Pod Insights a16z Podcast a16z Podcast | Engineering Intent About this episode "Young hungry and scrappy" is how Hamilton described his country, and it's how many -- including the guests on this episode -- describe startups... or more precisely, the mindset that engineers in startups need to balance both creativity and efficiency. But what happens as those startups scale, accrue technical debt, standardize their frameworks, and hire even more engineers? How do they deliver on their product while also staying on top of -- or better yet, using and also pushing forward -- new tech? (Even if that "new" tech is really the old, much-promised-before-but-finally-here, machine and deep learning?) And how do they do it all without getting mired in philosophical debates? Every Hamilton needs a Washington, after all...
Transcript Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:00 Speaker 00 The content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund. For more details, please see A16Z.com/disclosures.
0:19 Sonal? Hi, everyone. Welcome to the A6Z Podcast. I'm Sonal. Today's episode is about the tech that comes into our lives in unexpected ways. But more specifically, we talk about the engineering that's hidden behind and that drives consumer products that people use every day, like Airbnb, which people use to book homes, vacation rentals and experiences, and Pinterest, which people use to discover ideas or inspiration for things to try, from recipes and home decor to style and more. Joining us to have this conversation, we have Lee Fan, Head of Engineering at Pinterest, and Mike Curtis, VP of Engineering at Airbnb. In this hallway-style conversation, we cover everything from what it takes to manage engineering teams and the myth or the reality of the 10x engineer, to recommendation systems, data,
1:02 Sonal? and machine learning, to the camera as a vector for understanding intent, especially given the challenge of connecting the digital and the physical. So you both uniquely connect to the offline world, online to offline, and then back and forth. It's not just unidirectional. So a question I have for you is, how do you know what people are actually doing in the physical world? Isn't the biggest challenge that they essentially lose data when the person acts physically? You don't have any freaking clue what they're doing. So how do you solve that problem technically?
1:31 Speaker 02 I think it is an interesting challenge. You know, the experience that you have offline is like your true engagement with the product, right?
1:36 Sonal? Yeah.
1:36 Speaker 02 Airbnb built really early on this review system, right? An ability to be able to say like, what kind of experience did you have and what happened? And one of the areas that we're experimenting with now is being able to use the reviews that were given on a place as a predictive signal in the matching model all the way back. So we're able to fit-
1:52 Sonal? And why does a matching model matter?
1:54 Speaker 02 Meaning the prediction for like, if you book this listing, how likely are you to give it a good review? So we can actually sort of collect some amount of offline data from the actual experience that you want to have.
2:03 Sonal? Right.
2:03 Speaker 02 And then use that as a feedback mechanism into our ranking model in terms of what listings we're going to show you next time. Because again, if like your objective function for ranking is really like, how good of an experience are you going to have? Then all this data you can get about what kind of experience did you actually have out there in the world can be then used to your next booking, but also to like, be able to look at using that data towards how other people are going to book.
2:25 Sonal? Those reviews are not just reviews, they're much more, they're a vector to more data for you to be able to do different things.
2:30 Speaker 03 In some way, we have a unique advantage that user, when they use Pinterest, a lot of them start with explicit action. They ping something they like. Those explicit signal user give us is a unique asset. And this is fascinating data set our engineer can work with so that we know one out of 10 times, they're likely to go down deeper and a click by click. And then sometimes when I would say 20, they will buy something.
2:56 Sonal? You know, you guys are both talking about something so fascinating, which is the complexities, not just of search and matching and relevance, but that a lot of people don't explicitly express what their intent. I'll have a Georgia Nelson desk and a vintage French country bench. And those two categories do not go together. If you were to do an explicit search, one would do mid-century modern and a different category for country or cottage chic. But now you have to infer this collection, this, this cluster of traits of what people are interested in. Can you guys talk to me about some of the challenges of doing this?
3:30 Speaker 03 Um, first difficulty or challenges, don't know what they don't know, right? For me, like for example, very hard to describe. It is this contemporary, classic, traditional. And also you don't know whether this is ultimate thing you want to match upon. Maybe they, you should put in the two different rooms, right? So I think one powerful thing. We help users to do is chances are, some users have the same struggle with. And the hard part is maybe that user is in Scotland and that they use different language to express. That's why lens can be powerful because you don't know how to describe. Instead of asking user to input a text as a query, take your camera, point to whatever you want to understand,
4:14 Speaker 03 and we will give you ideas, what are the related ideas.
4:18 Sonal? So you're actually saying that people, because people don't actually always have the language, even when they do, there's different languages.
4:23 Speaker 03 There's a different language.
4:24 Sonal? A camera and image can actually level that.
4:26 Speaker 03 Exactly. And also, you think you are looking for contemporary. Speaking, you're one element contemporary, you take a picture. We know, okay, maybe you like those styles.
4:35 Sonal? You think you know you're looking for this.
4:37 Speaker 03 Exactly. It's not necessary you have to describe in the English text.
4:41 Sonal? Right.
4:42 Speaker 03 You give me the image, I will tell you like thousands of other users match your cabinet with this one or match this sofa with the other table or something like that. You may not think your white sofa match is a red table, like outrageous, but when you see the image, you say, wow, that looks good. We want to give you that inspiration and in some way, this is different from Google. There's no right or wrong.
5:03 Sonal? This is not the classic Google rabbit hole of you click on a bunch of page because you're interested in one topic. For me, it's fascinating because it's actually understanding yourself in a weird way. Computers augmenting humans. I'll give you a concrete kind of dumb example, but this literally happened to me on Pinterest in the early days. I had no idea, I love the combination of dark green, black and gray. So whenever I did searches on shopping sites, I would always put my favorite colors, but I never knew I love that combo. And then I noticed one day that one of my Pinterest boards for dresses was all dark green, gray and black. And I was like, oh my God, I love this combo. And in a weird way, the system kind of taught me what I like.
5:37 Speaker 02 I love that idea and like the, the sort of act of like helping inspire somebody for what they might like that they don't necessarily know that they're looking for, I think is really fascinating. And it's, it's, you know, it's actually kind of analogous to like the, the challenge that we have to solve on Airbnb because you think.
5:51 Sonal? How so?
5:53 Speaker 02 Well, I mean, you've got millions of travelers who are traveling. You've got millions of homes that they could potentially stand. Every home is completely unique, right? Like totally different, like we're not selling like a block of hotel rooms, right?
6:03 Sonal? There's so much variety in the style even.
6:05 Speaker 02 Every single one is unique. And so, you know, a similar problem for us is like, how can we look? Through like the click behavior of somebody who's going around and looking to stay in a certain area, if you're thinking about staying in Paris next weekend, you might have to sort through like 50,000 different places. Like, obviously, we can narrow some of that down with filtering. But then we can also look at what are the similarity and characteristics between the listings that you're clicking on or the ones that you're expressing interested in? And then the deeper you go into it, the more we can be re-ranking and surfacing other things that share some of those attributes, some of the areas that we're exploring now is like, how can we find other embeddings in those images that can take the unstructured data of the image, turn it into something that can actually be tagged and labeled and then used in that ranking algorithm? Like, maybe you particularly like places where you have views of trees from the
6:50 Speaker 02 window and all the listings happen to have those attributes.
6:53 Sonal? Yeah, actually, I do like that. I want that.
6:55 Speaker 02 I like that, too.
6:55 Sonal? Tell me how you technically solve that challenge of moving from unstructured to structured data in that case. How do you now extract that data, feed it in?
7:02 Speaker 02 In the same way, like, you know, you can hold up your camera and it'll detect your face, right? There's-
7:07 Sonal? Like on Snapchat and Instagram and everything with filters.
7:10 Speaker 02 That kind of technology can be used to detect all kinds of different objects that could potentially be correlated with what would be interesting for you. And so it's that same kind of computer vision techniques that can be brought in to bring like that unstructured data forward and turn it into something you can actually use.
7:23 Sonal? And what does it take, though, to like move it? What's the pipeline?
7:26 Speaker 03 First is you start with data. And I think both company have massive interest in data that user explicitly shared their interest, right? You click or you book. The first step is really understand that data. We have computer vision technology to identify, there's a sofa, there's a table, and we can train, label those, all those like millions, actually billions of things. We will have a subset of labeling data, say, this is a white sofa, and the computer will pick up all those labeling and start to learn, this is a white sofa, this is a leather sofa, this is a table with a brown color, things like that. Once they learn, you will then connect your user data and say, okay, this user like to click, for example, the open loft feeling of house. And these are the things connection, now you learn the connection, you say-
8:10 Speaker 03 We can do personalization for you. Now you go to say, search the same query, for example, red shoe. We know if you are male, you are likely to wanna have maybe a running shoe instead of a high heel pump for women. So those are the things we have to solve for personalize your favorite thing. And all those are powered by massive data set. Understanding of the image, understanding connection between image and a user, and use machine learning technology to make those connection.
8:38 Sonal? Okay, so you have, let's say you have users that are clicking what they're interested in. They're browsing, you know, room listings, they're browsing pins. How do you distinguish between the aspirational and actual outcomes? With both, you can have people who are pinning things because they wanna have their dream house one day. Or this is a type of Airbnb you're staying in because you can't afford to live in a beautiful industrial loft. But hey, when you go away, you can at least do it for a day or two, and that's a great way to get this experience. Do you actually weigh what they booked or actually pinned and bought more highly than the things that they might just be browsing and clicking through? Because one of the tricky things is differentiating intent when it's aspirational versus actual. And sometimes you don't know, is get an outcome that you can link to the choice that they've made. Because you might not know until weeks later that they've bought that dress after all.
9:23 Speaker 03 Oh, yeah, exactly have the same problem. I would say it depends on user's state of mind. Sometimes users are ready to book or purchase. Sometimes they're just explore, right? I plan my vacation months ahead and I start to explore. At that time, what I need is a creative inspirational ideas, what's possibilities.
9:42 Sonal? That's how the mind works, like things sit in the back of your head for a while.
9:44 Speaker 03 At that time, I, I think what we could do as a product, expand your horizon and help you to discover new interesting ideas. We don't have to push you to into deep, say, purchase this or book this. But then later, we can tell users' behavior, they're going narrow and narrow. Now we, I have this living rule, I do plan, looks like you are going down and down. Where is this sofa? I know I wanna get leather, I wanna get the, have the recliner, things like that. Now we see those signals, we know, okay, you probably have the right intent. You're ready to do something more real. Then we will help you to drill down. All those need a computer to understand the user intent and which state you are. Not everyone is, is ready to purchase right away.
10:29 Sonal? You know, similarly, when you think about differentiating between aspiration and, and actual intent to do something. There's also another dimension of like feelings. When you think about what in the early days of Facebook, you could only like or not like something. And now you can express a variety of things. We have emojis, a new communication form. And so we're essentially making emotions more machine readable. And that's a really useful thing, but you don't get those signals, actually. Because you're talking about people both in the physical world. And maybe, I don't know if you guys actually have the range of emoji, not just a star for how you liked or not liked something, or pinned or not pinned something. How do you guys think about that dimension?
11:02 Speaker 02 No, I think, I think there's actually a lot of facets to this. Could be like, you know, I'm looking for something that's a little bit more secluded and out of the way. Or I'm actually looking for something that's like in the middle of the nightlife. And like, we can start picking up on those signals again, based on like how you're searching and browsing through it. And then, so that's sort of at the front end. But then, you know, all the way at the other end, we can look at, again, with that review information, there's some of it that is structured, right? Like, give a star rating. But then there's also the content of the review itself. So we can do sentiment analysis and, you know, some natural language processing on that to sort of suss out which aspects of this, like what feelings did it evoke? You might have given it a five on cleanliness, but maybe you felt like, oh, it wasn't really like the right neighborhood.
11:43 Sonal? It lacked energy or didn't really have a warmth to it. Like, those are words that are interesting.
11:48 Speaker 02 People have such emotional connection with these experiences that they have traveling. And I, and I think like, as a result of that, like the reviews that people write and the comments that they have and everything are very rich and like filled with emotion about these experiences that they had.
12:01 Sonal? Don't you guys have this problem, which is typical of all review behavior, I think, where you have this sampling from the extremes? You have this natural skew where only the people who are most extremely motivated because they loved it so much, or extremely frustrated, like, God, I hated that place. You regress to the mean when you sample from the extremes. How do you think about that?
12:21 Speaker 02 You know, there's, there's such a personal connection when you're staying in somebody's home that I think there's, there's sort of almost like a social contract a little bit. The percentage of trips that get, get reviewed are incredibly high.
12:31 Sonal? By the way, you might have a unique. Benefit here too, which is that you have the both sides of your marketplace. Host can also be supply and supply can be demand in the sense that people who are hosts are also guests and people who are guests can also be hosts.
12:43 Speaker 02 We get a high review rate from the host side as well. And I think that's also part of being a host and hosting, right? One thing we were worried about a while ago was that the reviews were really nice. You know, one of the reasons for that, of course, could be like fear of retribution, right? Like if I reviewed you, the host, and said like, oh yeah, this place wasn't that good. Then you say, well, this guest wasn't great either.
13:01 Sonal? How'd you address that?
13:02 Speaker 02 So, so one of the things that we did was we changed it so that we have a simultaneous reveal of the reviews. So both the host and the guest have two weeks to write a review and then they're revealed at the same time. So that way you don't have to worry about like the retribution.
13:13 Sonal? They can review asynchronously, but the review is synchronous. So therefore it's they don't know until they're they know.
13:18 Speaker 02 Right, exactly. And the reason that we went down that path again is also to make sure that we have good accuracy in the data. Cause there's a lot of information in that review that can come through and be useful signal to us later on.
13:27 Speaker 03 I would say Pinterest is unique in that way. Cause when we design, we really design an app or a product for user to feel positive.
13:35 Sonal? Yeah, inspiration should be a positive thing. It's not necessarily a I hated it thing.
13:39 Speaker 03 Use Pinterest, the main amazing thing, I figure that's every paint you see like a so beautiful looking, right? And this is something actually surprising enough, we are training computer to learn. Why this image of the same living room? You, you take a picture of this way and that way, they look so different. One just looks so inspirational, the other like maybe just boring. And the computer will start to, to learn those cues. And they already learn certain colors, say yellow is more, you know, inspirational, a happy than other.
14:09 Sonal? This reminds me of the old days of like color coding moods and color coding design. But it's actually amazing because you can now actually do it in a way that's effective.
14:16 Speaker 03 Exactly. You, you can actually model. So what are the images? It's more inspirational than the other. Of course, it's not science, it's actually art. We learned, for example, if a same dress in a, in a stock, very boring white background. Didn't sell. We also want to add a personalization signal. We have some experiment in-house, like if you you take a picture of yourself and we learn your skin tone. But by the way, we can also learn from the pins you like, and we know there's certain style, certain shape of a model you like to see.
14:48 Sonal? And usually it's people that are it's similar.
14:50 Speaker 03 That you've related to. So then then we will know those are type of dress or something you look for. Right? It's not super fancy. For example, for me, it's more kind of business-like that I can wear. So I I think all those, you know, signals we can pick up. And we also, I think one luxury we have is a user super engaged with our platform. So sometimes they're willing to give more signals. Take a picture of your dish, take a picture of your living room, take a picture of your your uh kids artwork.
15:18 Sonal? Just you think about the future where we're going with this connection between digital into physical. There is a world of sensorification happening all around us, whereas more and more sensors are embedded in our environment. You don't only have to rely on your smartphone to then do that. That example of colors is fascinating because you can see colors, you can see sofas, white leather, the style, this type of thing. One of the funniest memes that makes its way on the internet, and I love this meme, and it's actually a fun new genre, is can computers tell the difference between a dog and a bagel? Or the latest one I saw was like Kentucky Fried Chicken and like these little like poodles that have the same kind of texture. And a human can instantly spot the difference. And this is one of the challenges in deep learning, because colors are easy.
15:57 Speaker 03 Mm hmm.
15:57 Sonal? But some of this is not.
15:58 Speaker 03 I would say if you give a classic picture of dog versus classic picture of a bagel, the computer can solve at this point.
16:07 Sonal? Right.
16:07 Speaker 03 However, if user generated random pictures with a dark background or with the dog in a weird shape or whatever, it is hard. Remember, computer learns things from the data we train them. Right? If you have millions of pictures of a dog and with different shape, a different color, I'm sure it will get there. Right now, I would say a lot of domain is limited by the data. If you only have a limited data to teach, let's say fashion, how can we know this is a fashion that are high end and more for the runway instead of a daily? It's a lot of data because it's a subtleness. It's very subjective, everyone view differently. And then to have the computer really get there, it would take a while.
16:52 Speaker 03 I think the breakthrough recently in deep learning is that we can learn more sophisticated way. But it's a fascinating move so fast.
16:59 Speaker 02 Yeah, I mean, at this particular point in time and like on the spectrum of point of time, like, you know, a year from now, two years from now, or five years from now. If you look at like the advances that have been happening, just even really in the last couple of years in deep learning and AI technology, it is like an explosion over the 20 years before that. It's like these two huge forces sort of coming together and reaching a tipping point. The availability of massive, massive data sets. And then the processing power to be able to actually train models on them in a time reasonable way, like the broad availability of GPUs. So I think what we're seeing is that we're at this moment of, you know, potential like exponential growth in this field that has been promised multiple times before in history, only to kind of not really have the ingredients there. The applications of this type of AI technology can go to every facet of how we live and work
17:44 Speaker 02 and, and you know, how people sort of exist in the world. And what technology companies are really going to be great in the futures are the ones that think about this as this needs to be core to what we do. And we're not just utilizing this technology, but we're figuring out ways to push it forward.
17:57 Sonal? Natively.
17:58 Speaker 02 Because we're going to be have to, we're all going to have to be in leading positions in this technology in order to be competitive in the, in the world that's coming in short order.
18:05 Sonal? I think that's so true, and I agree with you. And let's shift gears and talk about that. So you guys essentially are startups. You're still a startup technically, which is really weird to think about, given how big you are. And one of the biggest challenges is obviously getting to scale. And when you have that scale, you grow very fast. Who do you think about balancing building this kind of competency as heads of engineering? When in a lot of traditional engineering jobs, I would imagine that you kind of know your roadmap already. And here in this kind of business, you shift direction. How do you think about building this organizationally and operationally? How do you hire people?
18:36 Speaker 02 There's such high demand for people who have skills around, you know, computer science, data, you know, AI, machine learning. The technical challenge is so important and that has to be there. It's like the foundation for, as an engineer, am I going to be fulfilled by my work? But when you can satisfy that at multiple places, because you know, we all have interesting technology challenges to solve. Then it starts coming down to like, what is the purpose behind my work? People more and more are choosing their work based on the mission behind the work and the purpose of that work as opposed to like the specific thing that they're going to work on.
19:08 Sonal? Do you guys ever face like, you know, there's a lot of these funny religious debates with programming languages or like how you build your tech stack?
19:14 Speaker 03 Like open source versus like a non-open source and an own stack and a shared stack or non-shared stack?
19:19 Speaker 02 You know, open source can be kind of a lightning rod on some of these things.
19:22 Sonal? Yes.
19:22 Speaker 02 But I think we've, we've basically just taken a position on it and said like, you know, this is, this is the way we're going to manage it. We are strong believers in open source. We regularly open source our own internal technologies, particularly around data. Like everything we do from managing data pipelines to how we do data analysis, we pushed out to open source. And then internally, we kind of have the philosophy that if there's a great open source tool that can solve this problem that we're trying to solve, let's look to use it before we look to build our own. Because we want, we want a higher percentage of the hours, like engineering hours and thought and creativity of work that happens here being focused on things that are very unique to our business.
19:57 Speaker 03 It's a progress story, right? Because I think in my early days of career, like almost all the big internet companies do their own stack. And then we have tremendous progress in the open source community. Then you can say that principle without hurting your business efficiency. A lot of engineers, I would say, believe they can build better. And then it's a judgment call to, to encourage that or say, let's look carefully. I totally agree when there is an open source tool that you can leverage, let's pay attention and consider it. Sometimes I will smile when this question come up because I feel like engineer, by nature, they are so practical, but sometimes. They go to the extreme of philosophical discussion.
20:37 Sonal? Engineers are secretly poets.
20:38 Speaker 03 Yeah, and I'm like, this, let's talk about this case, right? Instead of say, are you in this or not?
20:44 Sonal? You're saying that instead of letting it be an abstract fight, you just go right down to the concrete of what you're actually working on.
20:49 Speaker 03 My strategy, when you go down the concrete things, and I always tell shared context, share the information you have. What most likely happen is when you shared exactly same information, your conclusion will be the same. But it takes a lot to force them to spend the time to communicate to each other and to spend time to understand, you know, that there's not so big a difference that you expected, so.
21:13 Speaker 02 We did a couple of things relatively early on to try to head some of that off. The first is that we standardized our supported technical stacks and said like, we're just going to operate within these stacks and not deviate from them except for in exceptional cases. And we also got, you know, actually a pretty broad engineering effort a couple of years ago to come up with, from an infrastructure standpoint, like what are the tenants by which we want to do development? And they're basically just guidelines. And the idea is if you're an engineer, if you're working within the architectural tenants for how we how we develop software, then you're pretty much good to go. And like, just stay within it. If you want to go outside, here's the group of people that you can talk to. And I think that that has like, sort of calmed down a fair amount of that chatter. Don't get me wrong.
21:53 Sonal? You're like, I fight this battle too.
21:54 Speaker 03 Yeah, I I think that's like definitely a good, a good principle. It's just the constant debate you probably agreed is really to encourage them to understand the benefit of a consistency and the company principle versus their own creativity, right? I think the engineer always want there like, I have my own idea.
22:14 Sonal? Individualistically by definition.
22:15 Speaker 03 Exactly, and you just have to constant balance.
22:18 Sonal? I want to ask you guys a question about technical debt. Big topic. You've had to grow very quickly and build things very fast and it means you leave this legacy of crap behind, frankly. And I want to know how you guys think about that.
22:27 Speaker 03 I strongly believe that we, you know, in some way, this is engineering excellence, right? You have to build a Assistant or infrastructure that lasts a long time, right? Because it's ultimate efficiency. However, there will be a lot of product features we iterate and we don't know whether it will stick for three months. Right? And maybe this features user don't want. Maybe something we will admit like shit, it's a failure, right? So it's okay. So we want to be, I use Hamilton quote a lot like recently.
22:56 Sonal? Oh, I love Hamilton.
22:58 Speaker 03 You know, want to be, you know, young, scrappy, hungry, just move fast. Um, at the same time, this is why we need a senior engineer leaders in house say, look, this is the system. We plan like I say, machine learning framework for the company. We plan to use for two, three years. Let's design carefully.
23:14 Sonal? Basically saying like you need a little Washington to balance out the Hamilton.
23:16 Speaker 03 It is definitely true, I would say. And also historically you accumulate all those that and let's face it. And I don't pretend it's there and not face it. We set aside time, say fix it week.
23:29 Sonal? I guess it's the only way you'll actually make things happen is if you carve out the time.
23:32 Speaker 03 Allocate, carve out time because otherwise you keep saying the principle. The engineer leads just want to shipping new things. This is human nature, right? You want to have fascinating new things coming out. And look at that's like, oh, system like, but you need them. You don't want that the reliability will hurt you in the future.
23:48 Speaker 02 A great engineering leader finds the right balance like between time spent fixing things that that came from before and doing forward, forward development. Because neither end of that spectrum is correct. Like if you spend all of your time having like a technically perfect system, you probably didn't do anything to further, further your business and vice versa. So I think like there, there definitely has to be, um, has to be a spectrum of time. The way we've been thinking about it and the way that we try to do it now is that the business leaders who are responsible for furthering the business also have goals that are associated with bugs, performance, like all the things.
24:17 Sonal? You're calling the engineers, business leaders.
24:19 Speaker 02 Yeah.
24:19 Sonal? That's actually kind of counterintuitive for me frankly, because I didn't even think about it that way.
24:22 Speaker 02 Because an engineering leader should be an owner of the business, right? Thinking about like balancing these types of trade offs. And the other angle that we think about technical debt on is really about what is our overall capacity as a technical organization to, to move? Like how much can we produce in any, any given amount of time? And you have to be watching that because if technical debt crops up, what happens is your overall throughput goes down. And the worst thing that you can do at that point is say like, well, we'll solve it by hiring more people, right?
24:48 Sonal? That's what people do.
24:49 Speaker 02 Yeah. And like you, you can't hire your way out of that. Every, every person that you hire just ends up becoming less and less productive. And so you have to be able to look at like, what is your contribution per engineer or per engineering hour work? And seeing that that is like increasing over time through retirement of technical debt, because then the organization has more throughput as a whole.
25:06 Sonal? One of our general partners, Martin Passato, wrote a post on why you have to hire VP of Engineering early on. And the big point he makes, because frankly, no offense to you guys, not knowing this world, I thought of it as roadmaps and deadlines and keeping track of schedules. And you don't realize there's just a big trade off between the big picture versus individual pieces, even if they're empowered owners. By definition, a lot of individual people or groups don't own that big picture. And so how do you sort of connect those dots in this way and including making the decisions about priorities and trade offs? Because he gave this great example of how a way to accrue technical debt is that someone will actually go to an individual engineer and say, how long would it take to build out this feature for this customer? And of course, that, that person will give them an estimate and say, well, I think it'd take two weeks, not realizing that actually that's an estimate in isolation, not trading off the bigger picture.
25:48 Speaker 03 And then you need to take into account who this engineer is.
25:52 Sonal? What do you mean by that?
25:53 Speaker 03 Meaning that engineer versus engineer, the throughput can be 10x or even 100x difference.
25:58 Sonal? Do you guys really believe in this? I mean, is it a myth of the 10x engineer?
26:01 Speaker 03 I believe in, I also believe that we should reward according to.
26:04 Sonal? Oh, interesting. So really reward that throughput. I love that idea. I want to be a 10x editor.
26:09 Speaker 03 But it is a judgment call. I think the leader to decide like whether there is a 10x or 2x difference. This is the beauty of why this field is so fascinating. A good engineer, the amount of the throughput or the contribution they can bring on table, sometimes it's beyond what you can measure or imagine.
26:27 Speaker 02 My thinking on that is slightly different in that I think that, you know, people have periods of time where they can be like producing 10x what their, their peers might be. And it's that magical moment when they're aligned with the right project, with the right skillset and like the right personal energy around it.
26:42 Sonal? I love that you said that throughput as this connotation from the semiconductor industry. But it's actually bigger than efficiency because it's about the social, cultural context that actually supports this idea. And I love it because it reminds me of creativity. You have moments where you can do like 10 edits a day and you have moments when you can maybe do one because you have a lot of meetings.
26:59 Speaker 02 And that's, yeah, that's exactly how engineers and like everybody works, right? Like these moments where it comes together.
27:05 Speaker 03 And a motivator or not to make a difference.
27:07 Speaker 02 This is like one of the most core fundamental premises behind engineering management is like understanding the person, what motivates them, what are their skills, what are they trying to develop? That magic moment that I talked about, like being able to find those moments for people and that being a core part of the responsibility that a manager has because everything good happens then.
27:25 Sonal? I have to ask you guys this, you know, you both have founders and co-founders who are not necessarily tech quote native, came from design product, but not classic tech. Is that sometimes a thing that you have to explain? Like, what's that like?
27:37 Speaker 02 I really enjoy it because Brian brings this incredible design sense. And it's not just to like how the pixels look on the screen, it's to like, how are things going to work? How does this ultimately pan out down the road? And I actually think that that in combination with a technical discussion can be super powerful.
27:50 Speaker 03 I would echo that. I actually really enjoy because they fascinated about the challenges and are very supportive. When you have everyday conversation with tech stack, like a people and you have formed a certain way because like implied this certain knowledge when you talk to a designer, ask a question like, wow, I never thought about this. And the process of thinking about how to answer this and then discover there's some hole in my logic. Or discover like, maybe we were too, you know, go down this path and maybe we should think in some way encourage you to think out of box.
28:22 Sonal? I think this is a wave of the future, in fact, because there's been this traditional adversarial portrayal. And this idea that like engineers don't collaborate well with designers and designers don't always collaborate with engineers. And I think that's changed a lot in the last 10 years as you have a lot of these products, examples we're talking about today, that they're touching lies, but there's deep, deep tech behind them as well. Well, thank you guys for joining the A6 and Z podcast.
28:41 Speaker 02 Thanks for having us.
28:42 Speaker 03 Thank you.
Previous Company Building in Crypto: A Conversation between Brian Armstrong and Chris Dixon Next What's Next for Education Startups Page updated October 9, 2026 · transcript and summary are machine-generated