How Snapchat is Evolving in the Age of AI with Saral Jain, SVP, Head of Engineering

(Intro Narrator at 00:00:00) Today, we're talking to Saral Jain, SVP and Head of Engineering at Snapchat, about how they're evolving as a company. You're listening to Joel Beasley, Modern CTO.

(Joel Beasley at 00:00:16) I want to know what is going on behind the scenes of Snapchat.

(Saral Jain at 00:00:21) Yeah, absolutely. I mean, look, Snapchat has been around for almost fifteen years at this point. And through all of this, our core mission has stayed the same. We really want people talking on Snapchat to feel like how you talk in real world. Ephemeral, casual, having fun with your close friends and family, essentially. And that core mission has not changed. It's still pretty much the same. What has changed is the scale. We now reach hundreds of millions of people who use Snapchat every single day. Almost a billion people use it every month. Trillions of Snaps get created every year. So I think that scale is really, really interesting because it helps us build on top of this core messaging platform to build features like our map and stories and spotlight and our very vibrant AR ecosystem. And so it's really fun to kind of deliver all of this. And especially with AI now, we are able to turbocharge all of these features on top of a very massive community. So it's a lot of fun, but that core mission of connecting you to your core friends and family is still the same at Snapchat.

(Joel Beasley at 00:01:19) That is awesome. So everything you guys are doing ties back to that core mission.

(Saral Jain at 00:01:23) Absolutely.

(Joel Beasley at 00:01:25) Do you use Snapchat a lot?

(Saral Jain at 00:01:27) A lot. I mean, at this point, I use it for everything. But yeah, I mean, I think one of the most favorite things I do is recruit people to come join me on Snapchat. But it's a lot of fun. It's a lot of fun.

(Joel Beasley at 00:01:38) What's your favorite feature on Snap?

(Saral Jain at 00:01:40) I mean, it's hard to beat those AR lenses. Most people call them filters, but these really, really fun ways to kind of transform yourself, but also the world around you using AR. And now increasingly more generative AI based lenses. It's a lot of fun for people to do. And it's actually a lot of fun to play around with my four-year-old as well. I have a four-year-old kid. Trying to kind of do silly face lenses with him is a lot of fun as well.

(Joel Beasley at 00:02:05) Yeah, they're way better than the default ones on Apple.

(Saral Jain at 00:02:08) Yeah, I don't use iOS.

(Joel Beasley at 00:02:10) On FaceTime, you can change myself into a monkey when I'm talking to the kids and stuff. But the Snap is, you've got the most creative different lenses. And I didn't know about generative AI for the AR lenses. So is that out yet publicly?

(Saral Jain at 00:02:25) It is out. I think we launched it in Q4 of last year, and it's one of the most popular lenses on Snapchat right now. And the thing is, I mean, you should definitely try it out. It's an open prompt lens, so you can basically give it any textual prompt on how you want to transform yourself or the world around you, and it'll do it. So if I tell it, "I would like Joel to have an awesome beard, and he should have my crappy glasses," it'll kind of make that change on the fly right now, and so it'll be really fun for you to try that.

(Joel Beasley at 00:02:54) That would be fun. That would be cool. I'll give that a shot after we record today because I think it's summer right now. I know yours is young, so I don't know if they're in school yet. But right now, ours just got out of school for summer. We got eight, seven, and three, so kind of right in there. And they always want to play. They want to play with something, and they want to play with some technology or one of Dad's gadgets. And so I'll play with those filters this afternoon if it would be fun.

(Saral Jain at 00:03:20) Yeah, those are fun ages, by the way.

(Joel Beasley at 00:03:22) Yeah, I know. I know. Yeah. One of the things that was interesting to me about Snap in relation, from a branding perspective, you've got, you know, Facebook, X, Snap. You guys are up there. But from a size perspective, you guys are like David and Goliath. You're a much smaller size competing with these much larger companies. How do you think about that?

(Saral Jain at 00:03:50) Yeah, it's something I get asked a lot, and I think it's been pretty clear that ever since Snap was born, we have always competed with companies that maybe have 100x bigger size, and it depends on how you define size. It could be budget. It could be headcount, whatever, essentially. But, I mean, honestly, Snapchat was born as the anti-social media company. And so we have actually never seen ourselves as a social media company. Sometimes the market and the industry might categorize us as one, but we always have seen us as a place where, as I said in the beginning, it's a place where you should have fun with your close friends and family. And we see ourselves as a visual messaging app that helps enable that. So it's all about that creativity and those conversations with your close friends and family versus this public theater, essentially, of, you know, gathering followers and stuff. But at the same time, what we have expanded is the number of features we can enable on our platform so that people can kind of have all sorts of experiences including turbocharging this close friend connection, essentially. But at the heart, we are still a visual messaging app, essentially, and that's how we categorize ourselves.

(Joel Beasley at 00:04:58) Okay. So instead of consumption, it's creation.

(Saral Jain at 00:05:02) Yeah, exactly. And that shows in every single decision we make in our product. For example, I am not aware of any other platform that reaches a billion people that opens to the camera. Right? Because when we want people to think of Snapchat, it is first about the creation part. It opens to the camera. That is where we reduce the friction, essentially. People take a Snap. They send it to their close friends and family versus on almost every other app in a social media sense. Essentially, it's a consumption app. You go to a feed of some sort. Right? You start consuming content, essentially. And versus in Snapchat, you start creating content, and then you obviously interact with your friends and family and chat with them or message them or Snap them or watch a bunch of content. But it's all about that creation, which is the heart of Snapchat, essentially.

(Joel Beasley at 00:05:47) I love it.

(Saral Jain at 00:05:48) It's better—

(Joel Beasley at 00:05:48) What was it? TikTok? I opened it, and it opens to an ad now. I don't use it much, but I go in there once a week to check the stats of the video. I have a video poster that posts there. And I noticed last month when I went in there for my one time a week that it started opening to an ad, and I was like, I didn't think I could want to open this less. But they found a way.

(Saral Jain at 00:06:15) I think so. I mean, honestly, I don't use TikTok a lot myself either. But, yeah, I mean, certainly they have done some things right, but it's one of those products where the whole point is for you to kind of spend a bunch of time consuming a bunch of content versus Snapchat is just different. Not only has Snap survived, but thrived in that competitive landscape essentially. And the only reason we have been able to do that is it's a very, very creative company and we move extremely fast. We understand our community very well. We love our community. We can kind of deliver features for our community much faster than anybody else does. And I think that core creativity has been that spark that has kept Snap thriving, and that will continue to be the case, especially with AI at this particular point in time. I think AI is probably the best thing that could have happened to a company of our size, essentially, honestly, because we can move much faster. Small teams generally can move much faster. And for us, the bottleneck had always been the ability to build, not the creative ideas. And now we can kind of, you know, turbocharge all of our creative ideas essentially with AI. So I actually think it's a perfect time to be at a place like that.

(Joel Beasley at 00:07:19) How has that impacted your engineering as far as, you know, have you built an AI agent that helps everybody? Are you plugging into existing stuff? How did you approach integrating this into your engineering workflows?

(Saral Jain at 00:07:35) I mean, AI transformation for engineering is really one of the biggest parts of my job right now, and especially the last couple of years, essentially. And it's not about an agent. We are basically changing every part of an engineer's workflow at Snap using AI at this particular point in time. The way we are calling it is the Golden Path of Engineering at Snap, essentially. And so, roughly, it kind of boils down to 14 agents that are managed blessed agents that all engineers use for their day-to-day workflows, essentially. It starts from how you write code to review code to investigate issues, to investigate crashes and bug reports, how you do data science, how do you do A/B testing. Everything has a managed agent, essentially. And so it's a pretty massive investment for us, but something that has really helped us move fast already.

(Joel Beasley at 00:08:22) Okay. I don't understand all of that. So, I mean, I'm a software engineer of 20 years, but when you were talking about managed agents, I've never heard this term before.

(Saral Jain at 00:08:30) Yeah. So let me give a very concrete example. We are building something called Casper right now. Casper is an autonomous remote coding agent. What it does is, traditionally, when you do software engineering, you have a workstation, you check out a repository, you write a bunch of code, you build that code, you deploy that code. Right? And then it goes through some verification. With Casper, what you do is it's a bunch of agents running in the cloud. You're just talking to your teammate on Slack, designing a new feature, brainstorming what that feature is just on Slack, essentially, and Casper is listening into that conversation right now. And at one point you say, "Casper, do you have all the requirements? Why don't you go ahead and build it?" And Casper has all of the context of the history of Snapchat in our codebase, all of the codebase, not a specific repository, but literally all of the tens or hundreds of millions of lines of code. It understands all of that. It understands Snapchat documents. It understands the conversation you just had. It understands you and your persona and the person you were talking to and their persona. And it actually, you know, checks out code and builds all of that, does a first pass of verification. We have a different agent called CodePal that then takes the code it had built, does a code review of it. They kind of talk to each other a little bit, make sure that the code looks good, and then it comes back to you for verification, saying, "Okay, does everything look good?" It creates a PR, a code review essentially. And the human, ultimately, is the accountable person. They take a look at the code, they bless it essentially, and then Casper is able to check in the code. So that's the nature of engineering, and that's how things are evolving. Obviously, Casper is one of the newer agents we have. We are still kind of working through some of the build-out of it, but that's the future of engineering at Snap, essentially, and I'm so excited about it.

(Joel Beasley at 00:10:12) I am so glad that you said this. So we released, I think, an episode today with this guy named Trevor from Soundstripe, and he had 30, 40 comments of people being like, "That's not possible. That's not true," because he's a small company. Soundstripe, they're like 30, 50 people or something like that. But he lives in my town, and he was telling me, he's like, "Oh, for the past quarter, humans don't write code anymore. We've built these AI agents that we, basically what you just said. And then they check everything, and then they deploy." And everyone was in disbelief on social media. They're like, "That's just not possible." And I'm like, as I'm talking to these great leaders like you and other people, I'm like, this is not only possible, but this is in practice today at a scale that's more than just one or two engineers.

(Saral Jain at 00:11:05) Yeah, absolutely. I mean, look, software engineering is changing in front of our eyes in the last few months, essentially. And we are still very early days. I mean, at this particular point in time, we are building agents like this that can do a lot of the autonomous coding. But we also have to understand that places like Snapchat have a decade of code already, a lot of legacy systems, and we can't transform everything overnight. So what I'm talking about are what we are doing at this point in time for net new repositories, net new codebases, even moving really fast with some of the legacy codebase, but not everything has changed. Right? And so I think that's probably where some of that practicality comes in. But it is very clear to us who are kind of closer to the details right now that this is the future of software engineering. That humans will be doing more of the orchestrating, managing the agents, but agents would be the ones building a lot of the code, essentially, and even doing the verification passes and stuff.

(Joel Beasley at 00:11:59) Okay. Tell me about the day Casper was created.

(Saral Jain at 00:12:02) I think I'll tell you about the inception. So I think this was one of those ideas that one of the engineers on the team had and said that there's a lot of the building blocks that already exist. Why don't we kind of get everything together into this managed end-to-end solution, essentially? And the team really went on a sprint, and within three weeks, the entire thing was created, essentially, in terms of a private beta, essentially, which is a closed launch within Snapchat. And the team worked very hard. The way Casper was created was Casper was at one point building a lot of itself. Essentially, you could tell Casper, "This is the functionality I want you to have in Casper," and it'll just build it. And so it's really a lot of fun to kind of see that whole evolution of Casper. But then we slowly started telling people about Casper internally at Snap engineers. And the demand was kind of insane. We have still not fully rolled out to all engineers at Snap, and the waitlist is hundreds of engineers already. And it's already generating so much code, but the amount of code generated is kind of a vanity metric. What's more interesting is generating really good quality code. It's like an AI teammate that has infinite memory, basically, and you can help it understand stuff and help it do stuff, but it's generating pretty good quality code. So we are very bullish about it.

(Joel Beasley at 00:13:13) Explain to me the waitlist. Why not press the button and let everyone have it tonight?

(Saral Jain at 00:13:18) Because these new solutions, new technology comes with a lot of friction sometimes, because not all features are fully built. Maybe it does not have access to this one particular dataset that a lot of engineers in a particular team might need to do their job well. And so Casper under the covers has a few building blocks. It has that ability to understand all of our codebases. It has an ability through our MCP gateway to understand all of the information at Snap.

Saral Jain at 00:13:45: But we are slowly adding that level of information and all of the code bases on it. So rolling it out to everybody would basically be very chaotic because it does not yet have all of the capabilities that everybody would need. So we are slowly rolling it out. There's also an aspect of managing all of the governance and the security and the auditing layer that we are slowly building over time so that we know how people are using Casper, right? Like, a powerful technology like this, we want to put all of the safeguards in place that it is not doing a bunch of things that we do not want it to do, especially creating AI slop in our code or, you know, exposing us to security that we are not comfortable with. And that's why the gradual rollout.

Joel Beasley at 00:14:25: That makes sense. You just gotta, kinda like a child, you kinda help them walk, and then you get more confidence, and then they can walk on their own. Right. Interesting. So structurally, as an engineering leader, do you look at Casper like a product and it has its own team and things like that?

Saral Jain at 00:14:43: Yeah, I think so. I mean, I've always looked at internal platforms and internal tools as products, essentially, because ultimately they are used by internal customers. And so it's not just about the AI tools. Like, our cloud infrastructure, for example, has a product mindset as well in terms of the team. It has a dedicated team. Although with something like Casper, there are so many people interested in building it. Like, I do a bunch of code reviews for Casper as well just because it's so much fun. And so it's less of a team, more of a community build-out, but there are dedicated people working on it.

Joel Beasley at 00:15:12: That's interesting. So, yeah, there was this five levels of AI coding article that we read, and I'll spare you each level. But the beginning, level one, was like autocomplete of a function name, and level five was you just turn the lights off and go home. And it's just this self-maintaining code base. It sounds like you guys are at like level three or four where you're getting there. But do you think that we're gonna get to level five?

Saral Jain at 00:15:46: I would need to understand those levels a little bit more. But like, based on what you mentioned, I think it's hard to put a largish organization like Snap on a particular level because different parts of the team are maybe at different levels, essentially. Like, for some of the legacy code, we are probably between two and three somewhere. Like, you know, people are certainly using AI everywhere, but there's still a lot of friction and guardrails in place. For net new apps we build or, you know, net new repositories or internal tools, we are probably between three and four. Like, you know, there's a lot of autonomous agent building happening. There is spec-driven development happening where people give the overall plan and the agent builds most of the code. That is happening for net new repositories. Level five seems dark to me. I don't know if it's gonna happen or not. It seems like a pretty real possibility given the pace of evolution. But I actually think, ultimately, it kind of is immaterial who writes the code. What is material is who owns the code, right? And that to me has never changed. Like, humans own the code. Essentially, we are accountable for the code because, ultimately, when things go wrong, we want to be able to feel confident that we understand what happened, right? And so to me, it's almost immaterial whether the AI wrote most of the code or the human wrote most of the code. The skill, the judgment, the traditional software engineering best practices that can help us understand all of the code is really still gonna be the key going forward as well.

Joel Beasley at 00:17:15: I 100% agree. And from an abstract standpoint, so I have no problem visualizing exactly what's occurring today, where we can interact with this and approve code, and we own the outcome. We own it socially within the company. We own the performance of it, and we use all these tools to make ourselves more efficient. The thing that I haven't yet created a mental model for or visualization of is that level five, which is actually referred to as the dark factory. Essentially, you shut the lights off and you go home and it self-maintains. Meaning, like, every role that a human would perform in that level four is now taken care of. And at that point, it seems to me, like, do all apps just disappear? Like, it's a weird thought to think about, but we should think about it a tiny bit. Like, how does that work?

Saral Jain at 00:18:14: Yeah. I understand. I mean, there's so many implications of something like that from a societal perspective that somebody has to really think it through. But, honestly, the way I think about it is a lot of powerful technology transformation that has happened in the history of human evolution have always caused change, but that has never led to an outcome where humans themselves have no role to play or are not important. Now the role will certainly be very different, right? Like, what we are doing as software engineers will certainly change. Every software engineer knows that today. So being able to reskill ourselves to still be able to provide the value at that particular point in time would be really important. But it, I would like to believe that the importance of judgment and taste and what to build will certainly still be there. And that is where the humans can play a big role even in that dark scenario you mentioned where, you know, the bulk of the work itself is happening through these orchestrated AI agents.

Joel Beasley at 00:19:14: I think you're exact—I think you nailed it with taste. I think that's one of the roles that humans are, because you have to, like a curator. We become more like curators of great things that were generated that were great, right? Because I don't want to spend my mental power filtering. I wanna find someone I trust in movies and, you know, or like in different categories in music. And then they're curating the best of, and I think that's a good visualization for the future.

Saral Jain at 00:19:50: Yes. I mean, I think, ultimately, there is one certainty in life. The number of hours in a day any human has will never change, right? Like, we all know that. And so if we know that is the one constraint factor, essentially, in terms of people who consume any output of anything that a software engineer creates, then hundreds of apps get created, AI apps get created daily. Most of them get zero traction because that taste is not there, the judgment is not there, what you're actually trying to build or solve for is not there. I think that's where humans add the most value—is building those creative, beautiful experiences that people actually want to spend time with, essentially. And that is not going anywhere.

Joel Beasley at 00:20:25: Do you think that we will—well, I'm beating up my own question as I ask it. I was curious about the atrophy of the muscle of engineering. Like, I used to joke when ORMs and stuff came out, you know, years ago, I was like, oh, my retirement plan is I could always write SQL code because no one's gonna know how to do that because they're all on ORMs. Turns out, it's not a great retirement plan. The AIs and everything has gotten so good. You don't even need to know how to really write SQL code in a lot of basic applications. And so do you think, like, as humanity, we might lose these very important skills, or are they more like archaic skills that we just don't need anymore?

Saral Jain at 00:21:16: I think Joel, the most important thing I'll tell you in this podcast is you need a different retirement plan at this point. Like, writing SQL code is not gonna serve you well too much. No. I mean, I think, look, it's natural. I think as new technology comes in, essentially, there are certain skills that do erode, essentially, over time, and you have to develop new skills. And so what I would say is writing things like boilerplate code or even debugging issues that used to be core part of what engineers do—well, over a period of time, you know, you'll start seeing erosion of skills. But there are certain other skills that become more important, like, you know, that product-driven development, writing specs, being able to communicate your plan very well with agents, being able to have that judgment and taste we talked about. Those are skills that are not really eroding. If anything, I think the bar is increasing in terms of how important those skills are. And so I think there's gonna be a little bit of a mix shift, essentially, in terms of the skills that we expected engineers to have versus the skills we will expect going forward.

Joel Beasley at 00:22:15: That's interesting. Let's talk about your leadership and, like, what are you learning right now as a leader?

Saral Jain at 00:22:22: I think the most important thing is being able to adapt. I think things are changing in front of our eyes, and every leader in every industry, but certainly for a company like ours, AI is not a side project, right? It is fundamental to how we build, how we operate, how we win going forward. And so it's really important for me as a leader to be on top of it, be hands-on, be using all of these tools, not just saying people should use all of these tools, but at the same time, helping the organization and the company overall, in a sense, work through the change management of what do we need to become as an AI-native company or as an AI-first company, essentially. That is the most important skill that I'm working through right now and working with my peers on it as well.

Joel Beasley at 00:23:03: Have you changed how you hire engineers? Like, your questions that you ask or—obviously, we were just talking about looking at these traits of taste and judgment and all that. But has that actually affected your hiring process in a tangible way?

Saral Jain at 00:23:18: Yes. Somewhat. I mean, we are definitely running a lot of pilots right now in terms of how to get more signals on these kind of things versus, you know, at this point, if you are going into an interview and asking somebody to write a reversal list or any other LeetCode kind of questions, essentially, it's not very useful. And so we are piloting interviews where we are giving the candidates access to their favorite AI toolkit, essentially. And we are giving more open-ended problems like build X, essentially, and see how they work through requirements gathering, planning, how they do the back and forth of how to build a good product, how to spec it out, how to work with agents as part of the interview itself. And then also when an agent generates code, how to have that adversarial conversation with the—we have also piloted something like give them AI-generated code and try to see if they can find issues with it. Those kind of things are happening, but they're still pilots because I do believe that some of the original aspects of what we look for in candidates and interviews still are very valid. Leadership skills, system design skills—those are certainly as valuable as they've ever been. And so it's a bit of new pilots, but also doubling down on what we used to do before.

Joel Beasley at 00:24:37: When you're bringing on a new member to your specific team as a direct report, what type of hiring questions do you ask to figure out their leadership skills?

Saral Jain at 00:24:46: There are a few things that are very important to me. First of all, is this a person that I would want to work with on a day-to-day basis? Like, what is the problem they are trying to solve? And it's really, really important for me that we have people who are saying what they do and doing what they say. That basic integrity is there. So that is the number one skill I look for in a leader on my team, essentially. But at the same time, strong ownership, bias for action, making sure that people always raise their hands. Because in any company, and certainly in a company like Snap as part of my leadership team, there's always gonna be more problems to solve than the number of people available to solve them. So the people who raise their hands actually do really well. And so that's the second skill I really look for. And then, obviously, competence and being able to roll up their sleeves and do the job themselves is really important as well. We want leaders who are very hands-on and who are able to build stuff themselves. So those are the few things that are really important to me.

Joel Beasley at 00:25:41: How do you—let's talk about the integrity one. I won't make it too ambiguous. So, like, if you're trying to figure this out, it's on your checklist, this integrity. What type of questions do you ask in the hiring process to figure it out?

Saral Jain at 00:25:54: I think the most important thing is just using their own past experience to ask an open-ended question about how they resolved conflicts, how they have described situations where they had to have difficult conversations with peers, with their employees and stuff, and see how they managed the situation. And peeling the layers of the onion as they give their responses will almost certainly tell you how they think. And so that first answer is almost immaterial. It's the follow-ups on that answer that can help you understand that thought process of how they navigated a complex situation. And if a person who's seen enough, who's reporting to me, is not able to come up with a complex enough situation, that to me is a signal itself. And so it's one of those things that, I think over a period of time, you get good at getting that signal.

Joel Beasley at 00:26:36: Yeah. I have figured out how to do it. I haven't necessarily figured out how to teach it. I can identify people who have it and empower them, but I haven't been able to figure out how to take someone who doesn't have it and teach it to them. Have you?

Saral Jain at 00:26:52: I think there are a lot of teachable skills. Not being an asshole is not one of them.

Joel Beasley at 00:27:01: Yeah. Accurate. That checks out. Saral, this has been fantastic. We made a podcast. How do you feel?

Saral Jain at 00:27:09: I feel great. It was a lot of fun conversation. And, no, I mean, I really enjoyed it.

Joel Beasley at 00:27:13: Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague that you think would get value from it. And if you have topics that you'd like to hear discussed on the podcast, either add me on LinkedIn or send me an email, [email protected]. Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.