured the teams and the way you've built the code base, does that allow you to have a certain product velocity and then philosophy on building products?
Rahul Sonwalker:
Yeah, I think our philosophy is primarily two things.
One is good ideas can come from anywhere, and if there's a good idea, we should implement it and give it a try. Not all good ideas work, but there's a good chance a good idea might work and make a real big difference, and we should give it a try as quickly as possible.
One of the things that suck is having an idea backlog. You always have this backlog of ideas that you are really passionate about and you are waiting to try them, and you never really end up trying them. That is not cool. You should try to get those ideas out of the door as soon as possible in the most minimal way possible.
The second thing is AI is moving really fast. There's new advancements being made every month. And if you're building these models or building with these models, you need to have the mind of a child that's always tinkering with things and playing with things and trying to understand how things work and trying to integrate all of that into your product, into the things that your users see in really useful ways. Because if you don't, then there's other people out there who will do that.
So with AI, I think you need that mindset to constantly play with the models, experiment with ideas, and integrate with those things as quickly as you can. An example of this is OpenAI just released a Gpt-4-turbo-04-09, and-
Turner Novak:
It's going to be outdated by the time this is published.
Rahul Sonwalker:
That's how fast AI is moving right now. But they released that two days ago, and within an hour we had that integrated, because we just play with models so much, swapping out models, testing them all the time in our free time, that we know when there's a new model, how do you make it work? How do you test if it actually breaks things or not? That mindset is already there.
So with Mike's example, he proposed an idea, which I can't share on the podcast, but it related using models in an interesting way to growth hack Julius to 5 million users from where we are today. And to me, that seemed like a really interesting idea. It was a very valid idea. And Mike is super sharp, he's awesome. He's built a great company.
So there was no reason to drop everything that we are doing and go build that idea. And we basically did that. And in 12 hours we had a prototype, and then over the weekend we started getting feedback on the idea. So it was kind of a no-brainer.
Turner Novak:
Yeah, I mean, and talk about using marketing ideas, using models. I mean, Zapier had the ultimate automated ... AI wasn't really a thing back 10 years ago. But they had the ultimate automated marketing strategy, basically anytime. They just built a bunch of different features and example use cases of the product and people would Google for it. And they'd just find the products and start using it. I mean, it was beautiful, their go to market.
Rahul Sonwalker:
It's so amazing. Yeah, I think both Zapier and Notion, incredibly capital efficient, bottoms up, our loved by the community, great products, great founders.
Turner Novak:
Yeah, and I think you seem to do a really good job of elevating your community. It's kind of like an overused word, but I just see sometimes you'll tweet something like, "Julius seen in the wild," at the XYZ newsletter. And you give your customers and your users a shout-out. Is that intentional? Does that help you think?
Rahul Sonwalker:
A big part of it is morale. When you build something, by default, nobody cares, even if you're probably Elon Musk; maybe for Elon, people will care.
Turner Novak:
People also just shit on anything he does also.
Rahul Sonwalker:
That's true.
Turner Novak:
Everyone has haters.
Rahul Sonwalker:
Everyone has haters, yeah. Every single product has haters. iPhone has haters, and Tesla has haters, and there's not a single great product that doesn't have hater. Like Vercel, Guillermo, his company, I'm a big fan of him. I was a user before he funded us. But every great product has haters. Even OpenAI has haters.
What's even worse than having haters is nobody caring. Nobody cares whether you succeed or fail. And that really sucks. That was true for podcastcomments.com that I built. Because I built this thing and I really wanted to have a comment section on each podcast episode that I listened to, so I could read the comments, but nobody cares about the thing that I care about. And that really sucks.
So a big part of it in building startups, you keep going even when things aren't working out as you expected. All the ideas that didn't work out before Julius, was that people didn't care. And so today, when people care, it feels so good.
From a morale standpoint, it feels great. While we are actually building something that people want to use. And they care about it enough that they will write about it or share their experience on social media about it, and it feels great. So I love elevating those people.
Turner Novak:
Yeah, it's good to just bask in it, because you probably have days where they just suck. Things do not go as planned. Your ChatGPT plugin gets banned or whatever, and you're like, "Man, are we done? Is it all over?"
Rahul Sonwalker:
Yeah, all the graphs go down after that.
Turner Novak:
In terms of just the future of AI. It's probably like the stereotypical tech bro podcast topic right now, but you're actually building something and you're in it.
So how do you think about that? The next 10 years? If I'm someone who maybe I don't really understand AI or I'm afraid of it, or I don't fully grasp it. Just help me think through over the next 10 years.
Rahul Sonwalker:
Totally. So our vision with Julius is that very soon, AIs will write more code than human beings do. In fact, Jensen talked about this in his GTC keynote.
He predicted that AIs will be writing all of our code. Not just some of it, or more of it, but all of our code will be written by Ais. And that when that future happens, it's going to allow people to do really powerful and useful things that they previously couldn't.
Because if you think about it, code is so powerful. You can make not just websites with it, you can do really deep data science with it. You can make apps with it. You can make software with it. You can solve problems. You can even use code to control a computer or a hardware device.
All of these things are possible with code. It's the most powerful tool built by humans yet. And AIs will soon be able to work with that tool in really useful ways that previously you had to be an expert in. You had to be an expert software engineer to make an app or make a website or make a really common piece of software.
So when that future happens, people are going to be doing really useful and meaningful things. And we are building that feature with Julius. We think today these models are good at writing code for data science and data analysis. But as the models get better, we will allow people to do really more powerful things like make a website, make an app, make software. In fact, the real powerful use cases of Julius aren't even being conceived of right now. Because when you look at a model four years ago, people thought, "I guess I can use this to write copy for my website."
Turner Novak:
That was kind of the official first big use case.
Rahul Sonwalker:
Right. Or Piñata Farms, two years ago? Three years ago? They were taking Will Smith and making him do cool things or sing a song or lip sync with a song. But now they're able to do way more cool things.
So as the models improve, more and more useful things will become possible. And just four years ago, people were conceiving of copywriting as a main use case or SEO writing. And today it's like, "Oh, I can get an AI to do regression for me, or I can get an AI to plot millions of rows of data for me." And soon there will be more and more use cases.
And Julius, our hope is that we build the interface and the tool to allow these models to do really useful things with code.
Turner Novak:
So then this is a question from our mutual friend, Roon, who you kind of mentioned this earlier, but he said, "What happens when you give an AI a computer?"
Rahul Sonwalker:
Totally. So that's pretty much what's happening with Julius today.
So on the surface, people see this AI that works as their data scientist. And it can give them data insights, create data visualizations for them. But really, what we are building under the hood, behind that interface, is we're taking the best models in the world and giving them access to a computer where they can actually write code and do all the things you ask it to do.
So if you give it a spreadsheet and you say, "Okay, can you find me my 10 highest paying customers from four years ago," it will write code to do that. It will spin up a code environment, it will write that code, look at the file, deconstruct it, look at the output of the code, and get you that answer.
So that's how basically Julius works today. We're giving these AIs their own computer. Soon, that computer will have a browser. Today, it just has the ability to write code.
Soon, it will have a browser where it can look up documentation on the internet. So when you say, "Hey, I have this hardware out there in the field, and you can connect to this with this API," the AI will make a network request from the computer that we give it and manipulate the hardware, or you tell it, "Hey, can you go to Shopify and get me all the reviews of my competitor and put that in a spreadsheet and give me the main insights out of the review?" It will spin up a browser, scrape Shopify, get you those reviews, and then give you that insights, all with the computer and the code.
Turner Novak:
So for some of these examples, your interface previously used to be a spreadsheet. Or the interface was literally the whatever the interface of that hardware device was. The future of AI, artificial intelligence, it creates new interfaces where the computer can talk to other pieces of the computer. And then the interface changes from, just simply being graphical, to being text-based. And then text can also change into images, or make a video, or make a different type of spreadsheet or something? Is that a fair way ... it's like this whole new interface that has a bunch of different things you can do with it?
Rahul Sonwalker:
Sort of, yeah. You basically have an AI or a bunch of AI models that are using a computer to do the things that you ask it to do.
So Guillermo recently tweeted about Julius because he was trying to upload a video to Twitter for the most recent launch they did with Vercel. And that video needed to be a certain dimension, and he couldn't find the tool online to resize the video.
He could have downloaded Adobe Premiere Pro or one of those hardcore video editing softwares, or he could just uploaded it to Julius instead and said, "Hey, can you make this video this aspect ratio?" And Julius just did it. It imported FFmpeg, which is a Python module to work with videos.
It wrote a code script and resized the video for Guillermo. And he tweeted about it and got thousands and thousands of likes, a lot of views. And that was the power of giving these models a computer to do these things.
Turner Novak:
Yeah, because if you think about it, when you're in Adobe Premiere, whatever product, and you're changing the dimensions, you need to access a certain interface that allows you to then change what's in the code. It's just zeros and ones. Ultimately, everything that we see that's electronic, it's all just zeros and ones.
Rahul Sonwalker:
Yeah, totally.
Turner Novak:
So it unlocks new interfaces, shortcuts, ways for computers to talk to each other. Data to talk to each other. Ways for information to flow. New interfaces. All that kind of stuff?
Rahul Sonwalker:
Yeah.
Turner Novak:
So then what do you think happens? I mean, there's so many AI companies right now, it would probably take us a full podcast episode to just list every AI company. So what do you think happens to all of them just over the next couple years? I believe this was a question from Aashay Sanghvi, at Haystack. One of his questions was, "What happens long term with all these AI companies?"
Rahul Sonwalker:
I think the future is kind of hard to predict. I have my own negative sort of takes on it. But as long as you're building something that is useful to people and people resonate with and really like using, I think these companies are going to end up in a really good spot. For companies that are not focusing on that, and they are simply trying to do something with AI because it's hot, are possibly going to have some kind of trouble.
So an example is Notion, right? Notion is for many people almost indistinguishable from Google Doc, like, "It's literally a Google Doc. Why do I need Notion?" But then there's this whole other side of people or a community of users that live and die by Notion. Everything to them is a Notion Doc or a Notion template.
So if you can find those for you, people who love what you're building and people who can do useful things with what you're building, I think these companies are going to end up fine.
Turner Novak:
So will OpenAI kill every AI start-up? Because that's kind of the meme that some people think.
Rahul Sonwalker:
I've actually thought about this quite a lot. And I love reading early tech history. And if there's one lesson you can draw from early tech history, it's that the future is really hard to predict. You just have to ride the tiger and see where you end up.
In terms of OpenAI, OpenAI is really friendly with startups. We work with OpenAI. We work with Anthropic, and other model providers. OpenAI is really friendly. A lot of my friends work there. And what they're building is really useful.
If you can find useful things to do with what you're building, I think that's worth a lot more focus and attention than worrying about whether OpenAI will kill your startup. Because there's a lot of additional work that needs to be done to take these models and to make them useful in a friendly way, to make their outputs intuitive to users, and to leverage these models to their full potential.
I'll give you an example. So today, over 200,000 images get uploaded to Julius every month. And when an image gets uploaded, because people take photos of their computer to analyze some data, people want to take photos of math and solve math using Julius.
But when they upload an image, we actually pass it through multiple models, from multiple model providers, to extract out the most information as we can. Oftentimes, we do that in a bunch of steps. So we use ... for example, Cloud's Haiku is a really fast model that, within milliseconds, will give you a very brief abstract of what's in the image. And then you can take that output and put it in a different model, like GPT-IV Vision, and get more richer look at what's in the image.
So, if it's like a graph, you can first filter out that, okay, there's a graph in this image. And then you can customize the question you're asking GPT-IV Turbo about the image.
So there's a lot of these tricks that go into making a useful product. And anyone who has curiosity to play with these things and anyone who has the passion to talk to users and solve their problems, will end up building something that is used by a lot of people.
In terms of tech history, I think a lot about OpenAI. Their story is so interesting. They started in 2015 when Google had the world's best researchers. They had billions and billions of dollars.
Turner Novak:
Hundreds. $100 billion, probably.
Rahul Sonwalker:
Hundreds of billions of dollars. They had the world's data. There's a billion queries plus happening a day on Google. There's YouTube, video data, audio data, text data. They have the most data in the world, and they have the best researchers. They have the most money.
Not just that, they have their customized hardware. They have custom hardware to train these AI models, now TPUs. They didn't have that in 2015, but the project was already underway.
Turner Novak:
They should have won. On paper, they should have. Every AI company that was started, every transformational tech that's come out in AI in the last decade, Google basically invented all of it, right?
Rahul Sonwalker:
Yeah, exactly. Google invented all of it. They should have won on paper.
Yet, today, we have OpenAI and ChatGPT just crushing it. Even though Google has distribution, they had distribution through search, and YouTube, and Android, and Maps, and Google Images. They just have everything, and yet they're not able to catch up or keep up.
With technology, things move really fast and there's ways in which startups can build things, talk to users, keep their ears close to the ground, and innovate and build really cool things.
Another example of this is Nvidia. I think that might answer Aashay’s question of what happens to these AI startups? There's so many, you might take a whole podcast to list them all.
But when Nvidia, 1993, this was the year when, I believe, Doom and Quake came out. Everyone had the insight that consumer gaming is going to be really big, because for the first time, you can play 3D games on a PC. 3D is just amazing. You have a whole world on your computer that you can explore and build Minecraft or whatever. 3D is just so powerful.
Not just Jensen and the two other co-founders of Nvidia had the same idea, so did 89 other companies who started in 1993 doing the exact same thing. Which is to build a graphics card that would help you play video games on your PC.
Turner Novak:
So why’d they win?
Rahul Sonwalker:
A big part of why they won was persistence, and not dying, and always thinking on your feet.
They were prescient enough in 1993 to realize that, "Wait, we have 89 other companies doing the exact same thing. We need to get the next generation of chip out first before everyone else does, because within months, everyone else is going to catch up anyway."
That was their strategy for many years, where they just kept getting the next generation of graphics card out first before the competition could. Microsoft tried to kill them with Direct3D. Intel tried to kill them by integrating graphics card into motherboards. They said, "How about we just give you a motherboard that already has a CPU and a graphics card?"
And then Nvidia realizes that, "Wait, if we call out a thing a graphics card, people are just going to buy this thing from Intel." They completely rebranded GPUs, which is like a whole new thing.
Turner Novak:
Instead of a graphics card, it's a processing unit.
Rahul Sonwalker:
Exactly. There's a lot of early history about this, about how GPUs, for the first few months, were a marketing term. They weren't truly programmable, but they realized that, "Hey, if you call a thing a GPU, it's like a whole new thing. It's not the graphics card that you already get with your Intel motherboard. You need a GPU to play video games."
And then about a decade in is when scientific computing became really obvious. But it wasn't until 2004, 2005, 10-12 years in, when most of the startups out of those 89 companies died. They just didn't stick around long enough.
So persistence is a big part of it. Understanding where your industry is moving, what the competitive advantage, what do the users want. The users want the fastest chip at first, even if it is really buggy. Turns out, that's what gamers wanted in the 90s. It wasn't the cheapest. It wasn't the most reliable. It was that they just wanted the next chip the fastest. So you have to understand your space.
Turner Novak:
Yeah, that's the takeaway. It's just what is the thing that your customers actually want, and just do that. Sounds easy enough.
Rahul Sonwalker:
Yeah. I feel like startups are really simple algorithm. There aren't many 4D chess ideas. The algorithm is really simple. It's just hard to execute and continue to execute the algorithm.
Turner Novak:
So then how do you think about just, I guess we'll call it building a moat, if that's fair to say, and where do you get some of that inspiration from? Just throughout history or other startups?
Rahul Sonwalker:
I think startups, by default, don't have a moat. When you first start, you have to carve your moat out of things. Each startup has its own unique mode. All startups are outliers.
You can't take NVIDIA's playbook and execute that playbook exactly how they did it and end up with the same outcome, because Nvidia is a very unique company.
Startups are all about power lies. It's the outliers that win.
With Airbnb, for example, Airbnb and couchsurfing.com are just pretty much the same idea. Couchsurfing was venture-funded, Airbnb was venture-funded. Couchsurfing had more distribution, but Airbnb just focused on different things. What if the houses that are on Airbnb, they just had really nice photos? What if we took the houses that are on Airbnb and cross-posted those listings? They automated that apparently in the early days, and that skyrocketed the growth even though they had no distribution.
Turner Novak:
Do you have any things like that with Julius? These are the couple things that are just opinionated product differences.
Rahul Sonwalker:
We try to use a lot of different models for their strengths. We have a feature where, oftentimes, the user will ask the AI to generate our data visualization, and that data visualization is not what they actually wanted. Maybe the color is different than what they wanted. And either they can just chat and give the model of feedback.
Or, what we actually do is we take the image and we take the code that generated that image, pass it through a different model, get that model to give us parameters and toggles, to change that graph.
So that attention to detail of what the users are trying to do and how do we let them do that in a very easy way by using the best models out there is going to be, I think, really useful at least in the next few years. The hope is that if we continue building what the users want, we will have a good business.
Turner Novak:
Do you have other AI companies that you feel like are doing a really good job of this or any that you really look up to?
Rahul Sonwalker:
There's Pika Labs that is doing text to video. With AI, they're building something that is really cool and useful to a lot of people.
Another company that comes to my mind is Midjourney. They're a really unique company in the sense that it's just a Discord channel or a Discord server. They have no generally available UI even until today.
When they first launched, they used publicly available models and used that data and feedback to train better models.
They focused on one thing and one thing only, which is, "What if we made it really easy to make art using AI?" That sounds really simple, but in practice, it's really hard to do.
But I have my friend, Kyle. His mom, she's 50 years old. She's never been on Discord. But she's on the MidJourney Discord server trying to generate art that she can use in her recipe blog, because she just wants to make this blog. When she's talking about pasta, she wants to show Italy and somebody who eat pasta in Italy with The Colosseum in the background or the Leaning Tower of Pisa in the background. You can just take your imagination and turn it into art. Turns out, there's massive demand for that. I think those are two really cool companies.
Turner Novak:
I always think it's crazy that they've got... I don't know what the latest public number is. I think the highest public number I saw was $400 million in revenue or something? But whatever the number is, an astronomical amount, and it's literally just people typing in texts to make art. It's all in Discord too. That's the most insane part to me. Who knew you could make a mid to high nine-figure revenue business on Discord? Insane.
Rahul Sonwalker:
It's crazy, because no market map could have predicted it.
Turner Novak:
Yeah, exactly.
Rahul Sonwalker:
No Substack that could have predicted it.
Turner Novak:
Yeah. Well, the thing, I think, is amazing about Discord is they were the center of crypto when Web3 was a thing. Everyone was on Discord. Now, with AI, it's like all these ad companies have Discords and some of them are built on Discord. Their business model is processing the whole business and processing the images and everything on Discord. It's crazy.
Rahul Sonwalker:
Yeah, 100%. With Julius, we are focusing a ton on what do the users want. As long as you build what the users want and what these models are capable of, I think we'll have a great product that will serve millions of users.
I'll use a couple examples. Today, when you generate a data visualization on Julius, we use a different model than the model we use for code generation to change the graphs and give you toggles to play around with different parameters on the graph. It's hard to do that with natural language.
We, on the fly, generate UI components that we can surface to the user that they can play with and then change the graph with.
Not just that, there's another example is computers. We're giving these models really powerful computers and code environments where they can write code and do useful things.
Building and maintaining these computers is hard because these models have a certain amount of knowledge in their training data. They remember certain modules and code packages up to a certain date. And making sure that your environments are compatible with that. And what the model expects is what you actually have in your computer for the model, is hard. We're hyperfocused on building that infrastructure of computers, maintaining those, and making those very easy to use for the AI.
Turner Novak:
It's basically that the whole product is almost AI-native, if that makes sense.
Rahul Sonwalker:
It's completely AI-native.
Turner Novak:
Yeah. I mean, it sounds like a pie-in-the-sky description, but I guess it's kind of true.
So the UI that you present to the user actually changes based on what they want to do with it?
Rahul Sonwalker:
Exactly. We generate UI components. Well, we let the models decide what UI components to surface to you based on what you're trying to do.
When you open a tool like Excel, you see all the controls at once, and that gets super intimidating. It's really hard to use.
In Julius, the main component is your chat box with the AI, and then the AI can populate more and more UI components on the fly do to do what you want to do.
Turner Novak:
Fascinating. I have two other questions. I feel like this is a question from someone else, because I don't know why I would ask this, but it's on my list. I can't remember who gave me this one.
You lift, why? You could argue it's a distraction from building.
Rahul Sonwalker:
Who asked that? That's so funny.
Turner Novak:
Yeah. I didn't make a note of who told me that one, but it's interesting one. Oh, you know what it is? I think I added it because we're friends on BeReal. You just always post your BeReals from the gym and you'd be working out. Obviously, it's a big thing.
Rahul Sonwalker:
Yeah, I love BeReal.
Lifting is fun because it just helps clear your mind. It helps you get physically tired at the end of the day, so that when you have a lot going on in your head, all these possible paths that you're exploring about where Julius could go, what are the things we could do. It helps you shut that off and just fall asleep for six hours, or five hours, or whatever amount of sleep you're trying to get at the night.
On a more philosophical level, it also shows you how progressive overload, It applies to so many things in life. When you first go to the gym, you try to lift less than a plate and you really struggle, but then you do that over time and you get to a point where you start warming up with the weight that you used to max out at a few months ago, a few years ago. It's very philosophical too, in a way.
Turner Novak:
Yeah. I felt like I got a lot of benefits. When I really started working out consistently in college, I just noticed I was just better at everything else. Your blood moves faster, your reactions are better, you're more confident. You're just better at everything that you do, basically. I mean, that's how I would describe it to people.
It's healthy. Do you want to be out of shape or in shape? You'll probably live longer or whatever. It's healthy. It's good.
Rahul Sonwalker:
Totally.
Turner Novak:
I mean, could you argue that it distracts you from making progress on your startup or it's a worthy trade-off?
Rahul Sonwalker:
I think BeReal is a bigger distraction than lifting.
Turner Novak:
The one-second on that you post. Yeah.
This is another question from Aashay, assuming there's a story behind this, but what do you love about SF and what do you hate about SF?
Rahul Sonwalker:
I love SF. I don't think where I am today, it would be possible if I wasn't in the US. Or I wasn't in SF. So I'm grateful for both of those things.
With SF, I think there's this community of people who are very high agency, and open-minded. And always wanting to try new ideas, and trying to be at the edge of cool things. You see a Waymo driving around, and then you go to a party on a Friday and there's a Waymo engineer who helps build Waymo in the party.
Turner Novak:
Waymo is Google's self-driving car, just for people completely out of the loop. They work, by the way. They're real. They're really cool to ride in.
Rahul Sonwalker:
They're really cool. Yeah. They're really cool to ride in. They work. If you go on a 30-minute walk around San Francisco as of April 11, 2024, you'll see at least five Waymos in the 30-minute walk. I love that about San Francisco.
It’s hard to find a lot of high agency people like that anywhere else. But the downside of it is there's also low variance of people. You go to a coffee shop, people are talking about startups or AI. Or you go to the gym, everywhere around you is a tech worker. Sometimes I miss a high variance of people where someone's training for a marathon or someone's writing a book. So I think there's pros and cons to SF.
Turner Novak:
It's the beauty that I get living in Ann Arbor, Michigan.
Rahul Sonwalker:
Hell yeah.
Turner Novak:
Rahul, last question. I know you take a lot of inspiration from history, other companies, other founders. Do you have a favorite one, favorite story?
Rahul Sonwalker:
I think it's really important to read about the non-PR stories of startups to know actually what they went through and what happened. Not the eBay story of, "Hey, I wanted to sell best dispensers, and I started eBay." Turns out, Pierre just built this site auction web, and just posted a bunch of links on online forums until they got users.
But my favorite one is Microsoft, because of how dominant they were at some point in time in history. And how they basically invented and controlled the software industry out of no existing moat. I have a bunch of Microsoft posters in my room. These are actual ad posters from when they first started Microsoft. This one is a Microsoft BASIC. For the first seven years, Microsoft was in the business of programming languages.
Turner Novak:
They invented, they created programming languages?
Rahul Sonwalker:
Not really. They took programming languages that already existed. These computer makers like Altair, and Macintosh, and IBM, all these computer makers wanted these programming languages to work on their hardware depending on what chip they have or what processor they have. Because if these languages work, then developers can build applications that actually make this hardware useful.
If you're a computer maker, you make the computer hardware, but if no one actually makes something that works on it, then no one will buy the computer outside of a hobbyist. So Microsoft would sell licenses of Basic, Fortran to Altair, Commodore, IBM, Mac, Apple, and that was their bread and butter for the first five or six years.
This is before operating system is what they really became known for, but that didn't exist. They were in the business of programming languages.
In fact, CPM is an operating system that we've never heard of today, but that used to be the dominant operating system in the '70s. Everyone had CPM until Microsoft bought DOS from a different company and then made it work on the IBM PC, and the rest is history.
What's interesting is that they went through so many iterations to get turned into a trillion-dollar company, become really dominant in a market that before the '70s didn't exist. There was no software market. The TAM for software was zero before the '70s.
Turner Novak:
Yeah. All those VCs would've passed. The TAM wasn't big enough.
Rahul Sonwalker:
This is true. Bill Gates, he built Microsoft Bootstrap. He took no outside capital until, I believe, the year before they went public.
Turner Novak:
It was basically he just wanted a board member and just someone who could be an adult in the room, I think is maybe a way to describe it. Just like, "We're going public. I probably need some help with this."
Rahul Sonwalker:
That's funny. Yeah. He is just, "I want to go public. I'm ready. I just need a adult in the room." It's crazy.
I talked a lot about Excel just now, but there's this thing Microsoft built before Excel. It was called Multiplan, which I have a poster for. This is the Multiplan ad poster from the '80s. But this was Microsoft's spreadsheet software before they built Excel. They were trying to compete with Lotus 1-2-3 and VisiCalc.
Multiplan went nowhere. Today, Excel is basically on every single PC out there. Any kind of desk job you do, you probably work with Excel.
Turner Novak:
Yeah. Excel runs the economy, really.
Rahul Sonwalker:
Right, right. Exactly. It took them 1987 to get the first Excel out. 12 years into the company.
That, to me, is a really interesting company because they built a really successful product. A series of really successful products, that didn't immediately work out of the box, but ended up being used by billions of people. And they invented a market out of it.
Turner Novak:
Yeah. It's a lot of good inspiration for founders. Never give up. Stick with it. It might take you 12 years to figure something out. Now, it's a trillion-dollar company.
Well, this was awesome. Thanks for coming on. This was a really fun conversation.
Rahul Sonwalker:
Thanks, Turner. This was a great chatting.
Stream the full episode on Apple, Spotify, or YouTube.
Find transcripts of all other episodes here.
Giving an AI a computer with Rahul? This sounds like a recipe for either brilliant innovation or hilariously unexpected outcomes! I just hope the AI doesn't get too adventurous trying to 'mix' its own programming. If only finding the perfect blend, whether in AI configurations or just a tricky color, were as simple as using a [Color Mixer](https://duckcolor.com/en/color-mixer)!