Operators Ep 12 Transcript

TRANSCRIPT Thomas Dimson (ex-Instagram)

Delian:  Hi, everyone. My name is Delian. And I'm a principal at Founders Fund, a venture capital firm based in San Francisco. This is Operators where I interview non-VC, non-CEO, non-founder operators that make the start up world go around. Today, I'm interviewing Thomas Dimson, former director of Engineering at Instagram.

Thomas joined as one of the first 50 employees and the 16th engineer there. During his time at Instagram, he developed prominent features, such as the polling sticker on Instagram stories, hyperlapse, emojineering and author of algorithmic feed. He left Instagram in February of this year, and is currently working on a new project. I hope you enjoy the show. Cool. Thomas, thanks so much for taking time to hop on the podcast today. Excited to, chat some of the things that you've worked on.

Thomas:  Yeah. [laughs] excited to be here.

Delian:  Maybe, if you like to, take it back a little bit. you used to, study, computer science, at, Waterloo, and then later on, got a sort of master's degree, at Stanford. maybe, walk us through what got you excited about, computer science, originally. how you got into this whole field?

Thomas:  Yeah. I think it's been part of my entire life. I started actually maybe when I was 12 or 13 building half-Life modifications, which, was more involved I think than, maybe it is even now. So you have to, learn C++ figured out all these weird, esoteric KPIs that were in there, and, play around a lot. so a friend of mine built some mods together and, released them online.

And I always had that interest. I think the interest was more about product interest, it's like, "Hey, I wanna see something cool. I don't wanna play with my friends." and see as I was amazed to do that. and then when I- I got to school at Waterloo, I was, I was reluctant to go into it, but then I got, I- I joined, Waterloo in the math department, and it's very connected to the c- to computer science. And then I got my assed kicked in my early classes, which was, like, good for me because I was like, "Actually, I don't know everything," and I went back there [laughs].

Delian:  [laughs] When do you think Half-Life 3 is gonna come out?

Thomas:  Half-Life 3, man, we've gotten like a- a surprise with this VR, Half-Life: Alyx. I was [laughs] excited to see that. I played through it. It's a ... Yeah, I think it's probably the best VR experience I've had. ...

Delian:  [laughs] Very cool. and so your first co- ... let's say, professional, work was an internship at Bloomberg. what did you end up interning in while you're there? W- ... did you have a good experience there? And how did it end up influencing, future career decisions that you made?

Thomas:  Yeah. I don't know. It's like interesting. So that was my first ... At Waterloo, you do, six internships, at least in my program. Th- they're mandatory, so you alternate school and work. Bloomberg was one of those ... it was my first s- step into the US. so I think it was a good ... It was good for that. part of it, I think it was more of a culture experience than anything.

I also joined Bloomberg, during the financial crisis, actually just as [inaudible 00:03:08] was breaking. [laughs] interesting being behind the scenes of what's going on there. I had no awareness of what was happening. Honestly, like I saw Bear Stearns collapse and people were like, "You're gonna remember this." And I was like, "Oh, yeah. Sure, whatever." [laughs] "I'll just continue programming." and we are building like an attribution and risk management stuff. So very related although, like, all those models completely failed. yeah.

Delian:  And so then, your next couple of, ... I believe you guys call them, externships were, at Amazon. And then you, I think, went to, Amazon after- after, graduating. What sort of drew you to the company ... why- why did you end up deciding to, go there, and how- how is that experience?

Thomas:  I honestly can't tell you what drew- drew me to Amazon. I think ... A- at Waterloo, you just have a list of all these companies you can apply to. And I applied to a bunch of them. I got jobs, but most of them ... And, the job at Amazon I think said something about learning machines, which I was like, "I wonder if that's machine learning?" [laughs], but it was ... So I'm like, "Ah, it sounds interesting." Okay, ass- assuming it's a four months commitment. So I jumped on that and then ...

Actually the team was much more interesting than I- I had even imagined, because it was way involved with the early Kindle. And so Kindle was this new platform, and, we were a little bit behind the scenes of, what can we do with the data from this platform? and so my internship project was this ... it actually still exists. It's these little popular highlights. I don't know if you've ever read a Kindle book. yeah, it's got these little underlines under there, and that was like my claim to fame, 'Cause it was like, "Hey, we've gotten all these, this data about highlights that is like clunky. Can you make this into some product?" And I was like, "All right."

and it was a good enough experience. Liked the team, I had a great, ... I think Amazon gets a bad reputation in the industry sometimes, but I had a great experience there. No problems. And so I was like, "Yeah, I'll go back." and went there full-time.

Delian:  Yeah. I've definitely- definitely use that product on the Kindle. super cool to h- hear that you built it. and so from leaving Amazon, you then added an interesting sort of, you'd say, journey from, leaving Amazon to, Stanford, to ... it sounds like a startup, acquired into Instagram. Can you just walked through, that whole journey to glory? How the hell that all happened and how you eventually ended up [crosstalk 00:05:18]?

Thomas:  Honestly, sometimes I get lost myself. I joined Amazon. I was working there on the same team for quite a while. it was a great team and I was excited to be there, but I think I just had an epiphany, and it was like in a coffee shop. And it was, ... I overheard some conversation about mathematics, which is what I studied in addition to computer science. And I was just like, "I need to go to grad school. Like I, I'm just not ..." I could probably be an Amazon for the rest of my life and have a very comfortable life, but it's not really being true to myself. I rather would learn something new.

and so I applied to Stanford, for a master's, which is like a two-year program. It's mostly coursework based, but, got in and it's "Okay, I guess I'll go to Stanford." [laughs] I- I ... being Canadian, and I, I've obviously heard of Stanford, but it wasn't like so ingrained in me. It wasn't something I would have ever considered applying to for undergrad.

so yeah, I went down there, studied AI, which was just as ... it was slightly before deep learning, but like right- right in there. and was most interested in natural language processing because of this Kindle interest. and then, I guess I met a friend of mine from undergraduate f- for beer, and he's "Hey, you wanna ..." he's, Ivan Karpenko. But he's ... He had these two brothers and himself were forming a company. the amazing Karpenko brothers, we used to call them.

And he's "You wanna meet my brothers?" And I said, "Sure. Okay." And I met the brothers and I was like, "These are the smartest people I've ever met in my entire life." [laughs] And so I was doing my masters. I was like, "I'll just work part-time for these guys." they were making video stabilization for the iPhone. And the product that they had was using, measurements from the, gyroscope, which just made it, cinematic quality. I was like, "Okay, yeah, for sure, I don't care what they're doing, I just want to be around these people." 

Delian: And so the company got acquired into Instagram and Instagram at the time had already been acquired by Facebook. 

Thomas: Y- yeah.

Delian:  So you [crosstalk 00:07:18] acquired by the big behemoth at that point.

Thomas:  Yeah. It's ... it's I don't know, like a big fish eating a little fish, eating like a minnow or something like that. So it- it was acquired about a week before I started there, to be honest with you. And then, so this- this company, which was called Midnox, they made a product called Luma. it was a YC company. And they were acquired right- right at, ... yeah, just at the time that Instagram was launching video, which is about six months after the acquisition by Facebook.

And I was still doing my masters. I was midway through it and I was like, "Okay, I guess I'll just go along for this. I have no idea where this is going to lead." so this seems super silly in retrospect, but I was like ... honestly, I didn't think much about Instagram. I was like, "Yeah, it's cool product, Let's see what that's all about." and at the time, Instagram was 40 people.

So again, in retrospect, I was like ... I should've just been ... had the attitude of a rocket ship that I- I would have just grabbed onto and hang on, but I just thought ... didn't take much. I came there and they gave me the option of either being an intern or quitting to return to studies in three months. I chose three months quitting option, and then after, after that summer ended, they were like, "Okay, you can just stay along part-time and continue working. So finished my masters and continued working there.

Delian:  Nice. yeah. What was it like, when you first joined? What were some of the early you know projects that took on a while- while you were there?

Thomas:  It was a weird time. and again, in retrospect, like Instagram had 40-ish people, and it already had more than 150 million users. And so when you think about that kind of ratio, it's there are very few companies that have that. and so I didn't ... again, no awareness at the time ,just felt "Okay, sure. I'm just an intern at some company." But like that level of influence is it's pretty unique to Silicon Valley, I'd say.

and the company, it was small, everything was on fire. We were still in AWS. It was still very isolated from Facebook. and so a lot of my time was just spent, fixing servers. our deployment process was a mess. It was just like everything that you would think would be fixed by a hundred million users was not fixed.

And, I remember even ... My first diff was reviewed by Mike Krieger. And I Googled his name and was like, "Why the hell was the co-founder of Instagram reviewing my diff. That's crazy." [laughs] Like ... yeah. So pretty hectic time. I guess I had a machine learning background because of my schooling. And so I was able to, look for projects and that kind of area, everything was horizontal. It wasn't like anything was organized into divisions or anything like that. So it was a good chance to play. ... Yeah

Delian:  It sounds like one of your, most impactful projects was like the, algorithmic feed, project. Was it primarily because of your, machine learning that they asked you to go tackle That? How did that sort of project come about and shift? and it's feels like it's been obviously one of the most meaningful, changes to the core Instagram product.

Thomas:  Yeah, a little controversial, [laughs]. But, ... Yeah. that was like actually a little bit later, later on, but, I had worked on an explorer pro- the Explorer project for Instagram, which was just like the second tab, which just showed you cool shit on the platform. which made sense and was pretty successful, especially relative to what was there before, which was this like popular page. [crosstalk 00:10:46].

Delian:  And how do you, define ... How ... Did you just define success was basically just like engagement on that, tab, time spent on it?

Thomas:  [laughs] so part of being early in those companies is you, barely measure anything and like measurement is a challenge. that was fortunately a pretty unambiguous product. We ran a small test, ninjad this thing. It wasn't very announced, by a personalized explore tab. and there is no metrics that were particularly down. It was some, ridiculous number of ... even just engagement with the tab or coming back to the tab, that kind of thing, it was just like ... it was off the charts. The P value rounded to zero and all the experiments were like, "Okay. Sure, this is, good news."

and then feed was a little bit later on. It was actually in New York, we were looking for a project. We just moved out there, and the company ... we had a company or had a- an office of maybe five engineers. and wandering around in circles, not sure what we were supposed to do. And there was this ... It was in the air. People would always talk about Instagram having an algorithmic feed, because Facebook did, and it was pretty successful for them.

And I was just like, "Okay, let's just fucking do it." like [laughs] ... it's- it's been talked about, it's inevitable that I think we're going to try it at some point and I'd rather this be done right than be done wrong." And so we just ... we- we went for it. And so that was like the origin at least of that, project.

Delian:  And for both the projects, let's say, walk through how you went about solving them. was it mostly like academic problem of, "Okay, what are the various signals that we're going to read into so that we can feed this into some neural net and figure out basically, how- how to recommend?" or was it like, "Hey, we're actually going to use like, super- super simple, ML, and it's mostly figuring out the, systems engineering, like, how to scale this, so that you can like generate, algorithmic, recommendations for, again, 15 million, different users with different habits"?

Thomas:  Yeah.

Delian:  Like, how did- how did ... walk me through just like the step by step of the architecture of what types of, engineers you needed and how you solve those problems?

Thomas:  [laughs] it's, it's a little bit of everything. certainly, I think if I was more ... I think when you're in these things, you just don't realize what you're doing in a sense you're like really just ingrained and you ... It felt like we could tackle these- these problems, even though now, in retrospect, it's that's kinda crazy, At the time we were doing feed ranking, I think it was something like 500 million people on the platform. and we were just a small group of engineers. I don't think it was technically motivated. And I think that's- that's something I've always taken. it's not the idea of "Hey, this is a cool technical challenge, let's go after it." The ... It's more of "There's a product issue here and what does it ... what needs to be solved." And particularly or, let's just pick our feed 'cause I don't think people always talk about motivations and stuff. It's just the network had ...

the motivation was that you ... when you went to Instagram, you would miss some of your feed 'cause ... when you scroll down and see everything. And over time, there were more and more people that posted, and there were some people that posted their regular amount, maybe one a day or something like that. And the people that posted more and more often started to dominate your feed more and more.

And so we're like, "Okay, we need to somehow fix this challenge and the product." and there are a couple of solutions that are not algorithmic feed, but the algorithmic feed one is like probably the one that works where you say, "Okay, instead of keeping this in a stable order, we're going to bring up the stuff that you would actually enjoy from the post that you otherwise would've missed."

And so that's like your product motivation, now you actually go out and solve that while you try to figure out a way of modeling the users so that you try to see what they would actually want to see in their feed. what would they want to ... not want to miss? And that tends to go towards these things that you can model, different engagement things that you can do, to model with whatever machine learning things you- you can.

And, yeah, the models, they're interesting. We can go deep into what these [inaudible 00:14:48] and stuff like that are. But I think it's much more interesting to think of this as "What are they solving and what are we actually trying to do?" Which is make the product better, 

Delian: And I feel like you're an example o- ... I've heard this from others that Instagram, of someone that was able to, have pretty meaningful impact on the product, both from this explore page and, algorithmic ranking of feed, despite not necessarily, being the person that like tried to, have as much head count reporting directly to them.

Thomas:  [laughs] Yeah.

Delian:  can you talk about was that intentional or was it more just this is the type of style of work that I like to have? and then what do you think let's say, e- enabled you to do if somebody else were listening to the podcast and, is an IC at a company, but like really wants to have meaningful changes without necessarily, having people report to them, w- what- what's your recommendation for how they go about doing so?

Thomas:  Yeah. You could do it. I think it's m- mostly, it's a bit of a personality type for me, I'm definitely a generalist, and I like my scratch my itch in interests. And generally, that means that I go very deep into something, and then I snap back and hop on to something else. I think that, just like having some interface and having the ability to actually change the product, and, ...

in some ways, engineers are really the best people to actually consider what the issues are in the product and how to solve them. And then the question that I think sometimes comes up at these big companies is like, "You have no opportunity to actually, show people what you're doing or make a product argument."

and so I think I was just loud and annoying. And usually the way I would do this is I would just build it. if I thought something should be built, I would built it ... build it, show it to people and be like, "Okay, it's already built. Is this something we should do or is it not something we should do?" And at that point it was pretty usually unambiguous. there were certainly things I felt they kinda sucked, and the answer was no, but, there was, I think, more often than not, I had like actually had success there.

Delian:  Do you have a favorite project that, you had sucked, that didn't quite make it as a product?

Thomas:  [laughs] boy. i- I remember I wrote something down and I was like, "What was the worst project I ever worked on?" And I guess maybe one thing that I- I ... This has shown up in Instagram over and over again. I won't to say it just une- unequivocally sucked, but one thing we- we've built so many iterations of the Instagram explore tab at this point. So V1 was the popular page, V-2 was the ... the V1.1 was personalization.

There was this lost product we had called Explorer V2, which was this, ... the idea ... I can make a sane product argument for it. It might sound good, maybe. Which is, as we did more personalization, there was less shared context. And so we really wanted to integrate, more- more of a sense of culture, more of a sense of Zeitgeist into the explore tab.

How do you do that? we were going to introduce trending hashtags. So it's like the cultural zeitgeist of the day and, editorial content in Explorer. So things that we decided, this is what everybody should see. so that- that sounds good. except when your platform is 500 million users, and there's no way editorial content scales [laughs] to people's tastes and trending is a hot mess that you need to clean up all the time.

And, it w- it was out there for- for quite a while, but o- over time, what we discovered, just it's hard to be what you're interested in what we're trying to imply, modeling your behavior. And that was probably one of the more difficult projects that were ... we worked on. And plus we were interfacing with editors, content editors that worked at previously news companies. And so that was just like ... It was definitely [laughs] a culture clash to say the least. so it was kinda interesting. Yeah, that's probably my- my example of a failure. There might be other ones too, but, yeah.

Delian:  You talked a little bit earlier about, starting the New York office and having, a five person engineering team there. can you walk through ... you said that you don't necessarily love going into like management, but obviously, like starting a new office and hiring engineers, that there as ... the world [inaudible 00:18:56].

Thomas:  [laughs].

Delian:  I- I imagine there were also challenges of being a ... I think you guys were one of the, the first remote engineering offices that had, multiple people.

Thomas:  Yup.

Delian:  what was the process like? Anything that you'd do differently, now ... if you were ... your new thing, let's day, I start, new remote engineering office?

Thomas:  Yeah. I- I think it went, so in retrospect, very well. it was the first remote engineering office for Instagram. I think there's a theme that's emerging in this conversation, which is just I don't think we realized at the time, why it would be successful. It wasn't intentional. it wasn't like, "Hey, we planned for this to work exactly right.

I had always signaled that I'd be interested in working in New York 'cause I like the city. and I signaled to this probably three months into my career there, but over time, and I would bring it up periodically and, it's just it took maybe two or three years, but eventually, the stars aligned where it was like, "Hey, the Bay area is getting expensive to hire."

There was encouragement from, big brother Facebook to actually hire outside of, Bay area. And so since I had already signaled my interest, it was like a natural fit. And plus, I had, the cultural context of Instagram. in terms of decisions that we made a very controversial decision at the time, which was ... So Instagram and Facebook were in the same office, office campus, whatever you wanna call it in, in Menlo park. And, I felt it was really important as we were getting started to actually pick our own space and not be in the same office.

I think it was more of just "Hey, we want to nurture something here and, it might be really distracting." Plus the desks that we were offered were, like, next to recruiting, and it was going to be very loud. and so we went off with, myself and then one other engineer, to just sit around and see what we can come up with. and so the separate office, we started hiring people.

And I think, the big inflection point is like, while we started working on this feed ranking, and that's a pretty big effort. And so we had the right mix of, we had our- our separate culture that was being nurtured in- in this office. We had camaraderie and everything like that. We had the people with the cu- cultural context from Menlo park, which is like myself and some engineer, Lindsey. And, then we had a very important project, which is often a mistake that people make with remote offices is that it's considered like an afterthought or "Oh yeah, we've put, something that's maybe not business critical there because the chance of failure is very high."

And so those three things made it so that we were important. We were recognized we had the cultural context from Menlo park and we had the awareness. and we also just grew, we were like a kind of a family as it was evolving. eventually, when we merged back into the main Facebook office ... but at that point, probably were about 200 people, 150 people at that point, it was ... we had already built our culture and it was like a st- strong foundation. So I- I think it went really well. it could have either ... lots of things could have gone wrong, not even starting feed ranking or something like that. it was ... It wasn't obvious that was going to happen, I'd say. ... Yeah.

Delian:  Yeah, Imagine just even the process of convincing, the execs of Instagram that, such an important product change, would be like run by a remote engineering team. seems like a sort of Herculean effort. Like-

Thomas:  [laughs].

Delian:  [inaudible 00:22:23] over the line or is it one of these things where you just, went and built and said, "Look, we're the remote team and we built this, and this is going to be good. And we need to like, run this ... ha- ... Wha- what was the process for convincing them?

Thomas:  yeah, I don't think it was as controversial as that. I had a very good relationship with Kevin and Mike, the founders of Instagram. We were very tight. and so I don't think it was like ... because it was my sponsorship, plus we had the ... we were ... we had a big team that was ... not big, but a s- small team that was ready to go on this. it seemed like a natural fit. Plus it was like the type of thing, this attitude of just like doing it and then being like, "Hey, we built this thing. Do you want it?"

it wasn't exactly like that, but it was closer to that and, a big top down product change. It was like more or less, we had some infrastructure that we had built from the explore tab and "Hey, we can probably just apply this. We should see if this works or not." and so it- it just naturally evolved like that.

but I think it was the ... it was a good decision by them to let that happen. I don't know. It probably would have been more damaging for them to pull it away from us. I think that would have been, demoralizing and I think it was the right decision in the long run for the company too. ...

Delian:  you also had a ton of like side projects that ended up, I think having a decent amount of success from probably my favorite, the polling buttons on stories.

Thomas:  [laughs]

Delian:  Like I remember specifically [inaudible 00:23:41] being like, "They should have had this, forever," to hyperlapse videos and emojineering. you ... how did those ideas end up, coming up? What was the response from colleagues when you sorta suggested them or wanted to try to roll them out? and how do you think, engineering cultures can encourage that type of, experimentation?

Thomas:  Yeah. I think it goes back to that thing we were talking about, just having, having an ... a forum that engineers on the ground can actually interface with product managers, product leaders, executives. for us, it was like hackathons. And I was very .... I ran most the hackathons in New York, started that kind of culture and actually made it like a thing.

sometimes hackathons, especially now, are I will say, they're not in the forefront of anyone's mind [laughing]. but, I was very adamant that we actually produce products in those things. And so my personality is m- like, when I see an opportunity that I really think should exist, I make it happen. I'll go through polling stickers 'cause I think it's probably the most interesting story there. Maybe hyperlapse was also was initially started.

But like polling stickers in particular, we were working on feed ranking, and stories, had just come out maybe a year prior, half year. And, I liked a lot about stories, but one thing that always bothered me was that when you post, your post goes out to the ether. So- so you can post, I don't know, does anyone see it? Yeah, maybe we can go and check, but you don't get notifications, you don't get any feedback. It's just it's like casting a bottle into the ocean or something. Throw a bottle in the ocean and maybe somebody reads it, maybe somebody doesn't.

and I knew that from feed ranking that actually that signal of engagement, the like button is actually quite important. Not even from a machine learning perspective, but just like a person giving a like really enjoys giving that it's "Hey, I signal that I'm interested in this person receiving it," and it's "Hey, my audience is really engaging with this and they like this particular post."

And so I was ... always felt that we should have something similar into stories bothering me. at some, at- at some point I was talking to one of our data scientists and she was mentioning that she had this crazy idea for a tamagotchi sticker, where you like nurture a pet on your story. And I was like, "That is a crazy idea. It's probably the craziest idea I've ever heard." wait a minute, if we had op- ... the sticker idea was a- a light bulb. I was like, the problem with likes is they're obligat- obligatory, and, it can get, like a really negative feedback loop. But if you're opting into it, if you're like, "Hey, I want to post this on my story. I want you to engage with it," that's a much more healthy way of doing that.

And so I hacked together, like with a small team that I ... the people that I work with a like sticker, for stories. And we- we tried it out, presented it to different execs and stuff like that. And it was cool, but what was interesting is that you could put multiple likes stickers. And what was quickly evolving was like, "Oh, actually the interesting way of doing this is to put one like sticker over here, one like snicker over here and you'd choose between two parts of the photo. And so I was like, "Oh, actually we should build a poll. That makes sense." [laughs]

And, yeah, so that's kinda like how that- that evolved. It was not particularly linear and I don't think people really anticipated the amount of success that it actually had. It was pretty significant for the business, and it opened a new domain of stories as well.

Delian:  And ... Yeah. so what ... when did you know that it was going to have success? How did that end up actually impacting people's usage of stories? What we're the- the future, let's say, iterations or things that kind of came out of that?

Thomas:  I'm ... It's I have- have to say, I- I hope for the best, but I kind of plan for the worst. And so I, never anticipated it being more than like a curiosity, scratch my own itch. and so other people were telling me, it- it's going to be big. but I've heard that before. And sometimes it's not big, sometimes it is big.

so really it was just a march day. We like, put it out there and it was a meaningful impact on people's creation habits. And I was like, "Okay, this is like a product that's resonating with people." And, the future iterations were obvious. It was like, we've just opened a new domain of interactive- interactivity into this platform, which is something that hadn't really existed in the format on Snapchat or on Instagram.

And so it was, what can we do with that? we can do anything really, voting is the- the star point, but there's you could do comments, you could do slidey stickers, you could do all kinds of crazy things. And the great thing is because they are stickers, it's not that much of a product commitment; it's much more flexible. and so that's ... That ... a lot of things, it was like, that was the- the starting point. And then I took myself out of that and there's like teams that were working on it from ... to- to develop it from there. And I- I think it was great. Yeah.

Delian:  I can see why, ... In a situation like that, I feel like sometimes people want to ride that wave, and they're like, "I want to become like the VP of stickers and like [inaudible 00:28:41]-

Thomas:  [laughs]

Delian:  ... engineering team, and, experiment with all the various, potential like interactive stickers. what, what makes you want to just be like, "Okay, I'm going to, le- leave this and go on to the- the next sort of experiment or thing"?

Thomas:  I think it's just a ... Yes. I think if I was motivated by, being a longterm career company person, maybe that's when I found THE success, I'd be like, "I may stick to this," but not. I just wanted to build cool things, and they were interesting to people and resonated with people. and I know myself enough to know that I'm much better in an early stage than I am in a later stage. and I can be a specialist. It's ... I'm capable, but I think it drives me crazy, and I get de-motivated by it. And so it's very clear in that case.

I think it's ... I was in a, in many ways, fortunate position 'cause I've been with the company for so long, I'd grown up with the company. I felt the reputation there that I had the flexibility to jump around and people not really questioned what I was doing. and so I just exploited that ... the ability to do that and fool around with things. Yeah.

Delian:  You mentioned like one of the important things of, Instagram's culture, or generally places that, experiment quickly, move, ship out new products is like engineers being able to like access ... engineers on the ground being able to access the exec team and founders.

Like how do you think ... w- what do you think like Kevin and the founders did to facilitate ... were there ... Was- was it just like having these hackathons? Was it just being like super open to having individual engineers pitch product ideas and setting aside time for that? what were the aspects of the culture that you think enabled, some of these, let's say features that you- you talked about?

Thomas:  Yeah, i- it evolved over time. when you're small, everybody knows each other, you can just fool around. And, I think the ability ... there's some things that are at even Facebook broader culture, like the ability to, work on the main line of the code- code branch and actually code up features that are not, deployed to everyone and be able to show people like, "Hey, here's what I built." Or even better, 'cause it's a social network, you have your friends that are working with you. And you build it, and then you say, "Try this thing out."

so it's like more of a grassroots kind of effort. I don't know. Kevin and Mike, like Mike was always very hands on and very respected by the engineering team. And he was that voice in upper, management, whatever you want to call it, speaking for engineers and had the respect of us. And I think he was like a very, inspiring figure to a lot of people that, hey, you can actually get to that point and really have an eye to the ground and be grassroots. And so I think just him being present there was a big motivator for people.

and then honestly, I was adamant about creating that hackathon culture 'cause I knew it worked for me. we didn't talk about hyperlapse very much, but it's ... I spent like a year trying to get that thing shipped, grinding gears with lots of people in the meantime. And really the last thing that I think we got over the edge was like, "Let's just make it into a product that a hackathon and make it a huge forum that everybody looks at it and gets excited about it, and then it will be very obvious that this thing needs to get out the door."

and so I think having that form is just, it's just important and not underestimating engineers in terms of I think you can get pigeonholed into other ... the coders that pick up features and, move them over. I'll say one more thing just briefly is, I also think you can go the o- opposite direction where sometimes engineers believe that every idea that they have is great.

The truth is that getting from an idea to a product is incredibly challenging. N- nevermind a successful product, but like a product in general. and you shouldn't underestimate the amount of work and- and sometimes stupid work. Like stupid work in the sense of you see something in your head and you're like, "This needs to exist," You need to convince every single other peop- ... o- other person around you, why. [laughs] And sometimes you might not even know, it may be gut feeling, but you have to justify yourself. and so I think that just because you've built it doesn't necessarily mean that it's good. ... Yeah.

Delian:  Yeah. Maybe we can dive into hyperlapse a little bit. I imagine that it was a much more significant ... I, just trying to understand the level of, engineering effort. I feel poll buttons and stickers feel a lot easier than, the times [inaudible 00:32:59]-

Thomas:  [laughs]

Delian:  ... during, video post-processing and, i- implementing hyperlapse. ... And it sounds like you, granted some gears with folks trying to get over the line. You talk about wha- what was difficult about getting it over the line. Why did you end up finally, y- deciding to do it in the hackathon, what ended up eventually getting people excited about it, and now it's- it's, it's its own standalone app? ... Yeah.

Thomas:  Yep. I- some- some things about these ... these journeys are always non linear and you'd like to take all kinds of unexpected terms, and it wasn't as intentional as maybe it would look in retrospect. but I had a chip on my shoulder 'cause I joined this company that did video stabilization and never got a chance to play there with their technology.

And so probably about a week into my career at Instagram, I was ... Honestly, actually my first couple of weeks on Instagram, I wasn't very happy. I was working in dev ops and I was like, I'm an AI person and I wanted something more interesting. i- if I had quit, that would have been a huge mistake, but it was actually in my mind that I might do that.

and so I was looking for things that were a little bit more intellectually stimulating. I was like, "Oh, we have this technology for video stabilization. I wonder what happens if I use it for time lapse?" And so it wasn't trivial. I got the iOS code base. I'm not really an iOS full-time type of person. But I grabbed it and I was, "Okay, let me like fool around with this. What's the simplest thing I can do? Let me just drop frames. Let me just ... If I capture eight frames, let me drop seven of them. And then I'll apply stabilization, on the non-dropped frames."

And then I tried it, I went ar- ... walked around campus, took a few videos, and I was like, "Holy shit, I have something in my pocket that nobody on earth has." And I felt like I needed to ship it. I posted it in kind of internal groups and stuff. People are always like, "Yeah, that's great. Cool." and I was like, "Why don't you see what I see? we need to get this out there."

and so honestly it actually died for about a year and I was sorta working on it in my spare time. And when we had this hackathon coming up, I was like, "We gotta get this out there." And so I grabbed a designer friend of mine and another coder of mine. and we're like, "Let's just build a product." And the designer figured out an interface that could possibly work for this thing. we built it and then I spent about a week before we had this, called Prototype Forums where you present.

I spent a week capturing footage, just going around, go to Stanford, went to the aquarium. I was like, "Let me just take anything that has not been done in TimeLapse and do it in TimeLapse." And I edited the video, got some music, and that was our- our prototype forum. So here's a amazing demo of this [laughs] thing.

I remember we presented it. We actually got invited to ... this was the Facebook Prototype Forum, not the Instagram one. And got invited to present in front of Zuck. And one day I rode my bike from Stanford campus over to Zuck, where I was in school and we presented it. And he was like, "This is cool." And I remember exiting that meeting, and a product manager came out. he was literally like, "Oh man, it was pretty cool, And then exiting that meeting somebody who was in the room came and said "Zuck loved your idea." [I was like, "Okay." and actually, fortunately, after that, both Kevin and Mike were like this is a good product. we should make it part of the Instagram family." And we polished it up and shipped it out.

Delian:  This is actually during your time where you're doing part time master's degree. Part-time, part-time-

Thomas:  Yeah. [laughs] yeah.

Delian:  Wow. ...

Thomas:  It was a, again, weird time in my life, I don't know how the hell I ever did it. But sometimes, being busy is good. it's just like you ... It's like sometimes work makes you more engaged and stimulated and exciting. And, learning stuff, and being able to actually apply it right away, it was really exciting for me. And so ... I- sometimes think being busy, not stupid busy, but stimulated busy is actually a good state to be in. You'll get all kinds of crazy ideas.

Delian:  it seems over the course of, your career at Instagram, you had to interact With a lot of different teams. You've mentioned, obviously, interacting with the exec team. I'm sure you got to interact with, product managers, designers, et cetera.

can you talk about, what were the, let's say, communication styles? what- what teams were easier to partner with versus not? what- what, and what sort of strategies would you recommend to folks that are maybe in a position like yours at a large company, 

Thomas: Yep.

Delian:  ... that are trying to figure how to best interface with, the various groups that, they need buy in from, and, and need to work with?

Thomas:  Yeah. So I think, ... I- I've definitely made a lot of mistakes around the way. And I- I think maybe it's- it's more helpful to start with mistakes that I've made. So when you're building products, you're interfacing with everyone. there's no team that you're not going to talk to. You're gonna talk to people that do content strategy. So what strings do you use in the app or how to guide people through it, designers policy, 'cause you're dealing with social media privacy, 'cause you're dealing with social media, spam fighting teams. th- this is the kind of thing you just have to go through. Not to mention all the- the wellbeing fonts you need to do about, reporting content and all that sort of stuff.

the mistakes I think I had made was just assuming that people had the same context that, I did. And I have a very, I have a very like strong user model about what- what works, what, doesn't, how this technology works, like how AI works. And I would show up to meetings and just assume that everybody had the shared context of understanding how a ranked feed, what that means. everyone sees something different. It doesn't ma- ... Like ... and the mistakes I would say I made is just assuming that context ... And then when people didn't have that context, it was like a little bit grinding or I might get frustrated or I might not hear them out the way I needed to hear them out. And so I think that you just approach it like everybody ... The truth is everybody does have, this cognitive diversity and it's like a good thing. It's not a challenge. It's it's more just like people are coming from a different point of view and you should see how that fits. And if they don't have the context that you have, you- your job is to translate that, and make sure that people understand these systems at the same level that you do, or at least can approach that level. So yeah, I think that's the main thing.

The other thing I'd say is, that's a very fuzzy answer. Sometimes the best thing to do is just show up with like hard data. That's well explained that you've actually researched yourself and can discuss the flaws with them and the- the good parts of it, because at the end of the day, that's going to trump, any- any kind of like opinion that is in the room if you can prove or disprove any, any questions about what you have. And that goes for anything, like anybody shouldn't be responding to data.

Delian:  Make sense. so you finally just had to leave Instagram in February of this year. right before the world crashed. what-

Thomas:  It's actually bad timing [laughs].

Delian:  Yeah. It's [inaudible 00:39:51] your stable job, instead of starting a new project. Yeah. w- wha- what led you to finally decide to leave? And, can you talk A little bit about what you're working on now or what you're exploring?

Thomas:  Sure, totally. yeah. I don't think I have a very good, story to tell. I've- I've been at Instagram for seven years. I've watched that platform grow from a hundred million to billion, users. And, I had a very good career that- that I was quite happy with, but, the loss of Kevin and Mike as founders was felt to me. and I had moved into management in the last year at a pretty high level of it, the principal engineer. And so I moved into being a director at this company, directly, which was challenging.

I also had a lot of personal challenges going on at the same time. And so all of these things aligned plus the fact that, I really didn't see myself as a lo- ... my career aspiration is not to be a VP at Instagram or Facebook. I like building things. I like getting on the ground and, failing on my own.

and so all that, I was like, I need ... if I'm gonna ... Maybe I'll come back with my tail between my legs and be like, "Okay, I- I wish I had a directive job at Instagram again," but I thought that I can do that when I'm ... 10 years from now much older and ready to have a very mature perspective on these things. and that was it. It was like, "Okay, time to go. It's scary, but, I'll- I'll make it, I'll survive. [laughs] There'll be some- something on the other end."

and in terms of shock projects, I don't think I can tell you enough that it was like a pedal to the metal experience, and I was in a bubble of ... The scale of that company can warp you. And, when you exit, I think it's extremely important to take some time to yourself. And, for me, I- I traveled around China a little bit, learned Chinese a little bit, and-

Delian:  During COVID or ...

Thomas:  or not during COVID. So I- I actually spent a couple of months doing that right before the ... my official leave in February. which strangely enough, yeah, it was, there [laughs], there was COVID in Wuhan. I was in Guilin, so it didn't really affect me. But, ... Yeah. And so that was important to me.

And then once COVID started actually, was, fortunate enough that I spent some time and being able to do remote learning at Stanford. And so I just actually went back to school a little bit for six months. I, studied things that I had ... AI changes all the time. And so actually I could take the same class or talk to the same professor that I did seven years ago, completely different, four years ago, completely different. And so I- I tried to get up to speed with some of the latest, that field's changing like crazy. And that's also what guided me towards doing products as well.

Delian:  you have this a website that I thought was funny though, This Word Does Not Exist that uses machine learning, like catalog, made up words. 

Thomas: Yeah. Yeah. Yeah.

Delian:  So- so this machine learning application, can you just, tell us a little bit about that kind of site?

Thomas:  [laughs] Sure. Same motivation, just like I, ... This is after I left. And so I was like ... I actually tried this ... Okay. Let me ... First let me start with a product. Product is, hey, ... it didn't exactly start here. What I started with was I want to type in a sentence and then I wanted to make up a word that represents the meaning of that sentence.

very clear there's like input and output and stuff like that. And, actually, there's a great training set, which is the dictionary. The dictionary is a perfect training set, so it has the definition and then the word that actually fits that definition. And so you might imagine if the- the definition says the word, "Russian" maybe it's a Russiany sounding name or something like that, or French if you have accents and stuff like that.

and I had that in the back of my mind, seemed like a cool product. when I say cool product, not necessarily useful, but just cool. and so I, tried that about ... When I was doing field ranking probably in 2016 or something, I tried, using these ML models called, character RNNs, and it was a miserable failure. I don't know if it was my problem or, it was just like the stuff I'd threw out kind of work. It was like Russiany, but it would repeat characters and be all over the place.

And so it was frustrating. I had this on the bat, it didn't work, I wanted to launch it. And it was just frustrating me that AI wasn't there. And so when I left Instagram, this is like GPT-1, GBT-2 had come out, not GPT-3 yet, but, it was about to come out. And I was like, maybe language technology has improved enough to actually make this work.

And I tried it, tried GPT model, GPT-2 style model, and, with a bunch of hacks to make it somewhat good. And I was like, "Holy shit, this works." and I guess the ... Maybe the more fun part was like, if you didn't have the conditioning, so if you didn't like condition on text, it would just generate the entire thing. So word and definition in and of itself. and I was like, "That's cool."

And so started brainstorming a little bit and was like, "Okay, if maybe I could just make this into a quick demo that generates words that are not real, but sound real, and that have the definition and also have an example." and, ... Yeah. It took a while from the lab to production, but then made it and watched it, got some press. I was excited about that. And now, GPT-3 makes that like a platform you can just generate these kinds of sites as a, [laughs] as a service almost. ... Yeah.

Delian:  So if you were to advise somebody that's maybe, similarly, leaving, let's say, Stanford, or, studying, AI, ML wants to apply that into the world of consumer products and have an impact on the world, maybe now with the more mature perspective than maybe where you were in ... at 28, 12, 13, when you maybe stumbled into it, what would be the path that you'd recommend for somebody?

Is it like, "Look for the place that has, super out-sized users relative to the size of team"? what do you think is the, ideal path for, if you were to advise yourself from eight years ago, maybe existing in the world of 2020 as opposed to 2012?

Thomas:  Yeah. There's- there's a lot there. I think that, if you can find that team, it's ... you definitely want to grab out of that. if you see an exponential growth curve and have the opportunity ... it doesn't even matter what job, what role you're in. I started in dev ops, like literally just writing scripts to auto scale, AWS tiers. you want to just be on that rocket ship? You'll ... It'll be fine. and ... But that's a rare opportunity. Not everyone's going to have that. Also, you need to make a prediction about the future about what's going to be successful. And so that's- that's a challenge.

A couple other things I just mentioned briefly. One is that what I've discovered now ... When I was in AI, like stra- strangely, even in 2012, it was like a backwater field. It was not a thing. and I ... So I had an advantage because I knew this skillset, let's just say a skillset that almost nobody else knew. and right now AI is table stakes. Every software engineer graduating should know, something about machine learning. it doesn't mean you- you need to go into it, but it's- it's a necessary component of any computer science education, just I don't know, operating systems or programming languages or something like that.

the other thing I think is like ... There may be some opportunities that are- are ... Okay. the last thing I'd say is just ... it's probably right now, I think, as a society, we're struggling with the ethics of AI. And I think that's actually an interesting question to take up. And so I think that not neglecting other parts of education, so having a more broaded perspective and actually just considering the questions that exist, and then going to get the experience of Facebook or Google or whatever, company that has, challenges, that's like the sweet spot to be educated and to actually have a very thoughtful opinion on these kinds of subjects.

Delian:  Makes sense. Appreciate, you giving that advice. Thomas, it was really nice having you on the podcast today. great conversation. Thanks so much for coming on.

Thomas:  Yeah. Thanks for having me.

Delian:  Sweet.

Thomas:  All right.

Delian:  Thanks for listening everyone. If you'd like to support the podcast, please sign up for a paid subsidized subscription, which we use to pay for transcripts, mics, and other improvements. If you have any comments or feedback on what kinds of questions I should ask, who should come on the show, or anything else, please do let me know. Have a great rest of your day.