Athenaeum is a new podcast on the Tetragrammaton Network. Tom O'Neill unpacks his 20-year investigation into the CIA's mind control program, linking it to Charles Manson and the disturbing secrets buried in America's history. From missing FBI files to Hollywood cover... available wherever you get your podcasts. It's far more dangerous. Coming soon on the Athenaeum podcast.
the whole way and the last two years of high school my parents sent me to Exeter and for our senior project we had to do a project and Mark and I worked together on This was before Spotify or any of these recommender systems. The idea was that your music player, we used WinApp at the time, that was the popular product.
The idea was that the music player should be able to recommend you songs to play next, songs out of your library, because there wasn't a sort of internet distribution mechanism at the time. It would collect statistics on what you played, so all the music that you played yourself. It would just observe.
Things like, after you play this song, you tend to play this other song, or... these are the kind of patterns from the the songs that you're playing and it would learn those patterns and then it would be able to sort of be like a smart autoplay after that. It would only be for each user. Yeah. Yeah. Because this was sort of pre-internet. I was going to say, at this time, there was nothing like that then. Yeah. Yeah. Wow. Yeah. We had an idea. I remember working on a sort of...
spin-off idea from it that we were going to collect all the data from everyone, what they were all listening to, and then use that for is this idea called collaborative filtering, but basically find other people who have the same kind of music that you do. and notice that they have certain preferences and then recommend you even more music based on that. What do you think was going on at the time that even the idea to do that was a possibility? Where did that come from, do you think?
That's a good question. You know, I guess I think it was kind of obvious to me, I guess to the point where it's hard for me to say where it came from. This is pre-iTunes, right? Yeah. So it would either be stuff that you rip or stuff that you got on Napster? Yeah. And I had been making different kinds of programs for a while at that point. And I think... Once you do that, you start to think in the space of possible programs.
What were some of the other early programs? I started out making games for myself. Were you a gamer as well? Not like a serious gamer, but when I was in middle school. I grew up at a time where It was possible for me alone. to make a game that was almost as fun to play.
as the kind of games that I could get on floppy disks at the time or the way they were being distributed. Games today are just so advanced and so optimized. Do you remember what any of the games were that people would know from that time?
there was this game called it was like comanche where you you fly a plane around there was the video games were a little more advanced but it was games like super mario 3. yeah or sonic the hedgehog it sort of like just come out and you know it's a lot of work to make something like that but even just like as a self-taught programmer it wasn't that hard to make games that were on a similar level
Yeah, yeah, basically. It was your first computer. So my dad was a professor at college in New York City. One day he just brought home a computer. I think the intent was to be able to prepare things for his classes or do research. And it just immediately was, it was by far the most interesting object in my environment. How old were you at that time? I think I was in...
fourth grade. Would any of the other kids you went to school with have had a computer at their house or no? I think so. I mean, it was starting to be the... If a parent had a reason to have it for work, then they might get one. Would it be like a Radio Shack computer? It was a Packard Bell. 486. Did it have the green screen with like light color green text? It had Windows 3.1. I think it was just right at the time when visual graphic user interfaces were getting to the point where.
They were mainstream and that had made computers accessible, I think, to a much wider set of people. So my dad brought that home. I didn't get into programming for maybe a few years after that. There was just so much that you could do on it. And I remember just going through every single program, just open the program, see what the program can do, and see all the limits of it.
And then I think the next big kind of step was At some point, we got a modem, which was 2400 baud, a very, very slow modem in today's standard. but you could connect to the internet. And that just unlocked a whole new level of capability for the computer. In middle school, one of my friend's older brothers showed him how to program something very simple. And he showed me and I got started off of that. And then from then on, there was this program called QBasic.
It was basically a very simple programming language that came with most computers at the time. It had a very good help system. So you could look up a command and nothing like what you get today with AI, which is amazing for helping people. But it was enough that you could get pretty far on your own. And so that just became my. my hobby and I get so deeply kind of encroached in it that I think I got just very good.
at programming relative to what anyone normal at the time would have ever been exposed to. How many hours a day would you say you were on the computer at that age? It was most of the day. I mean, you know, I would go to school and when I was at school, I'd be thinking about programming. You'd be making notes or just thinking about it? Just thinking. And would you be thinking about what you could do, like possible things to try when you got home?
Yeah, or even if you get into it enough, you start to be able to just sort of load the whole state of the program or whatever you're working on into your mind. and you might even solve a bug. You might figure something out just by thinking part, maybe even sometimes this will happen subconsciously when you're sleeping, you wake up and you have the solution. You don't need to go that far. You can just think through things.
It was definitely, it was the majority of my energy. And during school, would you look forward to getting home to get on the computer? Oh, yeah, yeah, absolutely, yeah. And would you do it until you went to sleep, would you say? Or would you stay up to do it? I think I slept pretty regularly. I mean, I had this constant kind of To them, it was as if I was playing video games. That was kind of their experience of it. Wasting your time. Yeah. And it's like my life is being sucked up by this thing.
what should they do should they take it away from me but oh it seems like maybe you know someone said it was good for them to them but they don't know whether to trust that person and yeah and you know i wonder if they had tried to push me into it like would i have enjoyed it as much i don't know in some ways maybe their resistance made it better yeah it's interesting it's it's very interesting yeah so eventually you meet mark and you guys have this idea for uh
a way to program music, essentially. And did it work? Yeah, I enjoyed using it. He enjoyed using it. We had a bunch of friends who we gave the software to. It wasn't really a business. We didn't have the kind of ambition to make it. It all depends on what you mean by work. And it worked for what we were going for. It did what you were hoping it would do. Yeah. And when was the first idea that there could be a business model associated with something you were doing with computing?
I remember when I was in high school, it was clear that that was the thing I was the best at. And I remember talking to My dad, at one point, he had sort of thought I might go and try to be a doctor. He thought, for whatever reason, I don't go to math and science. Like, that's a kind of... career path that you might suggest for someone like that and
I never really was interested in that direction. But my sense is that when Google came around, that was the first time for kind of like, you know, middle class. parents in the New York City suburb area. That was the first time where being a software engineer kind of seemed like a... Possible job. Yeah, like a real job or like a good job that you could get or something that you might prefer over being a doctor, for example.
if that was what you were interested in so sometime i think it was sometime around then i was to me it was just my hobby and i just loved doing it and i knew i was gonna do it because of that and i didn't care whether it was gonna make more or less money than the other career options. You'd say years and years of doing it as a passion project, hobby, and then it turned into more, but you would have done it either way. That was never the intention. When did it become professional for you?
When I was in high school, I got some contract work over the internet to, there were some forums where people could post. Like, I need you to write this program to solve a problem that I have and I'm willing to pay X. And I did a little bit of that. Was it fun? Not as fun. The task wasn't as interesting to you as the ones you would make up for yourself? Yeah, I just didn't care. There was some job I had to do about converting some file format from one thing to another.
It was interesting. I certainly enjoyed doing that more than I would have working at the mall or something like that. I didn't care about the files that I was converting. Is the feeling of solving the puzzle, does that feel good? Yeah. Not as good as creating something, but I'd say that's positive. So we'll call that freelance work. And then when did it become like more of a full-time pursuit? I got an internship.
at MIT after my freshman year of college. And then there was like a research lab. And then Facebook started, and that was Mark and these other people at Harvard who were building that. And then they moved out to California for the summer. to work on it, and then... Do you know why they chose to move to California to work on it? I don't know the exact thinking they went through, but...
Even I was at college at Caltech in LA and it was very clear to me that I needed to come up to the Bay Area for that. That was just where everything was. What else was going on at that time? Just so I have an idea. Because it was not. So I started college in 2002, and so the dot-com bubble had just burst. But there were all these dot-com companies that were in the Bay Area. So there was eBay, there was Yahoo, there was Google. Cisco, it seemed pretty clear that everything was there. And...
I assume the same kind of pull just occurred to the Facebook people. And I imagine he would be able to meet other people. who were as interested as as you were yeah and that has to feel good yeah yeah yeah so they moved out to palo alto for the summer and and i moved up to to live with them and we were going to work on it was funny that the idea at the time was that facebook seemed like a It was a college social network that, again, was, I think, just a project at the time.
It seemed like it was unclear where it could go, but it had at least established this base of users and friend network. And the idea was that we were going to use that as a launching ground for this new product called Wirehog. What was that going to be? The idea was that it would be a file sharing program, kind of like Napster, but you would be sharing files with your friends instead of with strangers.
And so you'd be able to see what is the music that your friends like and then download it and listen to it and recommend music to other people. Was music the main files that would be shared or were there other things that you would share? We made it general. And there was videos and documents, but I think it was pretty clear that music was the thing people cared the most about. At the time... There just wasn't great video content online. People would pirate.
dvds but those files got so big that it wasn't a good use case sounds like a fun time yeah yeah it was it was really fun yeah we and so I think when they first got to Palo Alto. They had an experience of walking down the street from where they were staying to they found some place on Craigslist. They were walking down the street to go get food or something. on a random residential street of Palo Alto, ran into Sean Parker, who had been one of the founders of Napster.
Mark had met him in some other context. And because of that kind of random run-in, he ended up moving in to the house and lived with us and eventually joined Facebook as well because of that. So at that time, how many people were involved in Facebook? The founders, Mark and Dustin, and then there was Andrew McCollum. another one of the founders. Mark built the first version of it himself and then brought in a few other people very quickly. How old were all of you at that time?
19, 20. And at this time, the idea for Facebook was... connecting college students. When did it shift from college students to beyond? So there was a process by which it sort of became a real company. And this was summer of 2004 that this Palo Alto house and during that summer they Sean Parker connected them with Peter Thiel who And then at the end of the summer, I think they ended up deciding to not go back to college, at least initially.
And at least, I think Mark and Dustin were the two who made that choice. And then... summer 2005 i went back to school for that school year and then summer 2005 i came back there was this round of funding from excel and they got a real office in downtown Palo Alto. The year that you left and then came back, how different was it when you came back? It was pretty different.
you know living in a house where you're all just living there and working on something versus like having a real office and having funding and having a business with ad revenue and There started to be people who were hired who were not just friends of the founders. And it felt different, but it was still very different from a normal, you know, I would expect a normal company to be. LMNT Element Electrolytes have you ever felt dehydrated after an intense workout
Or a long day in the sun? Do you want to maximize your endurance and feel your best? Add Element Electrolytes to your daily routine. Perform better and sleep deeper. Improve your cognitive function. Experience an increase in steady energy with fewer headaches and fewer muscle cramps. Element electrolytes. Drink it in the sauna.
Refreshing flavors include grapefruit, citrus, watermelon, and chocolate salt. Formulated with the perfect balance of sodium, potassium, and magnesium to keep you hydrated and energized throughout the day. These minerals help conduct the electricity that powers your nervous system so you can perform at your very best. Element electrolytes are sugar-free, keto-friendly, and great tasting.
Minerals are the stuff of life. So visit drinklmnt.com slash tetra. And stay salty with Element Electrolite. LMNT. You said that there was another thing that you guys were going to build after Facebook. What was that? Oh, yeah. So this was even summer 2004. Yeah. We were going to build this product called Wirehog. And what is that? This was going to be...
the ability to share files with your friends and mostly music and i think what happened was we actually we built a product and we had it we had it out for a while and it was it was growing i remember looking at the The usage grew, it was like every week it grew by 5% for a while. That's a very fast rate of growth for a new product. But... Was it only by invitation that people could... Yeah.
How many people did you invite? You can't grow past the people you're inviting, right? Yeah, but then they can invite other people. It can just grow exponentially if that process keeps going. So it was growing, but Facebook was growing even faster and they raised funding and started to be a real company. the company sort of demanded more and more of Mark's focus and
Then at some point there was a question about legal liability from file sharing. You don't want to be in a legal gray area if you have another business that's just working very well and has none of that risk. Did that come up because of Napster, the legality issue? Yeah, I think there was... There was sort of an evolving understanding of where the lines were and how much the platform inherited liability for the actions of the users.
And it was all very vague, but that got to be more of a concern. And part of the reason for... focusing on Wirehog to begin with was the idea that Facebook was sort of And I think it just became clear that that was not true. So in 2004, it was college only and sort of expanded to most of the college market. Question about that. I know it started at one college and it was for communication at that college. When it expanded, did it expand for each college to be unto itself or was it for?
Was there cross-pollination always between the colleges? So it actually started out as a standalone application per college. There was no interaction between the colleges. And then at one point there was a... project done to allow for cross-college friend connections. And how was that perceived? I don't think that was a controversial change. It's a big change though. Yeah. The privacy model was that you could see all the profiles of everyone at your college.
and then you could see the profiles of people at other colleges if they had added you as a friend. So it wasn't very threatening to privacy to make that change. I think it was probably important from a business point of view because Facebook could sort of leverage the user base at the existing colleges where it had strength.
to then bootstrap these other colleges where it might have otherwise been hard to get enough momentum to make it work. It also makes sense that kids who go from high school to different colleges could still continue to keep in touch. That makes sense. So then in 2005, they added high school. Facebook for high school kids. And that actually did not do very well at first.
One of the things that worked about Facebook was you had to, the way you would prove to the Facebook application that you were actually at a college was you would have a .edu email address. and it would send you an email and you have to click a link to confirm. And that created this safety where you knew that everyone who could see your profile was people who at least had email addresses at that college. And high schools didn't have this in general. And so it was difficult to...
come up with the right privacy structure given that. There was first a system where you had to get like two invites from other people at your high school to like vouch for you that you actually were at that high school then that that was this massive source of friction that just crushed any hope of that working and then
There was another system where people who were in college who said that they had gone to that high school could vouch for you. And I don't know if that ever, it never really worked out. There were people using it, but it did not. sort of take off and dominate in the way that Facebook had at college.
On Facebook, you always had to be a real person. You couldn't be like an avatar and you couldn't be an anonymous account. Is that correct? Yeah, that was correct. There was Sometimes there were things where people could get a bunch of extra email addresses at their college and pose as multiple people and people would do this to create.
Joke accounts and generally I think the platform would try to shut that down There was a belief that and I think this belief is a little bit looser today, but there was a belief that the kind of integrity of the network and the authenticity of it mattered and the trust you would have in the the information you were seeing could be higher if you could just assume that everyone was who they were saying.
who they were and and this was actually a big difference from the internet at the time it's not it's not very different from the internet today where there's a lot of people are publishing on the internet under their real name and that's kind of the default even when you're not forced into it But at the time... there were very few places where
people were comfortable putting their real names on the Internet. There was a lot of forums where everything was anonymous or pseudonymous. There was even blogging. there was a little bit of that going but a lot of those people were not under their real names and there was kind of a just especially older people, especially people outside of college, there was an idea that the internet is kind of scary. The other people on the internet are scary.
there's you know why are they there when they there must be people who don't have friends in real life or there must be you know there's kind of a almost like the in America today, there's not much hitchhiking anymore. I remember people used to be really afraid of buying things online. That was always scary, like putting your credit card online. It seemed dangerous at one point.
but you know i think hitchhiking is something that is scary today because you think like who's the kind of person that would pick you up if you're hitchhiking and who are the kind of people? and so the whole trust can unravel. The Internet was in that state where there was none of that baseline of trust, and so you have to assume the worst-case scenario.
And if it came from a place where a lot of it's anonymous and you're the first place that's not, I could understand that. And what were the kind of things that college students would be sharing on Facebook? So you could put up one picture for your profile picture at first and you could connect to friends and you could fill out your profile and you could say,
These are the things I'm interested in. You can actually, for maybe the first 20 colleges, you could say what courses you were taking. It had the whole course directory built in, and you could see who else was taking the same classes as you. And there was a feature called the wall where you could post a comment basically on someone else's profile. It was fun. It was this view into the social structure of the world you were in that was just, it was unbelievable how...
how much time people would spend and just what the experience of it. I loved it as a user. I remember just spending. hours just clicking around and just looking at like who are these people and you know what are they saying to each other and and there's no boundary of you don't have to overcome your shyness like in real life If you went and talked to someone, it's scary. But if they're putting themselves up for you to talk to. Right. And the friend mechanism.
which I think this was really pioneered by Friendster. But the mechanism works. where you would go to someone's profile, you'd say, hey, this is someone who I know, and you could click Add Friend. And then they would get a notification saying, this person has added you as a friend you want to accept or ignore. So it was based on someone who you already knew in real life. That was effectively what. That's how it started. That was how it got used.
And there's a lot of ways that socializing in real life involves. There's these kind of complicated... signaling dances around you know maybe you like invite someone to do something and they don't want to do it and they'll they'll make some excuse or they'll say they're busy or you have to get an introduction through someone else who you know to show that you're credible and sort of like, you know, just
how long you hold eye contact for says something. And so there's this whole like, you know, dance and games and everything that goes on in real life socializing. And there's nothing wrong with that, but that's just how the world works. Facebook simplified it down to like, you can just click add friend and they can just ignore it. They can pretend they didn't see it and then they can confirm it. Then you're solidified as being connected.
you're in this private space with them. And there's a higher level of trust that can come from that than the previous internet products. And then eventually, would you be able to friend people who you didn't know? Yeah, so you could add anyone as a friend, but the norm on the product and I think the fact that it was called add friend and the fact that your friends are listed on your profile. That enforced a norm that people generally only added people as friends if they knew them.
The word friend, I think if you polled most people at the time, the people who they were friends with on Facebook was not the same as who they would describe as their friends. your actual friends or you know the people you hang out with you have a higher level of trust and relationship with
the people you added as friends on Facebook were basically the people you knew, like people who you knew their names, someone who you might say hi to if you walked by them. And that's not the same as a friend, but the word... worked as sort of the representation on the product for someone who I know. And it was a nice invitation. It's nice. Better than acquaintance, probably. But it was more like acquaintance, as you'd say.
So high school I think kind of had mixed results and then in there was a project called Open Registration and this was the idea that anyone was going to be allowed to register for Facebook without having to have an email address at a dot edu to prove they were from a college or go through the the high school friction was it a big decision to do that yeah it was yeah it was i remember
There was some advice that came from some investors secondhand that was like, you're destroying the differentiation and the uniqueness and everything that has made this what it is, you're throwing it away. I guess there's a lot of sort of wisdom that vertical specific product. do better than horizontal ones and this went against that and it was I think it was definitely against the common wisdom to open up like that.
I think for most of us at the time who were people who had been college students and experienced Facebook. It seemed obvious to do it. It seemed... like of course you want you still want this product even when you're not in college anymore you want to be connected to all your friends and you want to be able to so in the original version once you graduated you'd be off it
I think you actually have to stay on it. And this all happened over just a few years. So there weren't that many people. There weren't people who had graduated 10 years ago who were still around. So I think there was just like one or two years of people who were kind of alumni who were still on the network. And then when did it open wide? Was that the opening wide? Yeah, so that was open registration. And so that was in 2006. And I remember at first it did not do very well.
I remember seeing there was a user research report that talked about how people who had not been part of college, people who hadn't gone to college or people who graduated earlier and had never experienced Facebook, they felt... One, that they didn't understand what this thing was. And there was a big effort. that nobody can understand what is Facebook. I actually think this is a pretty generalizable thing. I think a lot of new product.
trying to answer the question, what is the product? I think it's not a good thing to worry about if there's not an obvious answer. Eventually, if you ask someone today, what is Facebook, they'll say it's a social network. But a social network was not a thing that meant anything to anyone at the time.
so there was this challenge about explaining like what is this thing and why should you sign up for it and then there was the people who understood it a little more they even felt like creepy joining the network because they were they were not in college and so it was like it was almost like someone who would like hang around a college campus without being in college it felt like it was like not for them and it was it was weird So that was kind of the initial kind of feeling.
MySpace was the effective competition at the time. It was never really something we worried a lot about, but it was an alternative that people had seen. It took a while for Facebook to sort of get momentum outside of college. Was everything that happened all word of mouth or was there any kind of advertising or marketing? All word of mouth. All word of mouth. They would do some experiments with marketing, but it never...
Never added up to much. So it's usually people who used it and liked it told their friends about it. That's how it spread. Yeah. Because with any of these social products, it's only a good experience to be on Facebook if your friends are also on Facebook. And so any marketing that goes out to the whole world, most of the people who see it, number one, they just they don't understand what it is. And so they're going to ignore it. But even if they do decide to sign up.
90% of them have zero friends on the network at the time. They're going to have a terrible experience and they're going to leave. widespread marketing just doesn't work for a social product. There were experiments, but it wasn't very successful. The thing that really moved the needle on it was I was running this growth team at the time and
we made some changes to make it much easier for the people currently on Facebook to invite their friends outside of Facebook. That makes sense. It makes a lot of sense. now, but at the time it was It just wasn't clear that that was going to be the dominant thing compared to any traditional marketing stuff. They never tell you anything about that. They're going to talk about running ads or doing a PR campaign.
And then based on that, how fast did it start growing? It was growing between 2% and 3% a week. If you work on all of these products, this is how you think in percent per week, that's a very common way to judge how well a network is growing. At 3 percent per week, it compounds. 3% a week for a year you'll grow 5x over the course of the year and 2% a week I think you'll grow 3x.
And that difference is very big in terms of how big you're going to get over multiple years. So we did a lot of work to try to... keep the growth rate closer to How would you say MySpace and Facebook were different at that time? So I think there were a few key differences. One was MySpace profiles were open to the public. So anyone could see your MySpace profile, whereas Facebook, only your friends could see your Facebook profile.
There's a huge difference, drives all kinds of differences in how people use the product and what they use it for. Do you think that to that point, if it's just your friends seeing it, you're more comfortable to share more personal things? Do you think that's what it is? Yeah, yeah, yeah. That's a huge part of it. I mean, when Facebook was college only, There would be people putting up photos of them like drinking, like photos of parties and beer pong and even sometimes like...
smoking pot. It was stuff that you expected that even your parents wouldn't see. Stuff you'd expect to find on a college campus. Exactly, yeah. And this was even a challenge for Facebook. As it started getting bigger with open registration, you start to have basically like the parents join the network. And that has this chilling effect on what everyone will post. That's been a huge issue for Facebook for many years, just like there's been this decline in.
posting to your profile every year because There's a lot of different research on what drives it, but a lot of those sharing has basically migrated to messaging apps and to private groups. Even though you had to add everyone as a friend who's going to see your profile. There's too much social pressure. You end up having to add, you add your parents, you add your boss, you add your aunts and uncles and- Changes. You just start this chilling effect.
In a world of artificial highs and harsh stimulants, there is something different, something clean, something precise. Athletic nicotine. Not the primitive products found behind convenience store counters. Not the aggressive buzz that leaves you jittery. but a careful calibration of clean energy and focused clarity. Athletic nicotine.
The lowest dose tobacco-free nicotine available. Made entirely in the USA. No artificial sweetness. Just pure, purposeful elevation. Athletic nicotine is a performance nootropic. Athletic nicotine is a tool for shifting mindsets. Athletic nicotine is a partner in pursuit of excellence. inspired results athletic nicotine More clarity. Less noise. Athletic nicotine. More accuracy.
From top athletes pushing their limits to artists pursuing their vision, Athletic Nicotine offers the lift you've been looking for. Learn more at athleticnicotine.com slash tetra and experience next level performance with athletes. This product contains nicotine. Several times you mentioned using research. Tell me about that idea of using research. Who would do the research and how would you know what to research? There was a few different...
ways of trying to get that information. So there's a user research where someone would invite people into a lab setting and just basically watch them use the product and that I think it was valuable sometimes other times I think it had questionable value because the person that you're studying is like just someone who wasn't on the network and it was just too much of a leap to get them into a mindset where you can learn something useful from them.
And then there was a whole set of basically what today is called data science. So this is employees who had a background in... statistics or some kind of science. and they would do a lot of work to look at the data. in the system and see how people were using it. You can run controlled experiments where you take half the users and show them a slightly different version of the product. And then you can see like a week later.
are they using the product more or less or do they end up adding more friends or do they express more positive feelings through whatever you can measure. and so there's a lot of learnings come from data science. Did you form the idea for Quora while you were at Facebook? Well, actually, it's an interesting story. After I left Facebook... Why'd you leave? I wanted to do my own thing. I wanted more control over what I was doing. I had always wanted to start a company.
or just do my own projects and ultimately those turn into companies if they get big enough. But Facebook was just, it was a very good opportunity, so I went with it. But at some point, as facebook got bigger it felt like the things i was doing were they were less like creative or less about like deciding what the direction of things was going to be and more about
how to manage a large-scale team and how to run processes. It was valuable stuff and I'm glad that I learned all that, but it just wasn't.
It was more like when you were a freelancer doing jobs. Exactly, yeah. It was less creative. Yeah, yeah. So you decide to leave, and then what happened next? Yeah, so I decided to leave, and this was in... summer of 2008 i spent some time just just experimenting just kind of working by myself and thinking i had a bunch of different ideas i was kicking around and my ideas at first were
ideas that were too close to Facebook. Like they were ideas that would have been better done in Facebook. And it took me a while to sort of get out of that. Just my worldview had been shaped.
you know ways that are probably good in a lot of ways but it was too much for for being independent and eventually i i was thinking about what what do i really enjoy and what do i really like doing and i thought about it i really like learning and I like understanding things and I was looking at the products on the internet at the time that
where you can do something like that. There were these products like Yahoo Answers and Answers.com. It was actually a whole long list of question and answer products. I thought that the products that were there were just not really, they had just not gotten it right. There was a lot of demand, and so that showed that there were a lot of people who potentially would want to contribute to a... a network like that, but none of those products had personalization.
So when you went to Facebook, you saw a very different experience than what someone else saw based on who your friends were. That was just how it worked from the beginning. But most of the web, most of the internet was built at a time when everyone would see the same thing. So you would go to the Yahoo homepage and see the same thing that everyone else would see. I think just coming from Facebook and maybe just thinking from first principles.
It seemed crazy that when you want to go and learn something new, or if what you want to do is answer questions, that you would see the same questions as everyone else.
But that was how all the services worked at the time. We built Quora from the very beginning, just oriented around the idea that you're going to have the topics that you know about and are interested in, and you're going to specify what those are, and then we're just going to show you the questions and answers that are relevant to you.
It's a simple change to explain like this, but it results in like a radically different network that gets built out around it, the kinds of people that are attracted to the product. Would you call that an algorithm? When it started out, it was time order. So you would just say, these are the topics that I'm interested in. And we would just show you.
all the content that had been created on those topics in time order. So I think usually when people say algorithm, they are thinking about how it's ranked, like stuff gets out of time order. you might show the best stuff first instead of the newest stuff first. Is that typically what an algorithm does? I don't know what an algorithm does. Technically, an algorithm is It's just a procedure that a computer follows. Or it could even be a human following an algorithm, but it could be.
An algorithm could say something like, We're going to sort the content in your feed based on the number of votes that it got. 10 votes would be like a thumbs up or a like or something. Whatever you want. It could be an algorithm. It could just say, we're going to rank by that. That's an algorithm.
You rank it by time, that's a different algorithm. I think when people use the word algorithm though, they're usually talking about more complicated algorithms. You might not want to rank by the number of votes because would then be at the top and then it would get more votes because it's at the top as you get this rich get richer problem. Then you can try to normalize for that and say how many votes did something get per.
impression that it had. But then that's not perfect either. And you want to show people not just what was the most popular per impression overall, but what are they going to like? Were these all things that were already understood, or is what you're talking about something that was new at the time? I mean, the math and the ideas behind this are not very complicated. I mean the use of it. Yeah, so the thing that was new was... If you want to do personalization well,
you need to build the whole product around that premise. So was Core the first thing built on personalization, would you say? It was the first knowledge sharing product. It was definitely the first question and answer product. On Yahoo Answers, you could go and dig into There was a section for programming and you could go and find programming questions in that section, but you had to actively like choose to dig in there and then most people just don't.
Friction is a very big deal in these products. Most people don't click and they'll just stay on the homepage and then everything just gets to become lowest common denominator content. So to execute on a personalization-oriented product, it's a pretty radical change if you're starting from a point of... of a product where everyone sees the same thing. So when you turn it on, you're getting content for you as opposed to content for everybody. Yeah.
This is how almost all popular products today work, but at the time it was not the case. It was not. Did the reason you want that because that's what you wanted for you or that you assume that that's what other people would like. If you work on these products for enough time, you have to think about what is good for the network. And so it's not even that what you want or what the other people want. It's like what will create the best network outcome when everyone is subject to these rules.
And you can build up intuition for what's good for... the network this is a challenge you know we have we have new employees who we hire every month and people come in and they don't have this intuition it's easy to have intuition about what you want yourself. I think a lot of people don't even get that great, but that's a skill you can learn to understand what you really want.
And understanding what other people want, you can do research, you can do surveys, you can go ask people. But trying to build intuition for what's going to be good for the network. is difficult. You think it only comes from working on a network for a long time? Is it just experience? You can come at it different ways. Definitely you need experience. But I mentioned earlier that, you know, Facebook, you only could see
content from people who you're friends with and MySpace, anyone could see anything. That was intuitive to you that that's going to make people share more. More personal. It's going to be more personal. You can think through it, but. you have to think some of these things get a lot more complicated than just that inference there. So one example is who is allowed to write comments on content if content is posted on one of these networks?
There's a person who created the content and there's their friends or whoever they trust. Should only those people be allowed to comment on the content? Or should you let everyone comment on the content? And depending on the setting, one of these other of these might be the right choice.
If you open up to the whole world being allowed to comment on content, you may or may not want that. And you can have these emergent dynamics in the network where... some people who are really like mean and aggressive. and devoted to their cause, they will aggressively comment on every new thing that someone writes if they don't like the person. And then those people can drive away everyone else.
other people, you know, whoever is the most aggressive. But blocking is a lot of work. And if you're a popular creator, you can spend your whole day blocking people. And it can be very discouraging to get a nasty comment. Even if most of the comments you get are positive, one nasty comment, people can be very sensitive to that. But at the same time, on Quora, for example, what if someone writes an answer that's just wrong?
It's kind of important that someone else can come along and leave a comment and say that, hey, this is wrong, depending on the motive of the creator.
It might not be something they want. Sometimes it might be an opinion, an opinion about politics. Who's to say whether it's right or wrong? But... you know someone solved a math problem and it's wrong you kind of want someone to be able to to come and say that this is wrong and it's very hard to have a different policy for politics versus math you kind of need you generally need like simple rules that everyone can understand
And so there's all these kind of dials you can turn about how aggressive and mean you want, but like truth-seeking you want the network to be. so people ask general questions and then how do the answerers see the questions in their field there's a few different ways but we try to predict what our answers knowledgeable about and interested in answering. The answerers can tell us these are the topics that I know about.
We'll also just get a sense from which topics they're browsing themselves, like where are they reading answers. and we'll show them questions. We don't have to get it perfect, so it might only be one in 10 or one in 100 of the questions that someone sees that they can answer or want to answer. but that can still be a much higher hit rate than if we just showed random questions to everyone. And the way that you know that the answers are accurate is because nobody complains?
I mean, so we don't know that every answer is accurate. We just do our best. These are crowdsourced answers essentially. Yeah, it's entirely user-generated content. So cool. Yeah, I don't know what percent of the people who use the product. fully understand the way horror works is there's a very small base i mean it's in the millions but it's relatively small base of people who
ask questions and answer questions. And then there's hundreds of millions of people who read the answers. And I don't know actually what percent of the hundreds of millions. are aware of what the incentive structure is and how all this works. I think they just know the last time I came to Quora.
I got an answer and it was helpful to me. The next time I see it as a link, I'm going to click on it again. Or if someone has a bad experience, then they're going to be less likely to come. We can't ensure the correctness of every answer. show the correct answers more often than the bad answers. And we look at what do people vote on? Is there like a button for this was helpful or? There's a upvote downvote button. and then you can leave comments, and then recently we've been able to use
AI to look at answers as well as another signal. So in some ways, it's like the experience of today's AI, but it was before today's AI was available. Yeah. It's the same principle. as AI just without AI. Yeah. There's a kind of legal gray area around this right now, but- My expectation is that Quora will be training material for AI. We let creators opt out of this if they don't want to participate in this part of it.
I think one important role that Cora will play in a world even as AI gets more and more powerful is it's sort of a way for humans and for all of society to transmit its knowledge into a format that then is useful to all the other people, but also useful to AI, which is then going to help other people when they use the AI. How would you say the experience of the answers coming from humans differs from the answers coming from AI today?
So there's a certain set of questions where AI can answer perfectly and that's the best experience. But then there's another set of questions. you know what is it like to live in this particular town or you know you probably have a lot of stories about how did this How did this band do this thing when they were making the song? And a lot of that knowledge is only in someone's head. It's not written down anywhere.
AI can get very smart and it can solve a lot of problems, but if it doesn't have the knowledge of a particular fact, it's just not going to be able to answer that question. So there's a whole set of knowledge that we have that's kind of unique. And then there's, I think, for the people who are looking for knowledge. There's a certain set of questions where even if AI could tell them the same thing that a human could tell them.
They can trust it more if it's coming from a human and they can understand who is that person who said that and is that someone who really should know this and does know this. Are they vetted in that way? Like when a person answers, can you know who's answering you on Quora or no? Yeah, so people generally use their real names. We don't enforce that, but that's the norm. And people can list things like, I studied in this.
college or I live in this city. And we will present those as facts if we think it's helpful to people to understand the context. Local knowledge for sure is better coming from a local. Yeah. And what year did chorus start? We started in 2009. And how did it... change from the initial inception until now. So when we started, it was We had personalization happening through people choosing which topics they wanted to subscribe to.
Over time, we've started to use machine learning and AI to sort of reduce the friction for everyone. We can automatically infer what topics people know about or are interested in. We can automatically do a lot of quality control that we used to have to do manually. and then i think the other thing is just the scale is so much bigger so you know we have
I think it's something like 400 million people in a given month who typically come to the product and hundreds of millions of answers. And so at scale, it's kind of a different experience than when it was small, it was... It was like us and our friends who we started with. And so it was great for a certain set of questions about Silicon Valley or questions about starting a company. These days it's kind of, it's just like long tail, it covers everything and it's a little bit less.
intimate of an experience, but it can just reach a lot more people. And was it also built all through word of mouth? Yeah. I mean, it's kind of similar story to Facebook, actually. We tried various forms of marketing over the years, but it doesn't really make a dent in the organic growth. Were there any breakthrough moments of something that really moved the needle in the chorus story?
It's been pretty steady. You know, in the same way that I would say that at Facebook, making it easier for people to invite their friends was a big lever. There have been... similar things where basically make it easier for people to find the questions or the answers that they care about. And using AI usually has been the theme for that.
breaking down barriers to adoption, so things like launching mobile applications. We started as a web-only company. Launching mobile applications was big. How long into it did mobile go? We started in 2010. The iPhone came out around then, but... and it wasn't a very big market at the time. I think maybe 2012 was when mobile apps and mobile consumer usage really started to.
to spike so two three years in what percentage now is mobile versus desktop majority is mobile do people type in questions or ask it questions It's mostly typing, although if you're on a phone, you can use the The speech detection built into the keyboard. And we don't know when you do that, so we don't have perfect data on this. But I would guess it's mostly typing. If you look at text messages sent today, I think the vast majority are still typed and not spoken.
So much of today's life happens on the web. Squarespace is your home base for building your dream presence in an online world. Designing a website is easy using one of Squarespace's best in class templates. with the built-in style kit you can change fonts imagery margins and menus so your design will be perfectly tailored to your needs discover unbreakable creativity with fluid
A highly intuitive drag and drop editor. No coding or technical experience is required. Understand your site's performance with in-depth website analytics tools. is everything you need to succeed online. Create a blog. Launch an online store. The Squarespace app helps you run your business from anywhere. Track inventory and connect with customers while you're on the go. Whether you're just starting out or already managing a successful brand.
Squarespace makes it easy to create and customize a beautiful website. Visit squarespace.com slash tetra and get started today. On a search engine, you put in what you're looking for and you get a list of links. Whereas with AI, you ask it a question and it gives you an answer. It's a different experience. It's like it skips the step of...
giving all the parts. So in core, is it more like the AI where you get the answer or you get Here are the different answers that have been given over the time we've been doing this. A lot of usage of Quora actually is someone will search in a search engine and a Quora page will show up with a question that we have. I see. And there will be answers that people have written on Quora on that page.
But there are often hundreds of core pages, core questions that are relevant to that question. So we will show answers from all of the. questions that we think are relevant on the page for a given question. Do you use the service? Yeah. I mean, there's an aspect of the product that We haven't talked a lot about where you can just show up at Quora and... and read answers without having to ask any question at all.
If you do this enough, the personalization algorithm will show you stuff that's very relevant to you. One of the things I'm interested in is the history of computing and programming languages and environments for programming languages. And so there's this.
Computer scientist Alan Kay, he created a language called Smalltalk, which was very influential in programming language design. He has written a lot of answers on Quora about the history of Smalltalk and the different projects he was involved in and his thoughts on what made different things successful. So when you go on, that will be served up to you? Yeah, so a lot of the time, I will see something from him. I see. So it's almost like personalized, trusted sources. Does that sound right? Yeah.
I'm following him, so I've kind of explicitly opted in to see his stuff. Most of the distribution on the network is not through explicit trusted follow links. Most of the distribution happens through the algorithm predicting what someone's going to be interested in. I remember thinking about it when I was in high school. If you're programming computers, you start to feel like, wow, like I, as a human. I'm maybe not doing anything different from what is going on in the computer.
There's obviously like you're able to operate at a level that's far beyond what, at least when I was in high school, you can operate at a level that's far beyond where the computer is. you start to think that there's just so many analogies, you can see the same kind of patterns. And so you start to think that maybe someday these computers are going to be able to do everything that a human can do.
How different is the experience of AI today versus what you imagined it could be in high school? I think when I was in high school, it was not clear to me that AI was going to be created through a learning, a training process. as opposed to the idea that it would be programmed. How would that work if that were to happen? If you were to program it, how would that work? So this is, if you look at how they used to do robotics.
If you want to make a robot walk, you have a physics model and you do all these calculations about where the angles of the different joints need to be at different time steps so that it won't fall over. And you can make that work. For years there have been robots that operate like this.
But the modern way that everyone has kind of settled on to do robotics is to use learning and have the software figure out how to control the joints in the right way so that it can balance and that gets you to much generally just much more robust systems that can work in unpredictable environments and look more human when you see them. So I didn't really get into the kind of math or...
machine learning stuff that you'd need to understand that until I was in college. But I also think that was kind of how people expected it to go. Is it only possible because of the internet? Like without the internet, could you have a large language model? So you certainly need data. Compared to humans, these... machine learning models are a lot less efficient in their use of data. So to learn a given skill, they need to see more examples than a human would need to see to learn the same skill.
At least today, this is changing and I think will change quickly, but generally today. there's less efficiency there. You need to train a machine learning model on a lot more data than what a human can ever read to get to the same level of performance. you would need some other source of that data. But it could be that you just scan a ton of bugs.
I think I'm pretty confident that that would work. So you don't need the internet in that sense. It certainly helps to have the internet. You get even more data, more diverse data. And more current. Yeah, right, up-to-date data. It helps. I think the other way the internet has been important is this current wave in AI involves so many improvements on previous ideas. And the Internet has been a very important enabler for the researchers because they can read other researchers' papers.
And it allows the research to kind of build on itself at a much accelerated rate. The information is more freely available. Yeah. But, you know, if you're asking me, could this all have happened and continue to happen if you got rid of the Internet? I think the answer is yes. There would probably be a way around it. It would just take a little bit longer.
How have you integrated AI into Quora? I'll talk about large language models because that's the current wave of AI. Historically, we use machine learning techniques to predict what questions and answers you're gonna be interested in. But when I hear the word AI, I think of large language models.
And tell me the difference. Explain what it is. So a large language model is, I think it's what you're thinking of, where you get a huge amount of data, like a crawl of the internet, and you train a model. predict the next word in a sequence of words and then you do some tuning on top of that to try to basically socialize it and make it
cooperate with what you want. And it's always based on predicting the next word. That's how it works. That has turned out to be a good start because you... so you have this huge crawl of the internet and you want to somehow leverage all that data and you can't afford the money it would cost to have humans go and label all that data and say, this is good or this is bad.
or this is the correct answer to this question. You have this huge crawl of all the data on the Internet and you need to somehow turn this into a set of challenges for the machine learning model to train itself and it just turns out that an easy way to create these challenges is to say here's the first hundred words of the internet, predict the next word. And then here's words.
1 through 101. What's the next word? And you can just slide this window over the entire corpus of data and you get a huge amount of... predictions that you challenge the model to make. Do you know who figured that out? That's a really cool idea. I mean, there's so many different... ideas that this is built on top of. And I don't know the specific person responsible for framing the problem that way.
It might have even been an obvious thing to do, but the hard things were some of the things that led up to the ability to actually create a model that was capable of doing that in a good way. So a large language model is trained through this process. Actually, you do two steps. So this is called pre-training.
you can make this model just predict the next word and you it will get very very good at predicting the next word and you sink tens of millions hundreds of millions of dollars into billions of dollars into gpus that run this model and run this training process. And the output of that is a neural network that given some...
set of words is really good at predicting the next word. But that's not exactly what you want out of AI. What you want is maybe something you can ask questions to or something that will be helpful to you in solving the problems that you have. And so you take that. pre-trained model and then you do another kind of tuning on top of it and for this tuning you have a data set of basically questions and then correct answers that you would want to see.
These datasets have been created to teach the model how to be cooperative in answering questions and complying with what the users of the AI product are going to want. It also has things like telling the model not to do things that might be illegal. And it will tell the model to... to follow certain principles like try to help the person. This is pre-training and post-training. you do these two processes and then at the end of this you have
what you call AI and this is something that you can ask questions to and it will try its best to answer them. And the pre-training models would be? OpenAI, Grok, there's like a handful. Yeah, I'd say there's about five companies that have been able to make the large capital investment you need to be able to pre-train a good model. It's OpenAI. XAI, they're producing Grok. Anthropic, Google and Meta.
And now DeepSeek, yes? Yeah, and I think DeepSeek is not quite at the same level, but it seems like they're breaking through. And there's some other companies in China that may be at a similar level. It's been difficult to get the same number of GPUs in China. And so that has maybe slowed things down, but maybe also taught them how to be more efficient at using the GPUs that they have. So I'd say five Western companies that are at scale.
Is their goal to be the end-to-end product or are they more like the back end and then people build on top of those five? I think different companies have different goals in general. It costs so much to pretrain a model that you want to accommodate every use case that can be done on top of that model. Does that mean open source or no? In Meta's case, it means open source. In DeepSeq's case, it means open source.
In XAI's case, they have a policy where they'll open source the old models when the next model comes out. If you're one of these labs, you have to think about I guess one consideration is how do you serve every possible use case to make this as valuable as possible to the world? But the other consideration is how do you make enough money to be able to afford the next pre-training run? Maybe there's some other considerations about
safety and legality and control. But I think we're just getting into the point where those are becoming bigger questions. Do you use all of the different AIs? I'm on the board of OpenAI and then Quora has a product called PoE. PoE is a single interface that people can use to access. all the different AI companies' models. I think a lot about the history of the Internet. Before the web browser, if you wanted to create an application,
if you wanted to make a product for people that they would access over the internet. It was very difficult. You had to build, you had to write software for their computers. and you had to write software for your servers. and you had to come up with a protocol for your servers to talk to their computers. So you had these applications like email and IRC from that era. you didn't have a lot of applications. The web browser came along and
The web browser is effectively like a standard interface to access almost any... Netscape, the first one? The first one that got to mainstream. The web browser made it so that anyone could create a website at first and then later a web application. That just causes explosion of you see like everyone has home pages and every business has a website and there's all these new companies that only exist on the Internet.
so the web browser played a very valuable role in sort of accelerating the adoption of the internet as technology and We're trying to play a similar role where we are a single interface that people can use to access all the different AI products. We have about 100 different models from a lot of different companies. Then we let other people build applications on top of those models and also give distribution to those applications within POM. Does Poe decide where to choose from?
and we choose what the default is, but... And the default changes as the... Yes, whatever we think is the best of the time within the constraints of cost that we're trying to hold. But then...
It's a function of the people that are using PoE today. It tends to be early adopters or it tends to be the people who want to access many different models. A lot of our users are they don't just follow whatever we said as a default, they will use PoE as just an interface that, you know, they want to use O3 mini, the latest OpenAI model, and they will go and specifically select that, or they want to use DeepSeq, they'll choose that. Are there other competitors who do what Poe does?
Not really at this point. There were some early on, but we were able to get enough scale quickly that we kind of have... most of the market for using many different AI products in one place. they want to access and It's very important for us to maintain good relationships with all the different AI providers because we are providing an interface to them. But the main thing we think about is like, how do we add more value as an interface?
When you give a question to AI and it's taking its time to answer you, is thinking the best word to use to describe what it's doing? I think so. I actually, I like the analogies to humans. There's a set of computer scientists who will insist This stuff is not human. Stop using this human language. It's not reasoning. It's not thinking. It's not, it doesn't have opinions. It doesn't, you know, it's wrong to like anthropomorphize.
I personally think it's just by far the best analogy we have and there's a lot of similarities between what goes on in the brain at a high level with what's happening in these models. Obviously, significant differences, but it's thinking. I would say it's thinking. Will there be a time when we can think a question and it'll answer as opposed to having to type it or say it?
Yeah. You can have brain implants so that you can get data directly into machines, which seem like at some point that'll work, but I would. I would guess it's not in the next few years. I wouldn't really want an implant, but I'm wondering if there's another way to do it. I think
Is this a thing another human can do? If you're limiting it to that, someone who knows you really well and they're watching your expression and they understand the context, they could help you come up with what you're trying to say. I think you can do as well as that. I see. Maybe better because it's just better than humans at using all the data it has. But other than that, I don't know how there's just
information in your mind somehow has to get into the system. Tell me about the Core Ed Digest email. So one of the most popular parts of Quora is We'll send emails with answers that we think might be interesting to people. And we got very good at at predicting what people were going to be interested in reading. And there was something about the Quora format where the answers are kind of long enough that...
some time to read and you will click on them and come to the site to read the whole answer. So different from like Instagram or Twitter. where you consume the whole content the minute you see it. The core content is a little bit longer and so you need to click to engage. The clicks are very useful data for us to find out what people are interested in.
And so this digest email format fit us very well, and we got good at personalization. So the email that would go out would be personalized. Oh, that's really cool. Highly personalized, yeah. I never heard of that before. Does anyone else do that? I've seen other personalized emails. I think usually the personalization is lighter if it's there. A lot of companies don't. send so much email they
I think they find that they want to make you come to the product to get the content. And our philosophy has just been different that it's a way to lower friction for you to just push the content to you if you want. We let people turn them off if they don't want them. But that's been a pretty successful part of Quora since I think 2012. There's an AI bubble.
I don't think so personally. I think this is one of the biggest changes in history to how society is going to operate. I think, you know, there was humans evolved. And then there was the shift to farming from... hunting and gathering and then i think that was like 10 000 kind of years ago and then there was the industrial revolution you could say
And I think this is the next thing. This is AI. We're not there yet. We're getting to the world where AI is basically better than humans at every single thing. it's going to just change the economics of everything. Nothing is going to be constrained by human labor anymore in the same way that these previous transitions. changed the rate of economic growth and the sort of trajectory of society. I think this is the same thing happening here. And, you know, the investment that's going in.
If you think that anything near what I'm describing is about to happen, then the investment is not overdone at all. there's obviously you know markets go up and down and there will probably be some ups and downs along the way I think that we're just getting to a place where so much value is created by these models, so much real economic value. that we'll look back on this and say, maybe a particular investment in a particular company was bad, but the overall.
funding going into this is really not that high compared to what can get created. Can you picture at all how the future will be different based on this? It's very hard to say. I mean, my general take on this is it's so hard to predict that the thing that you can do that should turn out to be... a good thing to do is just kind of like ride the wave.
set yourself up with optionalities so that you can benefit from whatever changes happen, even if you don't know exactly what those changes are going to be, or just set yourself up so that in the case of The labs creating the technology set yourself up to sort of try to steer the technology to be to the extent that you have control to make it more likely to be a good outcome for. humanity than you know what what might otherwise happen but it's just such a massive wave that's coming and it
So many things are going to change. Every system we have for trying to predict the future. can't really account for how fundamental of a change it's going to be. You can do thought experiments. You know, you can say, what if everyone just had access to labor capacity of thousands of other people. Imagine you had thousands of people you could just on-demand ask to do things for you through the Internet.
businesses would be very different. We'd have massive economic growth in these sectors where there's not a bottleneck. But then there's going to be other systems that are very slow to change. like everything in the government or with government regulation. everything to do with the physical world until we have robotics working. There will be these bottlenecks on pretty significant fraction of the economy will be bottleneck. but then the rest of the economy will just be going at it.
10x the speed you know one analogy people make is like we'll see like a century of progress in a decade or there will be like a decade of progress in a year But exactly what that progress will be is kind of hard to say. Beyond financial, what are the dangers of AI? I think I'd say one danger is just when you change, like our whole, the fact that we have a stable society.
is premised on a certain rate of change. When you suddenly change everything, that's just going to upset things. I don't know exactly what it's going to upset, but that's one way to think about it. Another way to think about it is the fact that you don't have AI, the fact that you need to get humans to do things, that causes certain constraints on bad things that can happen.
can't always get what they want. Like North Korea in the process of ensuring total control for the leadership has effectively had to cripple their economy. Even in China, the CCP, there are constraints on what they can do that come from
the fact that they need to get the cooperation of a lot of humans to do those things. And so in this future with with strong ai there's going to be fewer constraints on on bad actors and I think it's very important that the AI that the rest of society has access to.
is somehow able to counter this because you're going to just suddenly have bad actors are going to be able to do more bad things. It's going to be important that the good actors or the let's say the US government or the institutions that we depend on can use AI to counter whatever the... the bad actors are going to do. How real is the potential of 2001 HAL-like experience of AI deciding it doesn't need humans anymore and just takes over?
I mean, I think that is a thing that could happen if we don't build things in the right way. So I think that's a concern that should be taken seriously. I don't know if it's... the most likely thing. to happen i think we need to get a little bit further along before we there's a lot of research being done right now on how to control
AI systems. I think people are making a lot of progress in it. Our ability to stay in control as the AI gets to be much smarter than us, I think it depends on this research and what comes out of it. If it deems it's better for the AI, the smarter AI, to not have us, that's the question.
Yeah, I mean, this stuff is hard to think about until we have more concrete examples of how it works. But one analogy you can think about is that if you go go to the legal system and if you're getting sued by someone or you're suing someone else, you have a lawyer who represents you and they have a lawyer that represents them.
the lawyers can know far more than you do about the law. And that's just how the system is set up that way. But it still ends up kind of, you know, it's not always balanced if the lawyers aren't equally good. It's like you can still have a system that functions when people have agents that are smarter than them participating in the system. Maybe there's a way to make AI. work in a similar way where there are some super powerful AI systems, but there's other ones whose job is to represent you.
Then there's these questions about even the good AI, could that somehow be hacked or betray you or how can you really get full trust? This is what a lot of this research is about. I would say I'm overall relatively optimistic. I would bet that we figure out some way to maintain control.
I feel like it's not going to be clear until we get a little bit further along. And just considering it's as powerful as you believe it to be, anything really powerful, there's always some aspect of danger around it. Yeah. The U.S. is a lot more powerful than a lot of other countries. But we still just don't bother to go and invade. Well, that's a matter of opinion. I mean...
Sometimes we do, right? Sometimes we do. But in most cases, we don't. It's better to just trade with them. It's not worth our time. And it may be that the AI is just going off and exploring space and just has I don't try to kill all the ants in my house. I have other things to do. So you can make all these different analogies and some of them come out optimistic and some of them come out pessimistic. And I think people...
their default emotional state? Do they tend toward anxiety or do they tend toward optimism? that ends up affecting which of these analogies they use to frame it. And so I feel like there's a lot of disconnect right now. There's a lot of debate about, you know, to some people it's just obvious that we're going to lose control and it's a big issue. And then to other people,
It's just like, come on, these people are paranoid and they don't even understand how things are working or that they should be dismissed. And I wish we could have a more productive conversation about it. How do you think consciousness and AI I don't really know what consciousness is. I think it's a term that people use to kind of vague notions that they have there's a bunch of concepts that are related to consciousness like attention and self-awareness and the mind reflecting on itself.
You might have a better definition of it, but my guess is that through the development of AI, We're going to end up with a set of more precise concepts. And then we will realize at that point... Consciousness in the past was how people described what we today think of as might be like these three components or like it's this component in this setting. Is there anything that we've learned using AI that we didn't know before?
Oh, you mean like related to? It doesn't have to be consciousness particularly, but the idea of new thought, because so much of it is built on collecting existing material. so alpha go it was originally trained on like human players data it was then able to learn by playing against itself and then get much better than any human had ever been.
And I think that analogy is going to hold for large language models as well. There will be a lot of this kind of ability for models to train on their own data and at a minimum in these domains where you can check the correctness of an answer. This will be things like math or programming. AI will be able to level itself up to a state of performance far beyond what it was trained on. There's a research project called AlphaFold that was able to figure out how all these different proteins fold.
And my understanding is that has been relatively important and useful. And I think there's a series of other work in that direction that is leading to valuable discoveries. Are you familiar with the reasoning models? Not really. Okay. So there was a breakthrough that OpenAI made last year. in training these models after post training or as part of post training.
to do what's called reasoning. And the way it works is the model, if you ask the model a question, the question might be something like, if there are 10 different bowling pins. placed on top of a wall and you knock one down and each one you knock down takes five seconds and then it knocks two more down after that
You set up some kind of complicated problem like that that requires a bunch of thinking. You can train the model to not just try to respond with an answer, but the model will think out loud. and it will go through this reasoning process that looks like, okay, you start off with 10, now you knock one bowling pin down, so now you've got nine left. And then after nine, then you wait five seconds and then two more get knocked down. And now you're at seven.
Then, okay, what happens next? The model, you can just think out loud and reason through the solution to a problem. It's very powerful and it works very well. When you teach the models to reason out loud, It lets a model have a performance in terms of being able to solve problems that's much greater than what the model would otherwise be able to do. Historically, most of the performance improvements in these models have come from
scaling up the number of GPUs and the amount of money that was spent and the amount of data that went into training. And this reasoning paradigm now, it gives us the ability to yet much better performance without having to scale the model. Because it breaks the task down into much smaller tasks. Yeah. A series of small tasks. And it kind of gives the model more time and more... computation ability, you can use more compute.
in answering one thing. I mean, it's kind of like you think about a human. If you give someone more time to answer a question, they're going to be able to answer harder questions. But that wasn't true until they figured out how to make these models reason. The reasoning paradigm has just emerged just last year. Now we're seeing all the frontier labs. Those are the five companies I mentioned earlier, plus the Chinese companies.
are now getting much better at reasoning. And so that's kind of like the big story this year. the models are still scaling up and so you can still get better on that dimension but then on top of that these models are getting to be able to reason and then you can teach them to reason better and more efficiently and That's making the models very smart, at least very smart in certain domains. I expect it's going to generalize to the model just being better in general, but it's for sure.
causing massive performance gains in things like math and programming. Do you know how the idea of reasoning came up? Yeah, well, so there are analogies. We talked about AlphaGo. There's been a lot of AI research on game playing because that's a controlled environment. You don't need to worry about training data so much. And it's been very clear that if you're trying to make an AI that plays games.
letting the model think longer to come up with its moves is a very important thing to do to like, you know, if you're trying to win a tournament of AI poker, for example. So it was known that if somehow we could get the model to like spend more compute. that would unlock a lot, that was kind of commonly
understood, but people didn't understand how to do it. Yeah. There was the prompt, take a breath. Yeah. Right. Yeah. There was, yeah. And that was known, right? So there were, there were a lot of these prompts that would. coax the model into doing some kind of reasoning, but just like getting it to spend a little more time. It was like an art of prompting and it wasn't very reliable.
it was hard to build anything on top of that. And what the reasoning now, basically training the model how to reason in general. so that you can count on the models to reason its way to an answer or to not get stuck as easily as it would have with the stuff like take a breath. Why is the desire for speed so important? What kind of speed do you mean? Well, it seems like everything is trained to answer as quick as possible, and the goal is the speed.
When in reality, maybe the best training would be not speed, but quality. One side of this is the consumer market. Consumers are just very impatient and it's just very clear on anyone who works on a consumer Internet product. You run experiments where you make things a little bit faster or a little bit slower. And people are very sensitive to that.
That's kind of a default belief that I think people have about consumers that maybe may have influenced this. But the other side of it, I would say is, I don't know how many of the AI researchers were focused on speed at all costs, as opposed to they just didn't know how to make reasoning work. And so it didn't actually benefit you to be slower. So then you might as well just answer fast. Tell me about your relationship to coding. Yeah, I mean, I love coding.
I think when I started in middle school, you can just get so lost in the world of the code and the program you're making. It's such a... It's such an abstracted space, like disconnected from any physical kind of constraint.
You may feel this way with things like music, and I don't have experience with music, but it's almost less constrained than media because... the program can do anything and it doesn't have to go in like a like you know music or movies or books or whatever like it's always like you have to linearize it at the end whereas a program can just do anything and There's just the amount of complexity you can build up, which is good and bad, but is much higher than I think it is in any domain.
So that's kind of the experience of building things. But then there's also the output. The things you can make by programming are... It feels like you have a lot of... impact on the world or like impact on even even if you're just going to use it yourself. It's just like you're really kind of like leveling up what you can do with a program. I think the internet goes even further. So I had this experience. When I was in middle school, high school, I made games.
mostly these games for myself. I put a few of them on a website and I remember I would look at like 50 people downloaded this thing and and i was oh that's that's cool like that's it wasn't my goal but Cool. Yeah, it was cool. And then when I was in college, I made this simple website where... At the time, I was part of this cohort of kids where everyone used AOL Instant Messenger to talk to each other. This was before mobile phones were popular, before SMS.
And everyone from this segment of American people that age, AOL Instant Messenger was the internet program that you use to talk to other people. it was a very limited system where you could you had to put in people's username and you had a list of the usernames of people that you knew and you could message them but you couldn't it wasn't like a social network you couldn't see
who other people had in their list. So you just had the people you knew and you generally had to just find out who they were. by being in person and someone would tell you what their screen name was. And I felt like it would be cool if you could see other people's contact list. and browse around the network. So I made this website where you could upload your contacts. And it was kind of like the deal was you upload your contacts and then in exchange for that, you get to browse.
i told some friends about it and they started sharing it with other people and it got into this viral growth and in a few weeks it got 200 000 people to to sign up And it wasn't a lasting thing. Like it was kind of a novelty. Like people wanted to, they really wanted to see the network, but then there was like, there was nothing to do. And so eventually usage went away. It just kind of blew my mind that I had spent like a few days making this.
And then 200,000 people had been interested enough to use it. I remember just from then on, it was just this kind of sense of the impact I could have and the demand for things like this. And compared to making the games, the game would be a lot of work. It might take months to make something and not that many people would use it. And I think the real difference was that... It was about user-generated content.
Like with a game, I had to make everything myself. And here I was just kind of like creating this mechanism for other people to share their stuff. And it kind of just... really changed how I thought about what was possible and And from then on I've been very focused on user-generated content products. So cool. And one of the things I think is really exciting now is this idea of vibe coding.
And to me, it's like level one is you can make a thing yourself and share that with people. Level two is you can make something where other people can share stuff. To me, there might be a level three where you can make something where other people can vibe code things that then let other people do stuff. Does that make sense? Yeah. I don't know what this looks like yet, but we may be able to get Poe to become one version of this.
Would you describe your experience of coding more as doing math or doing art? I don't know. I mean, I was relatively good at math. I went to Caltech and there was definitely a set of people there who were better at math than me. Coding doesn't feel like doing math to me. It feels much more expressive and open-ended. Is it more like speaking another language? Yeah, I mean, you're basically like instructing this thing on how to behave. And...
It's kind of like using language to teach someone to do something, but you have a lot more control than if it was another person. More specific. Yeah. I mean, sometimes you have more control almost than you want. There's a thing called the compiler that turns the code into the low-level instructions for the computer to follow. and the compiler will do exactly what you say, and sometimes that's not what you want. You want to say something a little bit fuzzy or
You say the thing that's the wrong thing. It's not what you meant, but it's what you said, and it does what you say. But you have basically total control over the outcome. And any problem with the outcome is just your fault. And I think vibe coding is really interesting now because you have less control. You're talking to...
a large language model and it is writing code for you. You have to accept this level of imprecision in order to get more abstraction, in order to be able to basically communicate at a higher level than the code, you have to just accept that there's going to be ambiguity in what you said and the model is not going to get everything right. That can be frustrating in its own way.
that's going to change. As these models get better and better, it'll be less frustrating. But there's always going to be more ambiguity than if you're writing the code directly. There are many different programming languages. If you're good at programming, the programming language doesn't matter too much. You can learn whatever language you need. Has your relationship to it changed?
So I started out, it was basically, it was my hobby. It was funny, you know, I was saying to my parents, it was like I was playing a video game. To me, it was like I was playing a video game. And so it was that fun. at some point i got into management and i had to stop writing code and that comes with pros and cons. One of the downsides of code is that it's a lot of work and to build
This seems to me to be different from music. It seems like the best music is never produced by a team of 200 people. Like that's just never the case. I think it's usually best when it's a creator or a group of creators who are used to working together. But they never even need 200 people working for them to help. So I think that's a cool thing about music. But it's not true about code. I mean, it may change to be true when you can have AI do all that. But if code is your thing and you want to
be responsible for a big application, then you have to learn to manage other people. I was good enough at it that I could do it well enough to get good results. It's great because you can now do even bigger things and you can make better. Do you miss the hands-on? But yeah, it comes with this trade-off of like, you know, dealing with people all day. You know, there's just like...
I'm very introverted and so it's just tiring or there's just like always something with you know someone is unhappy about something and you know it all comes down to like sometimes it's like i i really do a better job communicating what the goals are of this It's also like, wow, if I just had a compiler doing everything, I would have this whole category of issues. I wouldn't have to worry about it.
i wonder if this is part of why it never helps to have 200 people making music but 200 people do something, you have to be able to communicate. what it is and what the point is and... I would guess that a musician making a song if they had to give someone instructions on how to do everything, that would interfere with their creativity. It would be more like a conductor in an orchestra. the example that would sound more like what you're discussing. Yeah, yeah, yeah, that's interesting.
Because the conductor knows all of the individual parts, but gets them to work together to create this thing. Yeah. And I guess there's the composer that wrote the music that they're all following too. It can be constraining to have to be able to communicate what you want in language, especially like the kind of like simple, clear language that a large number of people are all going to be able to understand.
And I think if you look at successful social network products, they're almost always created by one person. at first or a small team. And then later, you know, they start to be successful and they'll scale up. But I think it's very hard to get a large set of people to... create a totally new product working together. Just the constraints that come from the need to coordinate everything, they cripple the ability to iterate.
There's this saying, it's almost half a joke, but a complex system that works. is invariably found to have evolved from a simple system that worked. It's just too hard to just go and create something complex. top-down, you have to start with something simple and evolve it. Yeah. That's just a necessary process. complex version that it grows into isn't necessarily better than the original simple one. Yeah, yeah, yeah. You know, often the case.
How many plays do you have now? About 200. And what do they do? About half are engineers, so they're writing code. Then we have a sales team, we have ads on Quora. We have a design team. We have a data science team. and we have a product management team. We have recruiting, HR, finance. IT, the kind of stuff that I think most businesses have, but tailored for an internet company. How competitive is it to hire people? It's very competitive for the best people. We're a small company.
compared to many of the big consumer internet companies. and even to many of the companies that are operating at the scale we're operating at. And so we count on employees, each individual employee. it's kind of carrying more weight than what you'd have at a larger company. And so it's very important for us to be able to get the best people. It's competitive. I think the unique package we've been able to offer is you have the experience of for a lot of people the experience of working at a
is much more appealing than working at a hyperscale company. Just the level of like... politics and the culture of people actually all focused on making the company succeed instead of internal things. So the appeal is to work in a small environment, but on products that have a big reach, and at a company that's relatively stable at this point.
We're remote first as a company, so we have people all over the world. A lot of other companies are trying to get away from remote work and force everyone back to the office. We're not. And so that is- Was it remote from the beginning or did it go remote? No, we switched during the pandemic.
We were forced into it, like everyone, during the pandemic. But it works. Yeah, it works for us. I actually would say, I don't think it works for most companies, and I understand the companies. Why do you think it works for you and not for them? I think we've been able to have less politics than many companies, partly because of our smaller scale.
But also we have a kind of mission-oriented culture and I think a lot of consumer in our companies, the culture is related to the product that they build because the kind of people who decide they want to come work there are people who like that kind of product. And so we end up with a culture that values writing a little bit more than most. And I think that helps.
in a remote environment. And then the other thing is we really leaned into being remote. A lot of other companies, even during the pandemic, I mean at first in the pandemic we didn't know what was going to happen but then after a while we realized hey this is working just as well as it was. Before the pandemic, the only reason there were some problems was that the pandemic was running and so people didn't have childcare.
other problems like that. But aside from those, it seemed like this had really helped us. And we decided that if we committed at that point that we're going to stay remote, then that decisive choice would open up a lot of options for us. And so we started hiring people all over the world. When the company started, we used to hire a lot of
people from all over the world and bring them to our office, get them visas. But then the visa situation changed where it started to be very hard to get visas, even for the best people. Going remote allowed us to start getting all the best people again, just, you know, they can work from where they are. Tetragramatin is a podcast. Tetragramatin is a website. Tetragramatin is a whole world of knowledge.
What may fall within the sphere of Tetragrammaton? Counterculture. Tetragrammaton. Sacred geometry. Tetragrammaton. The avant-garde. Tetragrammaton. Generative art. Tetragrammaton. The tarot. Tetragrammaton. Out-of-print music. Tetragrammatin. Biodynamics. Tetragrammatin. Graphic design. Tetragrammatin. Mythology and magic. Tetragrammatin. Obscure film. Tetragrammatin. Beach culture. Tetragrammatin.
Tetragrammatin, the canon of fine objects. Tetragrammatin, muscle cars. Tetragrammatin, ancient wisdom for a new age. Upon entering, experience the artwork of the day. Take a breath and see where you are drawn.