See all Lex Fridman transcripts on Youtube

Speaker 1

00:00

The following is a conversation with Kevin Systrom, co-founder and longtime CEO of Instagram, including for 6 years after Facebook's acquisition of Instagram. This is the Lex Friedman Podcast. To support it, please check out our sponsors in the description. And now, here's my conversation with Kevin Systrom.

Speaker 2

00:22

At the risk of asking the Rolling Stones to play Satisfaction, let me ask you about the origin story of Instagram.

Speaker 3

00:28

Sure.

Speaker 2

00:29

So maybe some context, you, like we were talking about offline, grew up in Massachusetts, learned computer programming there, liked to play Doom II, worked at a vinyl record store, then you went to Stanford, turned down Mr. Mark Zuckerberg in Facebook, went to Florence to study photography. Those are just some random, beautiful, impossibly brief glimpses into a life.

Speaker 2

00:54

So let me ask again, can you take me through the origin story of

Speaker 4

00:57

Instagram, given that context?

Speaker 3

00:59

Yeah, you basically set it up. All right, So we have a fair amount of time, so I'll go into some detail. But basically what I'll say is Instagram started out of a company actually called Bourbon and it was spelled B-U-R-B-N.

Speaker 3

01:16

And a couple of things were happening at the time. So if we zoom back to 2010, not a lot of people remember what was happening in the dot-com world then, but check-in apps were all the rage. So- What's a check-in app? Gowalla, Foursquare, Hot Potato.

Speaker 3

01:34

So I'm at a place,

Speaker 2

01:35

I'm gonna tell the world that I'm at this place.

Speaker 3

01:37

That's right.

Speaker 2

01:38

What's the idea behind this kind of app, by the way?

Speaker 3

01:41

You know what, I'm gonna answer that, but through what Instagram became and why I believe Instagram replaced them. So the whole idea was to share with the world what you were doing specifically with your friends, right? But there were all the rage and Foursquare was getting all the press.

Speaker 3

01:56

And I remember sitting around saying, hey, I wanna build something but I don't know what I wanna build. What if I built a better version of Foursquare? And I asked myself, well, why don't I like Foursquare? Or how could it be improved?

Speaker 3

02:10

And basically I sat down and I said, I think that if you have a few extra features, it might be enough. 1 of which happened to be posting a photo of where you were. There were some others. It turns out that wasn't enough.

Speaker 3

02:23

My co-founder joined, we were going to attack Foursquare and the likes and try to build something interesting. And no 1 used it. No 1 cared because it wasn't enough. It wasn't different enough.

Speaker 3

02:35

So 1 day we were sitting down and we asked ourselves, okay, it's come to Jesus moment, are we gonna do this startup? And if we're going to, we can't do what we're currently doing, we have to switch it up. So what do people love the most? So we sat down and we wrote out 3 things that we thought people uniquely loved about our product that weren't in other products.

Speaker 3

02:57

Photos happened to be the top 1. So sharing a photo of what you were doing, where you were at the moment was not something products let you do, really. Facebook was like, post an album of your vacation from 2 weeks ago, right? Twitter allowed you to post a photo, but their feed was primarily text and they didn't show the photo in line, or at least I don't think they did at the time.

Speaker 3

03:19

So even though it seems totally stupid and obvious to us now, at the moment, then posting a photo of what you were doing at the moment was like not a thing. So we decided to go after that because we noticed that people who used our service, the 1 thing they happened to like the most was posting a photo. So that was the beginning of Instagram. And yes, like we went through and we added filters, and there's a bunch of stories around that.

Speaker 3

03:46

But the origin of this was that we were trying to be a check-in app, realized that no 1 wanted another check-in app. It became a photo sharing app, but 1 that was much more about what you're doing and where you are. And that's why when I say I think we've replaced check-in apps, it became a check-in via a photo rather than saying your location and then optionally adding a photo.

Speaker 4

04:08

When you were thinking about what people like, from where did you get a sense that this is what people like? You said, we sat down, we wrote some stuff down on paper. Where is that intuition that seems fundamental to the success of an app like Instagram?

Speaker 4

04:26

Where does that idea, where does that list of 3 things come from exactly?

Speaker 3

04:31

Only after having studied machine learning now for a couple of years, I like, I have a...

Speaker 2

04:36

You have understood yourself?

Speaker 3

04:39

I've started to make connections, like, we can go into this later, but obviously the connections between machine learning and the human brain, I think, are stretched sometimes. At the same time, being able to back prop and being able to look at the world, try something, figure out how you're wrong, how wrong you are, and then nudge your company in the right direction based on how wrong you are. It's like a fascinating concept, right?

Speaker 3

05:12

We didn't know we were doing it at the time, but that's basically what we were doing, right? We put it out to call it a hundred people and you would look at their data. You would say, what are they sharing? Like what resonates?

Speaker 3

05:26

What doesn't resonate? We think they're going to resonate with X, but it turns out they resonate with Y. Okay, shift the company towards Y. And it turns out if you do that enough quickly enough, you can get to a solution that has product market fit.

Speaker 3

05:39

Most companies fail because they sit there and they don't either their learning rates too slow. They sit there and they're just adamant that they're right, even though the data is telling them they're not right. Or they're learning rates too high and they wildly chase different ideas, and they never actually set out on 1 where they don't groove. And I think when we sat down and we wrote up those 3 ideas, what we were saying is, what are the 3 possible, whether they're local or global maxima in our world, right?

Speaker 3

06:10

That users are telling us they like, because they're using the product that way. It was clear people liked the photos because that was the thing they were doing. And we just said, okay, like, what if we just cut out most of the other stuff and focus on that thing? And then it happened to be a multi-billion dollar business.

Speaker 3

06:28

It's that easy, by the way.

Speaker 2

06:30

Yeah, I guess so. Well, nobody ever writes about neural networks that miserably failed. So this particular neural network succeeded.

Speaker 2

06:38

This output. Oh, they fail

Speaker 3

06:39

all the time, right?

Speaker 2

06:40

Yeah, but nobody writes about it.

Speaker 3

06:41

The default state is failing.

Speaker 2

06:42

Yes. When you said the way people are using the app, is that the loss function for this neural network or is it also self-report? Like, do you ever ask people what they like or do you have to track exactly what they're doing, not what they're saying?

Speaker 3

07:00

I once made a Thanksgiving dinner, okay? And it was for relatives and I like to cook a lot. And I worked really hard on picking the specific dishes and I was really proud because I had planned it out using a Gantt chart and like it was ready on time and everything was hot.

Speaker 3

07:21

Nice. Like I don't know if you're a big Thanksgiving guy, but like the worst thing about Thanksgiving is when the turkey is cold and some things are hot and something. Anyway, you

Speaker 2

07:29

had a Gantt chart. Did you actually have a chart?

Speaker 3

07:31

Oh yeah, yeah, OmniPlan. Fairly expensive, like Gantt chart thing that I think maybe 10 people have purchased in the world. But I'm 1 of them and I use it for recipe planning, only around big holidays.

Speaker 2

07:44

That's brilliant, by the way.

Speaker 3

07:45

Do people do this kind of- Over engineering.

Speaker 2

07:48

It's not over, it's just engineering, it's planning. Thanksgiving is a complicated set of events with some uncertainty with a lot of things going on. You should be able, you should be planning in this way.

Speaker 2

07:59

There should be a chart. It's not over.

Speaker 3

08:01

I mean, so what's funny is brief aside.

Speaker 2

08:04

Yes. Brilliant.

Speaker 3

08:06

I love cooking. I love food. I love coffee.

Speaker 3

08:08

And I've spent some time with some chefs who like know their stuff. And they always just take out a piece of paper and just work backwards in rough order. Like it's never perfect, but rough order. It's just like, oh, that makes sense.

Speaker 3

08:21

Why not just work backwards from the end goal, right? And put in some buffer time. And so I probably over specified it a bit using a Gantt chart But the fact that you can do it, it's what professional kitchens roughly do. They just don't call it a Gantt chart, or at least I don't think they do.

Speaker 3

08:38

Anyway, I was telling a story about Thanksgiving. So here's the thing. I'm sitting down, we have the meal, and then I got to know Ray Dalio fairly well over maybe the last year of Instagram. And 1 thing that he kept saying was like, feedback is really hard to get honestly from people.

Speaker 3

08:59

And I sat down at after dinner, I said, guys, I want feedback. What was good and what was bad? Yes. And what's funny is like, literally, everyone just said everything was great.

Speaker 3

09:11

And I like personally knew I had screwed up a handful of things. But no 1 would say it. And can you imagine now not something as high stakes as Thanksgiving dinner? Okay.

Speaker 3

09:22

Thanksgiving dinner. It's not that high stakes, but you're trying to build a product and everyone knows you left your job for it and you're trying to build it out and you're trying to make something wonderful and it's yours, right? You designed it. Now try asking for feedback and know that you're giving this to your friends and your family.

Speaker 3

09:42

People have trouble giving hard feedback. People have trouble saying, I don't like this or this isn't great, or this is how it's failed me. In fact, you usually have 2 classes of people, people who just won't say bad things, you can literally say to them, Please tell me what you hate most about this and they won't do it. They'll try, but they won't.

Speaker 3

10:05

And then the other class of people are just negative, period, about everything. And it's hard to parse out what is true and what isn't. So my rule of thumb with this is you should always ask people, but at the end of the day, it's amazing what data will tell you. And that's why with whatever project I work on, even now, collecting data from the beginning on usage patterns, So engagement, how many days of the week do they use it?

Speaker 3

10:34

How many, I don't know, if we were to go back to Instagram, how many impressions per day, right? Is that growing? Is that shrinking? And don't be like overly scientific about it, right?

Speaker 3

10:44

Because Maybe you have 50 beta users or something. But what's fascinating is that data doesn't lie. People are very defensive about their time. They'll say, oh, I'm so busy.

Speaker 3

10:59

I'm sorry I didn't get to use the app. I'm just, you know, but I don't know you're posting on Instagram the whole time. So I don't know at the end of the day, like at Facebook, there was, you know, before time spent became kind of this loaded term there. The idea that people people's currency in their lives is time, and they only have a certain amount of time to give things, whether it's friends or family or apps or TV shows or whatever.

Speaker 3

11:27

There's no way of inventing more of it, at least not that I know of. If they don't use it, it's because it's not great. So the moral of the story is you can ask all you want but you just have to look at the data and data doesn't lie, right?

Speaker 2

11:45

I mean, there's metrics, there's data can obscure the key insight if you're not careful. So time spent in the app, that's 1. There's so many metrics you can put at this and they will give you totally different insights, especially when you're trying to create something that doesn't obviously exist yet.

Speaker 2

12:07

So, you know, measuring maybe why you left the app or measuring special moments of happiness that will make sure you return to the app, or moments of happiness that are long lasting versus like dopamine short-term, all of those things. But I think, I suppose in the beginning, you can just get away with just asking the question, which features are used a lot? Let's do more of that. And how hard was the decision?

Speaker 2

12:40

And I mean, maybe you can tell me what Instagram looked in the beginning, but how hard was it to make pictures the first class citizen? That's a revolutionary idea. Like, at whatever point Instagram became this feed of photos, that's quite brilliant. Plus, also don't know when this happened, but they're all shaped the same.

Speaker 2

13:05

It's like a-

Speaker 3

13:06

I have to tell you why. That's the interesting part.

Speaker 2

13:10

Well, why is that?

Speaker 3

13:10

So a couple of things. 1 is data, like you're right. You can over-interpret data.

Speaker 3

13:18

Like imagine trying to fly a plane by staring at, I don't know, a single metric like airspeed. You don't know if you're going up or down. I mean, it correlates with up or down, but you don't actually know. It will never help you land the plane.

Speaker 3

13:32

So don't stare at 1 metric. Like it turns out you have to synthesize a bunch of metrics to know where to go. But it doesn't lie. Like if your airspeed is 0, unless it's not working, right?

Speaker 3

13:43

If it's 0, you're probably gonna fall out of the sky. So generally you look around and you have the scan going. Yes. And you're just asking yourself, is this working or is this not working?

Speaker 3

13:56

But people have trouble explaining how they actually feel. So just, it's about synthesizing both of them. So then Instagram, right? We were talking about revolutionary moment where the feed became square photos basically.

Speaker 2

14:14

And photos first and then square photos.

Speaker 3

14:16

Yeah. It was clear to me that the biggest, so I believe the biggest companies are founded when enormous technical shifts happen. And the biggest technical shift that happened right before Instagram was founded was the advent of a phone that didn't suck. The iPhone, right?

Speaker 3

14:38

Like in retrospect, we're like, oh my God, the first iPhone almost had, like, it wasn't that good, but Compared to everything else at the time, it was amazing. And by the way, the first phone that had an incredible camera that could like do as well as the point and shoot you might carry around was the iPhone 4. And that was right when Instagram launched. And we looked around and we said, what will change because everyone has a camera in their pocket?

Speaker 3

15:08

And it was so clear to me that the world of social networks before, It was based in the desktop and sitting there and having a link you could share, right? And that wasn't going to be the case. The question is, what would you share if you were out and about in the world? If not only did you have a camera that fit in your pocket, but by the way, that camera had a network attached to it that allowed you to share instantly.

Speaker 3

15:34

That seemed revolutionary. And a bunch of people saw it at the same time. It wasn't just Instagram. There were a bunch of competitors.

Speaker 3

15:41

The thing we did, I think, was not only, well, we focused on 2 things. So we wrote down those things, we circled photos, and we said, I think we should invest in this. But then we said, what sucks about photos? 1, they look like crap, right?

Speaker 3

15:55

They just, at least back then. Now, my phone takes pretty great photos, right? Back then, they were blurry, not so great, compressed, right? 2, it was really slow, like really slow to upload a photo and I'll tell a fun story about that and explain to you why they're all the same size and square as well.

Speaker 3

16:18

And 3. And if you wanted to share a photo on different networks, you had to go to each of the individual apps and select all of them and upload individually. And so we were like, all right, those are the pain points. We're going to focus on that.

Speaker 3

16:31

So 1, instead of because they weren't beautiful, we were like, why don't we lean into the fact that they're not beautiful? And I remember studying in Florence, my photography teacher gave me this Holga camera, and I'm not sure everyone knows what a Holga camera is, but They're these old school plastic cameras. I think they're produced in China at the time. And there, I want to say the original ones were like from the seventies or the eighties or something, they're supposed to be like $3 cameras for the, every person, They took nice medium format films, large, large negatives, but they kind of blurred the light and they kind of like light leaked into the side.

Speaker 3

17:10

And there was this whole resurgence where people looked at that and said, oh my God, this is a style, right? And I remember using that in Florence and just saying, well, why don't we just like lean into the fact that these photos suck and make them suck more, but in an artistic way. And it turns out that had product market fit. People really liked that.

Speaker 3

17:29

They were willing to share their not so great photos if they looked not so great on purpose, okay? The second part.

Speaker 2

17:37

That's where the filters come into the picture. Yep. So computational modification of photos to make them look extra crappy to where it becomes art.

Speaker 3

17:46

Yeah, yeah. And I mean, add light leaks, add like an overlay filter, make them more contrasty than they should be. The first filter we ever produced was called X-Pro2.

Speaker 3

17:58

And I designed it while I was in this small little bed and breakfast room in Todos Santos, Mexico. I was trying to take a break from the bourbon days. And I remember saying to my co-founder, I just need like a week to reset. And that was on that trip, worked on the first filter because I said, you know, I think I can do this.

Speaker 3

18:17

And I literally iterated 1 by 1 over the RGB values in the array that was the photo and just slightly shifted. Basically, there was a function of R, function of G, function of B that just shifted them slightly. It wasn't rocket science. And it turns out that actually made your photo look pretty cool.

Speaker 3

18:38

It just mapped from 1 color space to another color space. It was simple, but it was really slow. I mean, if you applied a filter, I think it used to take 2 or 3 seconds to render. Only eventually would I figure out how to do it on the GPU.

Speaker 3

18:53

And I'm not even sure it was a GPU, but it was using OpenGL. But anyway, I would eventually figure that out. And then it would be instant. But it used to be really slow.

Speaker 3

19:02

By the way, anyone who's watching or listening, it's amazing what you can get away with in a startup as long as the product outcome is right for the user. Like, you can be slow, you can be terrible, you can be, as long as you have product market fit, people will put up with a lot. And then the question is just about compressing, making it more performant over time so that they get that product market fit instantly.

Speaker 2

19:30

So fascinating because there's some things where those 3 seconds would make or break the app, but some things you're saying not. It's hard to know when. You know, it's the problem Spotify solved, making streaming work.

Speaker 2

19:49

And delays in listening to music is a huge negative, even slight delays. But here you're saying, I mean, how do you know when those 3 seconds are okay? Or are you just gonna have to try it out? Because to me, my intuition would be those 3 seconds would kill the app.

Speaker 2

20:09

Like I would try to do the OpenGL thing.

Speaker 3

20:12

Right. So I wish I were that smart at the time. I wasn't, I just knew how to do what I knew how to do. Right.

Speaker 3

20:21

And I decided, okay, like, why don't I just iterate over the values and change them? And what's interesting is that Compared to the alternatives, no 1 else used OpenGL. Right. So everyone else was doing it the dumb way.

Speaker 3

20:37

And in fact, they were doing it at a high resolution. Now comes in the small resolution that we'll talk

Speaker 2

20:42

about in a second.

Speaker 3

20:45

By choosing 512 pixels by 512 pixels, which I believe it was at the time, we iterated over a lot fewer pixels than our competitors who were trying to do these enormous output-like images. So instead of taking 20 seconds, I mean, 3 seconds feels pretty good, right? So on a relative basis, we were winning like a lot.

Speaker 3

21:06

Okay, so that's answer number 1. Answer number 2 is, we actually focused on latency in the right places. So we did this really wonderful thing when you uploaded. So the way it would work is, you know, you'd take your phone, you'd take the photo and then you'd go to the, you'd go to the edit screen where you would caption it and on that caption screen, you'd start typing and you'd think, okay, like what's a clever caption?

Speaker 3

21:33

And I said to Mike, hey, when I worked on the Gmail team, you know what they did? When you typed in your username or your email address, even before you've entered in your password, like the probability, Once you enter in your username, you're going to actually sign in is extremely high. So why not just start loading your account in the background? Not like sending it down to the desktop, that would be a security issue, but like loaded into memory on the server, like get it ready and prepare it.

Speaker 3

22:03

I always thought that was so fascinating and unintuitive. I was like, Mike, why don't we just do that? But like, we'll just upload the photo and like assume you're going to upload the photo. And if you don't, forget about it.

Speaker 3

22:15

We'll delete it. Right. So what ended up happening was people would caption their Photo they press done or upload and you'd see this little progress bar. Just go It was lightning fast.

Speaker 3

22:29

Okay we were no faster than anyone else at the time, but by choosing 512 by

Speaker 1

22:35

512

Speaker 3

22:35

and doing it in the background, it almost guaranteed that it was done by the time you captioned. And everyone, when they used it, was like, how the hell is this thing so fast?

Speaker 2

22:46

But

Speaker 3

22:47

we were slow. We just hit the slowness. It wasn't like, these things are just like, it's a shell game.

Speaker 3

22:52

You're just hiding the latency. That mattered to people like a lot. And I think that, so you were willing to prop with a slow filter if it meant you could share it immediately. And of course we added sharing options which let you distribute it really quickly.

Speaker 3

23:08

That was the third part. So latency matters, but relative to what? And then there's some like tricks you get around to just hiding the latency. Like I don't know if Spotify starts downloading the next song eagerly.

Speaker 3

23:24

I'm assuming they do. There are a bunch of ideas here that are not rocket science that really help.

Speaker 2

23:31

And all of that was stuff you were explicitly having a discussion about, like those designs and argument, you were having like arguments, discussions.

Speaker 3

23:41

I'm not sure it was arguments. I mean, I'm not sure if you've met my co-founder, Mike but he's a pretty nice guy And he's very reasonable. And we both just saw eye to eye and we're like, yeah, just make this fast or at least seem fast.

Speaker 3

23:56

It'll be great. I mean, honestly, I think the most contentious thing, and he would say this too initially, was I was on an iPhone 3G, so like the not so fast 1. He had a brand new iPhone 4. I was cheap.

Speaker 2

24:10

Nice.

Speaker 3

24:12

And his feed loaded super smoothly. Like when he would scroll from photo to photo, buttery smooth, right? But on my phone, every time you got to a new photo, it was like, ka-chunk, ka-chunk, allocate memory, like all this stuff, right?

Speaker 3

24:27

I was like, Mike, that's unacceptable. And he's like, oh, come on, man. Just like upgrade your phone. Basically, you didn't actually say that.

Speaker 3

24:33

He's nicer than that. But I could tell he wished like I would just stop being cheap and just get a new phone. But what's funny is we actually sat there working on that little detail for a few days before launch and that polished experience plus the fact that uploading seemed fast for all these people who didn't have nice phones, I think meant a lot. Because far too often you see teams focus not on performance, they focus on what's the cool computer science problem they can solve, right?

Speaker 3

25:05

Can we scale this thing to a billion users and they've got like a hundred, right? Yeah. You talked about loss function. So I want to come back to that.

Speaker 3

25:16

Like The loss function is like, do you provide a great, happy, magical, whatever experience for the consumer? And listen, if it happens to involve something complex and technical, then great. But it turns out, I think most of the time, those experiences are just sitting there waiting to be built with like not that complex solutions. But everyone is just like so stuck in their own head that they have to overengineer everything and then they forget about the easy stuff.

Speaker 2

25:45

I mean, also maybe to flip the loss function there is you're trying to minimize the number of times there's unpleasant experience, right? Like the 1 you mentioned where when you go to the next photo, it freezes for a little bit. So it's almost as opposed to maximizing pleasure, it's probably easier to minimize the number of like the friction.

Speaker 3

26:05

Yeah, and as we all know, you just make the pleasure negative and then minimize everything. So

Speaker 2

26:12

we're mapping this all back to neural networks.

Speaker 3

26:15

But actually, can I say 1 thing on that? Which is, I don't know a lot about machine learning, but I feel like I've tried studying a bunch. That whole idea of reinforcement learning and planning out more than the greedy single experience, I think is the closest you can get to like ideal product design thinking where you're not saying, hey, like, can we have a great experience just this 1 time?

Speaker 3

26:43

But like, what is the right way to onboard someone? What series of experiences correlate most with them hanging on long-term? Right? So not just saying, oh, did the photo load slowly a couple times or did they get a great photo at the top of their feed?

Speaker 3

26:59

But like, what are the things that are going to make this person come back over the next week, over the next month? And as a product designer, asking yourself, OK, I want to optimize, not just minimize bad experiences in the short run, but how do I get someone to engage over the next month. And I'm not gonna claim at all that I thought that way at all at the time because I certainly didn't. But if I were going back and giving myself any advice, it would be thinking, what are those second order effects that you can create?

Speaker 3

27:30

And it turns out having your friends on the service is an enormous win. So starting with a very small group of people that produce content that you wanted to see, which we did, we seeded the community very well, I think, ended up mattering and So. Yeah, you said that community is 1 of

Speaker 2

27:48

the most important things. So it's from a metrics perspective, from maybe a philosophy perspective, building a certain kind of community within the app. See, I wasn't sure what exactly you meant by that when I've heard you say that.

Speaker 2

28:01

Maybe you can elaborate, but as I understand now, it can literally mean get your friends onto the app. Yeah,

Speaker 3

28:10

think of it this way. You can build an amazing restaurant or bar or whatever, right? But if you show up and you're the only 1 there, is it like, does it matter how good the food is?

Speaker 3

28:23

The drinks, whatever? No, these are inherently social experiences that we were working on. So the idea of having people there, like you needed to have that, otherwise it was just a filter app. But by the way, part of the genius, I'm gonna say genius, even though it wasn't really genius, was starting to be, marauding as a filter app was awesome.

Speaker 3

28:50

The fact that you could, so we talk about single player mode a lot, which is like, can you play the game alone? And Instagram, you could totally play alone. You could filter your photos. And a lot of people would tell me, I didn't even realize that this thing was a social network until my friends showed up.

Speaker 3

29:06

It totally worked as a single player game. And then when your friends showed up, all of a sudden it was like, oh, not only was this great alone, but now I actually have this trove of photos that people can look at and start liking, and then I can like theirs. And so it was this bootstrap method of how do you make the thing not suck when the restaurant is empty?

Speaker 2

29:26

Yeah, but the thing is, when you say friends,

Speaker 4

29:29

I

Speaker 2

29:29

mean, we're not necessarily referring to friends in the physical space. So you're not bringing your physical friends with you, you're also making new friends, so you're finding new community. So it's not immediately obvious to me that it's almost like building any kind of community.

Speaker 3

29:45

It was both. And what we learned very early on was what made Instagram special and the reason why you would sign up for it versus say, just sit on Facebook and look at your friends' photos. Of course we were live, and of course it was interesting to see what your friends were doing now.

Speaker 3

30:00

But the fact that you could connect with people who, like, took really beautiful photos in a certain style all around the world, whether they were travelers, it was the beginning of the influencer economy. It was these people who became professional Instagrammers way back when, right? But they took these amazing photos, and some of them were photographers, right? Like professionally.

Speaker 3

30:23

And all of a sudden you had this moment in the day when you could open up this app and sure you could see what your friends were doing, but also it was like, Oh my God, that's a beautiful waterfall Or, oh my God, I didn't realize there was that corner of England or like really cool stuff. And the beauty about Instagram early on was that it was international by default. You didn't have to speak English to use it. Right.

Speaker 3

30:46

You could just look at the photos. Works great. We did translate. We had some pretty bad translations, but we did translate the app.

Speaker 3

30:55

And even if our translations were pretty poor, the idea that you could just connect with other people through their images was pretty powerful.

Speaker 2

31:03

How much technical difficulty is there with the programming? Like what programming language you were talking about?

Speaker 3

31:09

What was- 0. Like maybe it was hard for us, but I mean, we, There was nothing. The only thing that was complex about Instagram at the beginning, technically, was making it scale.

Speaker 3

31:23

And we were just plain old objective C for the client.

Speaker 2

31:27

So it was iPhone only at first? IPhone only, Yep. As an Android person, I'm deeply offended, but go ahead.

Speaker 3

31:34

This was 2010. Oh, sure. Sure.

Speaker 2

31:36

So it's like,

Speaker 3

31:36

so it's getting a lot better. Yeah. So I take

Speaker 2

31:41

it back. You're right. If I

Speaker 3

31:41

were to do something today, I think it would be very different in terms of launch strategy, right? Android is enormous, too. But anyway, back to that moment, it was Objective-C and then we were Python based, which is just like, this is before Python was really cool.

Speaker 3

32:00

Like now it's cool, cause it's all these machine learning libraries like support Python and right now it's super Now it's like cool to be Python back then. It was like, oh Google uses python like maybe you should use python and facebook was php like I had worked at a small startup of some ex-Googlers that used Python, so we used it, and we used a framework called Django, still exists and people use for basically the backend. And then through a couple interesting things in there, I mean, we used Postgres, which was kind of fun. It was a little bit like Hipster database at the time.

Speaker 3

32:34

Versus MySQL. MySQL. Like, everyone used MySQL. So using Postgres was an interesting decision.

Speaker 3

32:41

But we used it because it had a bunch of Geo features built in, because we thought we were going to be a check nap. Remember?

Speaker 2

32:47

It's also super cool now. So you were into Python before it was cool, and you were into Postgres before it was cool.

Speaker 3

32:53

Yeah, we were basically like not only hipster photo company, hipster tech company. We also adopted Redis early and like loved it. I mean, it solved so many problems for us.

Speaker 3

33:07

And turns out that's still pretty cool. But the programming was very easy. It was like sign up a user, have a feed. There was nothing, no machine learning at all, 0.

Speaker 2

33:17

Can you give some context, how many users at each of these stages? We're talking about a hundred users, a thousand users.

Speaker 3

33:24

So the stage I just described, I mean that technical stack lasted through probably 50 million users. I mean, seriously, like you can get away with a lot, with a pretty basic stack. I think a lot of startups try to over-engineer their solutions from the beginning to really scale, and you can get away with a lot.

Speaker 3

33:45

That being said, most of the first 2 years of Instagram was literally just trying to make that stack scale. And it wasn't, it was not a Python problem. It was like, literally just like, where do we put the data? Like, it's all coming in too fast.

Speaker 3

33:59

Like, how do we store it? How do we make sure to be up? How do we, like, how do we make sure we're on the right size boxes that they have enough memory? Those were the issues.

Speaker 3

34:10

But can you speak

Speaker 2

34:11

to the choices you make at that stage? When you're growing so quickly? Do you use something like somebody else's computer infrastructure or do

Speaker 3

34:20

you build in-house? I'm only laughing because when we launched, we had a single computer that we had rented in some co-host space in LA. I don't even remember what it was called.

Speaker 3

34:34

Cause I thought that's what you did. When I worked at a company called Odeo that became Twitter, I remember visiting our space in San Francisco. You walked in, you had to wear the ear things. It was cold and fans everywhere, right?

Speaker 3

34:46

And we had to, you know, plug 1 out, replace 1. And I was the intern. So I just like held things. But I thought to myself, Oh, this is how it goes.

Speaker 3

34:54

And then I remember being in a VC's office. I think it was benchmarks office. And I think we ran into another entrepreneur and they were like, Oh, how are things going? We're like, you know, try to scale this thing.

Speaker 3

35:06

And they were like, well, I mean, can't you just add more instances? And I was like, what do you mean? And they're like instances on Amazon. I was like, what are those?

Speaker 3

35:15

And it was this moment where we realized how deep in it we were because we had no idea that AWS existed, nor should we be using it. Anyway, that night we went back to the office and we got on AWS, but we did this really dumb thing where, I'm so sorry to people listening, but we brought up an instance, which was our database. It was going to be a replacement for our database. But we had it talking over the public internet to our little box in LA that was our app server.

Speaker 3

35:47

Yeah. That's how sophisticated we were. And obviously that was very, very slow. Didn't work at all.

Speaker 3

35:53

I mean, it worked, but didn't work. Only later that night did we realize we had to have it all together. But at least if you're listening right now and you're thinking, you know, I have no chance, I'm going to start a startup, I have no chance. I don't know, we did it.

Speaker 3

36:07

We made a bunch of really dumb mistakes initially. I think the question is how quickly do you learn that you're making a mistake? And do you do the right thing immediately right after?

Speaker 2

36:16

So you didn't pay for those mistakes by failure. So yeah, how quickly did you fix it? I guess there's a lot of ways to sneak up to this question of how the hell do you scale the thing?

Speaker 2

36:28

Other startups, if you have an idea, how do you scale the thing? Is it just AWS and you try to write the kind of code that's easy to spread across a large number of instances and then the rest is just put money into it?

Speaker 3

36:46

Basically, I would say a couple of things. First off, don't even ask the question. Just find product market fit, duct tape it together, right?

Speaker 3

36:55

Like if you have to, I think there's a big caveat here, which I want to get to, But generally all that matters is product market fit. That's all that matters. If people like your product. Do not worry about when 50,000 people use your product because you will be happy that you have that problem when you get there.

Speaker 3

37:13

I actually can't name many startups where they go from nothing to something overnight, and they can't figure out how to scale it. There are some. But I think nowadays, it's a, when I say a solved problem, like, there are ways of solving it. The base case is typically that startups worry way too much about scaling way too early and forget that they actually have to make something that people like.

Speaker 3

37:41

That's the default mistake case. But what I'll say is, once you start scaling, I mean, hiring quickly people who have seen the game before and just know how to do it, it becomes a bit of like, yeah, just throw instances of the problem. But the last thing I'll say on this that I think did save us. We were pretty rigorous about writing tests from the beginning.