Every year of your career is not equivalent. This is one of the crazy ones. I guarantee you can just do things. It turns out you can just do things and it's fine. I used to wake up every four hours for two years and read an email report that I had this script send to me of what the top activity in the site was across every category. And if I saw a scraper go in the database, kick them off the network. So everything was manual, everything was hard. I think the major delta between even two years ago today is our belief that the software will arrive. Machine learning was the breakthrough and robotics was late to it. Robotics really only opened their eyes to modern machine learning like two or three years ago. In Karate Kid, he's like painting a fence and then he's like waxing a car and it turns out he's learning karate. Like I spent a lot of time painting fences and waxing cars to be the person who got to kind of be in that job. So my job is it's intense, it's stressful, it's competitive, it's fun.
Everyone, let's go ahead and get started. I'm not going to spend a ton of time on introductions, but it is a distinct pleasure to invite a dear friend, former colleague, an amazing father, but also the Meta CTO Andrew Bosworth to the stage here. Welcome.
All right, Bos. I've known Bos for 20 years, so we'll keep this light. We'll keep this fun. Hopefully a little spicy. Bos, all of us have a story about joining Facebook. My story is you interviewing me. Your question was implement a queue with a stack and then what's the runtime? I remember it vividly. I actually didn't do great on that interview if I'm being honest with you, but I did better than the other ones. But no, I mean, how did you find out about Facebook in the early days? I know you TAed Mark. But there was it was you, Fed Jang and Bob Rhan I think who were all at Microsoft. But tell us a little bit more about super early days.
Yeah, I joined the product Facebook on day two as a Harvard undergraduate. ID 1,681, which for those that know means I was user 1,678. It's the first three chest accounts. So that's my user ID for the product. So I used the product right away. And Mark had been, just a few weeks earlier, had finished the final in a class introduction to artificial intelligence, CS182, where he was in my sections. I was a teaching fellow section and I had had to vouch for him to the ad board after FaceMash. I didn't know this. Yeah, FaceMash, the ad board came to my professor and the professor David Parks came to me and was like, 'Hey, is this guy okay?' He's like, 'Yeah, he's okay.' All right, good. And so he was not really focused on my class. I'll say that he was building Facebook while he was in the class. Well, I didn't know this part of the story. Yeah. And so time went on and I had gone to Microsoft. It was a really, I don't know if we graduated, understand 2004 was the class of kids who had just seen the bubble burst when they were freshmen in 2000. So it was a very, it was an apogee. It was a small number of computer science graduates. Maybe 40 or 50 of us. And I think more than half of us went to Microsoft. Microsoft was just like cleaning up university recruiting in computer science at the time. And so I got a, I was at my desk at Microsoft and I got a message on AIM, AOL instant messenger, from a person being like, 'Hey, come, you want to come work for Facebook?' And I was like, 'I just got this job at Microsoft, I bought a house, I think I'm good.' She's like, 'You can buy 10 houses in the Bay Area with the money that we'll get from Facebook.' And I was like, 'Well, now I know you're lying because I know what 10 houses in the Bay Area cost even back then. You can't be doing that. So, this is some bullshit.' But wait, this isn't Dana Hayes, is it? Dana Hayes. Absolutely. Dana Hayes. Very good. But my family's in the Bay. So, I'm like, 'But I'll take a free ticket.' So, I fly down with Dave Federman. We stay at my parents' house. We interview. And we were blown away. We expected it to be a total joke. We actually used to have a lunch called 'Sweet Dudes You Love the Party Lunch' (SDWLP). Those were early days. And we were all making fun of Facebook, 'You go join Facebook, whatever.' And then me and my buddy went and we were like, 'Oh, we got to join this company, this is legit.' You know, Mark had such an incredible vision. I met with Jeff Rothschild. I got interviewed by Adam D'Angelo and Diddy. Do we call him Diddy anymore? Is that just me? Well, sorry, I think it was not cool for the last four or five months, but I think we're back now. We're back now. He's been exonerated more or less. Still a bad dude, but okay. Yeah. So, we were so blown away. So, we went back to Seattle and we went directly from the airport to a party that was well underway. Everyone was smashed and everyone was like, 'Oh, how was Facebook?' And we're like, 'No, it's amazing. We're going to join.' And literally Charlie Chver, who's another part of our crew, on his Blackberry, drunk as a skunk, starts typing a message to the recruiter like, 'I want to join.' And has us like proofread it because we are not yet drunk and then sends it. So within a few months we had seven of us who all went down together to join January 9th, 2006. Incredible.
It's interesting. One of the things I often get asked about even today is like, what did you see in Mark in the early days? And I think I tell them that two things are as obvious about him today within 10 minutes of meeting him as it was when he was 19. He is incredibly smart, his clock speed is just four standard deviations away from most people, and he's just incredibly intense. When you meet him, you kind of have this like he will shape the world around him in a vision of what he sees. And he's clearly doing it today still. I mean, I think the founder mode is bad for him. Yeah, maybe. He's honestly, you really keep thinking this guy cannot possibly get more intense, and he does. He does.
Well, if you already kind of mentioned this a little bit, but if you had to pick three words to describe the early Facebook culture, what comes to mind?
It was chaotic. Super naive if I'm being totally honest. Which I think is a good thing. It's a feature, not a bug. That's right. It was fun. I think it's important. It was fun. Chaotic is definitely the number one word. It's definitely the tough word.
Have we ever heard the story of Richie in the internet? I think we'll let Richie say that story, but you should act. No, go for it. This is how chaotic it was. At one point, the site goes down, but we can't fix it because our internal network is also down. And we're like, are we under attack? Is it DOS? What's going on? And it's like hours and this is bad. We don't have site downtime. We did not tolerate that. So, it was a big deal. Everyone's freaking out. And then finally, like one of our IT guys, I think Ton Air, starts walking around the wall. What's he doing? He finds a cable and sure enough, not her fault. Ruchcci had been trying to plug her laptop into the Ethernet and so she found like the extra edge of the Ethernet cable and what she did is she just plugged it into the wall, which is fine. Unfortunately, what she didn't do is it wasn't the right cable. So, she had plugged the Ethernet into itself and not her fault, due to a misconfigured server, we ping flooded our own network and then due to a misconfigured firewall, it bled over to the production network. And so, we still tell the story of this time Ruchi plugged the internet into itself and took the whole site down. I mean, this is so chaotic. Just imagine you're just working and then suddenly it's like a fire drill and everyone's freaking out and then a guy's walking along the walls like unplugging every Ethernet cable and that fixes the internet. It's an unbelievable accomplishment. So chaotic. At SPC, it's Finn and I who do the IT work, not Richie. It's not really her fault. Again, these are misconfigured things, but still hilarious.
But you know, it's interesting. There's one thing that you said there that I think is something that I often think about for startups, which is that startups are so hard for so many reasons. It's such a crazy roller coaster of ups and downs, but it should also feel like fun. No matter how hard it gets, there should be some intrinsic source of camaraderie, of fun, of a sense of optimism about what you're building. And if you're not having at least a small amount of fun, then something's off. And I tell people this all the time, that no matter how hard it gets, it should never feel completely soulless or joyless.
Yeah. It's hard to describe. Fun is one of the most ineffable qualities. You really can't make something fun. That's one of the worst attempts to make things fun, almost certainly backfires and makes things less fun. So, it's some quality of the people and the belief in the thing. We all felt like we were riding lightning. You just knew the thing you were on was rare, was valuable, was exciting. You felt humbled and honored to be there. Again, you were naive, otherwise you wouldn't have been there at all. You'd have thought you were possibly good enough to be there. It was electric.
Yeah. The word that I often use when people ask me what early Facebook was like, the first word that comes to mind is, 'It's the most fun I've ever had in my life.' It was actually just super fun.
It's the most fun you can have without sleeping. Yeah, exactly.
A lot of what we try to cover in these talks is a person's kind of minus one journey. And it's often easier when someone hasn't spent 20 plus years at the same company. But you have reinvented yourself five or six times within Facebook. So maybe tell us about a minus one moment where you came in as the boss who had gone to Harvard, spent a little bit of time at Microsoft, but in some ways had to go through some period of rediscovery or refocus. What is that for you? Newsfeed in the early days? What comes to mind?
Constant. No, you're absolutely right. I mean, there was so much. I worked on newsfeed for the first 18 months I was at the company, but I also did a bunch of other stuff. I did early integrity work. I did a bunch of early stuff because everyone kind of did a lot of jobs. Then I literally changed jobs every six months. I always kind of liken it to a reference that most of you will not get, which is regrettable: Karate Kid. In Karate Kid, he's like painting a fence and then he's like waxing a car and it turns out he's learning karate. That's how I often think about my career. I spent a lot of time painting fences and waxing cars, and then later on I got a job and I was like, 'Oh, I know how to do all this stuff. This is cool. I've done all these things before.' Newsfeed was a good example. The idea was simple enough. We know the behavior is everyone goes to Facebook, there's no feed yet, and they go profile, profile, profile, profile, and they see what's changed. What if we just take all those changes and put them in one place? Simple enough concept. And we wanted it to be ranked like a personal newspaper. Simple enough concept. No one's ever done it before. There's no book, no website, no article. It's just you doing a thing. I still have this notebook. I would just draw these curves. How interesting do I think an event story is? Well, I think an event story is super interesting right before it and right after it and then not very interesting. So that's a pretty easy curve. What about a photo album? Photo albums are only kind of interesting for maybe the day or two after the thing happened. So I just started running these curves. For a long time before machine learning finally caught up with us, it was just random arc tangents in the codebase of me being like, 'What mathematical function matches the curve I just drew in this notebook?' So random, so unscientific. And Mark would call me and be like, 'Hey, there's too many basketball stories.' I'm like, 'All right, turn basketball down. We're going to take that parameter, turn that parameter down.' It was just manual. But it was a tremendous learning opportunity. And again, the naivete we had. One of the greatest sayings last time is 'You can just do things.' It turns out you can just do things and it's fine. So that was the first one for me, definitely newsfeed. That was crazy. And then once I had newsfeed, I was cocky. I was like, 'Cool, integrity. People are spamming the site, people are scraping the site. I can do that.' So we started building these tools for anti-spamming, anti-scraping. I used to wake up every four hours. For two years, I did not sleep more than four hours at a time. I would wake up every four hours and read an email report that I had this script send to me of what the top activity in the site was across every category. And if I saw a scraper or a spammer, I would get up and go in the database, get rid of all this, kick them off the network. This was wildly unscalable. We didn't have people to do this for me, and we didn't have the systems and tools to do this automatically with high confidence. So I was just me as a human waking up every four hours and doing it. So everything was manual. Everything was hard. Everything was yourself. But it was cool because you were standing alone. Then as the company scaled, I didn't scale with the company. I didn't know how to work with people. I didn't know how to work through a team. I didn't know how to be at all polite. So I had a lot of problems. So you're like, 'Okay, being zero to one on the technology is not enough. You have to be zero to one in a lot of different vectors. You have to take other people with you on this journey. What's that like? And how do you lead through them?' So at every different turn, every different job I took, I was just building another piece of the ability. And now I kind of have a job which is professionally negative one to zero or zero to one.
A couple of maybe things to add to the anecdote there. It's interesting. So Bos kind of obviously came up with these hand-tuned weights for our feed story composition in newsfeed. And then Mark at some point asked me to go build out a machine learning team to go automatically tune this. Shouldn't we have some machine learning experts who figure out the coefficients? And we spent, I remember like Max Gubin and the whole teams, we spent 18 months on this and we couldn't beat the metrics that Bos had hand-tuned. So it's a useful learning lesson for me that humans are actually pretty strong with their intuition. Machine learning isn't going to solve everything. Or maybe it was that Google search only switched over to fully machine learned like a couple years ago. For a long time, the machine learning would beat the algorithm and then the hand-tuned team would go in, figure out what the machine learning algorithm had learned, take it back to the hand-tune, and they beat it again for another like six months. We underestimate the degree you can make a lot of progress with human intuition. And the second thing is something you said earlier which I often think about in the context of startups: you cannot tell someone to wake up every four hours to check a log that they have emailed themselves. That only comes if somebody actually cares enough. No CEO can tell anybody, 'Hey, I want you to wake up every four hours and not sleep.' So one thing that was really special about early Facebook is this culture of total ownership. Nobody had to tell you to care more. Every single person was just like, 'I will make this. This will not fail under my watch. This will excel under my watch.' I wish I could say how to recreate that easily, but I do think that is a key defining feature of the Facebook culture. I haven't been there in a while, but I suspect that is a key part of our DNA from that place.
Yeah. Everyone's competitive also. You want to do it for the team. You also want to do it for yourself. The spamming stuff, I felt like I was personally going to war with these spammers across the country and I was going to win, damn it. I don't think anyone not only told me to do it, I don't think anyone even knew I was doing it at that time. It wasn't the thing I was doing for credit. I just was like, 'Get these guys.'
It's interesting. I do think that I personally subscribe to the philosophy that companies' cultures end up being, at least in some part, a microcosm of the founder's DNA. And I do think that a bunch of our competitiveness and 'failure is not an option' comes from Mark. To this day, I think it is insane that you ever ask anybody to work harder than you are willing to put in yourself. And I think that the way that at least at Facebook it just flew from the top. You can never ask anybody to do something that you wouldn't be willing to do yourself. And frankly, if you look at SPC a little bit right now, giving Ruchi some credit, when Ruchi comes in here, the first thing she'll do is make sure that there's no overflowing trash because she's like, 'I don't have a problem taking out the trash. I did that for the first five years. I'm happy to do it today.' The message really here is that there's no job too big or small. When you come in, it's a culture of total ownership. It was a great Reed Hastings story. He used to tell that great story at Netflix. For a previous job before Netflix, he came in to work early one day and the CEO was in there washing mugs. And he's like, 'Oh, it's nice of you to wash these mugs.' He goes, 'Yeah, I do it all the time.' And he had realized that for the previous three or four years he'd been leaving his dirty coffee mug on his desk and the CEO had been the one washing it and putting it back on his desk clean. He just thought there was a service to do that. And he's like, 'Oh my god, I'm so embarrassed.' I was like, 'Don't be embarrassed. You work hard at the company. The least I can do is wash your coffee mug. It's not a big deal.' And he always told that story about how he wanted to set his company. That's awesome.
Bos, you touched on this towards the end of your last answer, but I kind of feel, to your point, being the CTO at a company that is trying out building all these cool new things, everything from modern AI, AI super intelligence, to AR/VR glasses, to some of the stuff around robotics, kind of feels like a pretty fun minus one job. So talk to us a little bit more about that.
Yeah, incredibly. I'm incredibly fortunate to be in the position I am. I think it's not just me, I think all of us are pretty lucky. We have this idea that human time is relatively homogeneous, like every year is roughly the same as any other year. It's just not true. There are periods of time where there are disproportionate amounts of opportunity to have change, to participate in that change, to shape that change. I talked about this 2004 group of people that came to the industry. There were relatively few of us at a time when the internet was finally delivering on its true potential. And it was an incredible match, and I was very fortunate to be alive at that time in the space that I was. I think you as a community are in the same kind of place with AI right now and the rejuvenated energy to think differently about hardware, which didn't exist for good reason for the last 15 years. Phones are great, and that was obviously the right hardware platform to focus all the energy on, so it got 100% of the energy. And phones haven't changed in the last five years, and people are starting to feel like, 'Yep, that's not the best place to put the energy. We should put it into different things.' So it's a very exciting time. There have been these windows. If you haven't read 'Dealers of Lightning' about Xerox PARC, or 'The Dream Machine', tremendous important history of pre-computing told through the lens of JCR Licklider, but it's not really about him. It's the history of our industry. We're at that kind of a time now. And I say I just spent enough time painting fences and waxing cars to be the person who got to be in that job. So my job is intense, stressful, competitive, fun as hell.
Something you said is really interesting. I got to the industry in 2004, which now is a lot closer to the original dot-com bust. But sometimes I wonder if you run the experiment of when I joined in 2001 when everyone was actually just sad, I think you might have taken away pretty different lessons around some amount of pessimism and cynicism. Instead, the time that I entered felt like one of quiet optimism. 'Yeah, we can build a lot of cool things. You can build things.' So don't underestimate the time you're in now. Even though it feels stressful, there's a lot happening, but it is a unique time right now to actually be optimistic and to build. It will likely go through some euphoria and then level off for five or six years, and that will feel quiet relative to now. So one of the things we talk about at SPC a lot right now is, 'My god, if this is a time to work really hard, you will get compounding returns on the work that you put in right now.' This is why I personally have probably worked harder for the last two years than I have in a while, because it just feels like this is a time where it's actually rewarded.
That's 100% right. I wrote this internal memo called 'Year of Greatness' which leaked, so you can all read it. We'll talk about some leaks. Yeah, there's some leaks out there. But it's kind of misunderstood in a lot of context, but it's exactly this point. I'm like, 'Hey, every year of your career is not equivalent. This is one of the crazy ones. I guarantee you.' Especially for us in wearables, we have the best AI wearable in the market. It's not a question. Everyone's coming for us. Everyone sees it. Google, Apple, pick a favorite, Samsung, all of them. They see it. They want it. They're coming for us. Now, we got to double down right now. So it's an unusual time right now.
The other point that you made was also really interesting. You're right, if you had been into hardware and robotics like three years ago, you wouldn't get into SPC. I don't want to talk to people building really stupid robots. But now, actually the most interesting now, stupid robots are great. Robots are great. So tell us, you touched a little bit on this when we were talking earlier, but what do you think has changed about the robotics landscape that makes today interesting? And your thoughts on hardware versus software.
Yeah, robotics. Hardware is not the limiter. What's attractive about robotics is that the software is the limiter. We know that because even robots from 10 years ago, these pneumatic or hydraulic robots that had a lot of issues in terms of how to stand up and locomote and had a lot of issues in terms of speed of execution, they could still do very complex tasks. They could fold laundry, which is famously a hard challenge for robots. They could do all these hard tasks. They could only do them if they're being teleoperated by somebody. So somebody who put on a headset and had little grippers could make the robot do complex things, which tells you the hardware isn't the limiter. I'm not saying it's not a limiter over time, but it's not the limiter today. The hardware has caught up further. They're on kind of modern electromechanical actuators, it's made a lot of progress, but it doesn't really matter. They still can't do anything interesting or useful. Locomotion through reinforcement learning has made tremendous progress. You can make anything locomote. You can make a pogo stick locomote. It's not that hard anymore. But you still can't make the robot do anything interesting or useful. It turns out being able to stand upright and locomote is not enough to affect change in the world. It's all up here in the grippers and hands. And that is an area where we just did not have a path through conventional roboticists, who are always building heuristics just the same way my old newsfeed algorithm worked. We just couldn't possibly build up sophisticated enough heuristics. Machine learning was the breakthrough, and robotics was late to it. Robotics really only opened their eyes to modern machine learning like two or three years ago. They skipped the entirety of sparse neural networks. They didn't do any of it. So they are now speedrunning conventional machine learning and trying to get up to speed on large language models at a time when world models are becoming popular, becoming possible. We've all seen Sora. We've seen what direction things are headed. Still a lot of challenges between those things and physics simulations that allow you to do textures and pressures and deformation, but still tremendous progress. So I think the major delta between even two years ago today is our belief that the software will arrive. So now you're building out hardware in advance of the software arriving, that you think the software will be able to come.
