This podcast is brought to you by the Albany public library main branch and the generosity of listeners like you. God daddy these people talk as much as you do! Razib Khan’s Unsupervised Learning
Hey, everybody. I Hope you guys are doing well thank you for joining me on this episode of the Unsupervised Learning podcast. Today I'm here with my friend and engineer an erstwhile Twitter engineer, Nikolai Yakovenko. And we are going to be talking about Twitter, social media technology, etc, etc. Obviously, as I'm recording this, there have been some changes in the ownership. But um, you know, we will get to that. But we will talk about, you know, what his experiences working at Twitter were, you know what he knows that maybe we would like to know, hopefully, but Nikolai, do you want to introduce yourself with anything else?
Yeah, sure.
Hi, Razib, Yeah, I'm a machine learning engineer and applied machine learning scientist, kind of, like stumbled into it out of out of college. You know, my, my first job was working on Google Search, which is, you know, a lot of fun and like, you know, very gratifying field to make tiny, tiny little changes to the search function that will be used by millions of people. So I got pretty spoiled there. You know, not going to go through my whole tech career, but kind of out of that got into applied machine learning for various different problems, and ended up some years later, at Twitter, you know, on the applied Deep Learning Team, a Team is still existence called Cortex, although, you know, the founding most of the founding people have left, which is part of the story. Yeah. And then finally, now, I have a startup working in machine learning for the crypto space specifically for estimating, you know, basically, we call it price Oracle's so, you know, historical and current value for NF T's right? There's, there's a large number of individual items that are not like liquid, like a Bitcoin or Aetherium would be so estimating what they're worth in the market. So you can sort of score your portfolio and negotiate trades, uses it as collateral, all kinds of fun things like that. So that's what I’m up to these days.
Wait, what's your startup called again?
it's called DeepNFTvalue. You can check it out, we have a public facing thing, and we're looking for people to join. So if you're interested in crypto blockchain and have at least some interest in NFTs we'd love to talk to you as well
Well, I mean, I mean, because you already raised before the crypto winter, right?
Yeah. I mean,
I guess you could call it a winter, definitely a crypto fall. Yeah, no, we raised a $4 million round, closed this late this summer. And, you know, we haven't spent very much of that yet. So that's good.
Well so before I go further, you know, I think maybe you should tell the listener. I mean, a lot of listeners already know, but some of them do not. What machine learning is, and then what deep learning is.
Totally. Yeah, absolutely. So so when I when I went to work for Google, like, I mean, obviously, machine learning as a concept existed, but it wasn't even something that we talked about. But you know, like, basically, the idea is look at something like Google search, or where then the next great machine learning product that we all used and sort of loved or not, which is the Facebook and then later the Instagram timeline. So that's really, you're using machines to sort of, at a high level, make decisions that a human might do, or to sort of solve judgment level tasks, essentially, automatically in real time. So I remember, you know, the head of Search, when I went when I was at Google, you know, he had a way of telling the story, right? He's like, Well, you know, when you type a query into Google, you're looking for something, if you had an expert in the field, who had infinite time and space and could sort of like go through the websites and rank them and give you the top 10, then that theoretical person would do significantly better than Google search. But what we're trying to do, you know, in our 300 person team, then right was do that very fast, automatically approximate what that person would do fast automatically, and also in a way that's like structured and reliable, right? So you can kind of predict what the machine is going to do. Whereas a person would kind of, you know, make more random decisions sometimes. So the best way you think about it, I mean, you could think about is anything from predictive modeling, you know, to things like well, like pricing hotel rooms, you know, like was what is demand going to be to like, obviously, like one of my favorite sort of, you know, applied machine learning problems in the wild is like, you know, it's like rental car stuff, right, managing rental car fleets, and pricing you automatically if you're going to drive from Vegas to LA, sometimes one way will be cheaper. Sometimes another way will be cheaper because they sort of know where they need cars - what they expect the demand to be. So these are all like problems that, you know, in the past were solve through human judgment, maybe are solved through some sort of like formula an if/then, but eventually they switch over to machine learning, which means you sort of look at the past, you look at the features, and you try to predict what would be the best decision for the, for the future for the near future. And in that way, you've asked about deep learning. So deep learning is a is a technology. There really, the idea is to, is to train these like very large models that have a huge number of parameters, and to really focus on like the structure of the network, to sort of generalize from experience. So like, the first problem, not really the first or the first, like popular problem to be really, like crushed by machine learning by deep learning, as opposed to previous things is, is basically identifying cats and dogs in pictures, like putting a box around it, if some of you use Facebook, I guess a lot of us don't anymore. But you know, remember was the first thing was cool identified your friends automatically, you would draw like a box around their picture, you just have to confirm, and it was good. It went, you know, it was getting it right. You know, basically every time pretty quickly, you know, whereas I think I remember when Microsoft and Apple had similar technology, you know, five to seven years earlier, and it was right, like 80% of the time, which is frankly not acceptable. You know, like for these things to work, they really need to be quite accurate, and really specific. So deep learning is kind of a technology. So yeah.
Well, I mean, so basically, you know, I think one way that I like to think of it is it's doing it scaling up and doing things that humans could do, right. I mean, in terms of, we take information training sets, we see patterns, and then we detect the patterns, and we make an inference. And you know, with machine learning, you write some code that processes these data inputs coming in and extracts out insights, patterns of structures in the data. And then things like, you know, generate conclusions from the inputs that are coming in later, beyond the training set, right? Is that I want to just think about it.
Yes, so you're
describing something called generalization, which is that you have past data, and you train the model on past data. And you sort of expect and hope that given completely new inputs that are a little bit different, or very different from what you've seen in the past, you'll still get reasonable answers to your point. One small sort of point that I would make though is that, yeah, two things. So one is yes - basically machine learning, especially with deep learning, which is really great technology, I would say, at this point, basically, if there's, if there's any tasks that you could, that you could teach to a human, like a cognitive task, putting aside robots and physical tasks, but any sort of cognitive task that you could, you could teach a person, you know, within a few weeks, you know, who's not a super duper, duper domain expert, we're not talking about, you know, surgeons with, you know, 20 years of experience, that is something that I think you would expect a machine could also be taught to do at a similar level. Now, the issue you're going to have is dealing with sort of, like out of sample, right, like, like an extreme situation that didn't happen in training. You know, for example, on like the Tesla, self driving cars, people would post these issues where it would have really strange responses to emergency vehicles, or snow or something like that it wasn't trained on while it would dry beautifully, sort of, you know, sunshine, right? With normal conditions.
Yeah. So I've heard with the self driving cars, and also sometimes with robots, that have to, you know, respond to novel conditions. Sometimes, they will just do crazy things that make no sense. And so,
yeah, it's out of sample, it's out of sample, right. Like, it just hasn't had any experience. And you as a person have, you can switch to like a certain level of judgment, right. And then also, to be fair to the robots, the roads were the roads were designed for humans, they were not designed for self driving cars. So we're like, expect, like, if the if the if the roads were designed for self driving cars, you know, then they would be much better than humans by far already, but they're not.
Well, I mean, like, in some ways, I wonder if this is just random, but um, like trains or airplanes are much better cases than self driving cars, because they're pretty demarcated and specific and they're already, like, heavily automated anyway.
Oh, totally. Yeah, I think a friend of mine was involved actually on the security side for the train system in Switzerland, which is, I think, one of the best train systems in the world. And it doesn't have any conductors. It doesn't have everything, everything is completely automated. With the exception of like, there's edge cases, that it just easier for humans to remotely operate than to design a computer system to do it. And the one he noted is like, he's like shoveling the snow. If there's like snow in the tracks, the train actually As a machine that can do stuff to clear the snow. But instead of having it like automatically clear the snow and use computer vision to figure out where the snow drifts are it just easier for humans to just wire around and remote control it, you know, but apparently, that's the only thing that humans do it all on the on the Swiss trains. But But to your point, that's a controlled environment, right?
Yeah, and Twitter is not and, you know, I'm, we're talking to you, we want to talk about social media, although I, you know, do appreciate just kind of like the the quick verbal primer there about machine learning, because I'm sure it's a word that a lot of people have heard. But it sounds quite exotic, but then when you break it down, I mean, some people, I think this is overly reductive, but you know, some people are like, Oh, you're just talking about regression, you know, like regression analysis. So, you know, there's a bunch of different techniques that are bracketed under machine learning these, like, formalizations of taking an input data, and just making sense out of it, right? I mean, that's the simple way to say it. So with Twitter. I mean, what are you guys doing? Can you identify? I mean, I'm assuming people want to identify what a troll is, can you automatically identify a troll?
Yeah, no, that's, that's kind of frustrating, because yeah absolutely - the, to the to sort of - to defense, Twitter a bit, you know, some combination of sort of things, I believe, and also devil's advocate, I think it's interesting to hear that side, I think, I think I'll be saying that throughout this conversation, in many ways. I think people don't have a lot of experience dealing with spam. So the issue of dealing with spam, and I learned this from Google, is, you really, you really have like, massive imbalances. You know, like, you look at the piece of spam that you see. And it's like, OMG, this is easy to identify, and in many ways, you're right. And in many ways, these, these companies should do a better job. They're actually those critiques are correct. The issue is that if you kick in an automated system, that you know, kicks out 1000 pieces of spam, and then one out of 1000, they like ban like a real user. You know, there's kind of an imbalance there, right? Because the computers can just create, you know, 10,000 times more messages than humans, right. So the real issue is at a high level is you really want to be careful not to - not to ban real people. Because you're kind of wanting new people to sign up to Twitter, or to Facebook, or whatever. And, you know, even something like two to 1. 10 to one spam to human is actually not the right ratio, if that's what you know, you're doing. And then on top of that, you don't know what you're doing, it's actually very hard to get feedback, right. So now it can be done, like Google, for example, has basically no spam. Like, when you search in Google, you will never see spam. That wasn't always the case. And it's not the case for new things. And any new service as soon as it explodes. You get like inundated with spam, I think a kind of a famous case that is well known in Silicon Valley is when Instagram was first bought by Facebook for the billion dollars, they kept running their own team and their own scaling their own stuff with Facebook support, but the one thing that they immediately turned on was like to Facebook, you know, like spam detection, you know, and that's ironic, given how much spam there still isn't Instagram, but that's because people are really incentivized to do it. Right. So these are difficult problems. I mean, I think part of it is a cost thing. Obviously, anytime you're, you're doing any security or reducing spam, it's not directly making you money. So like, you're not going to spend 80% of your revenue on something that is a, that's basically like a safety measure, that just doesn't make any sense. So it's always it's inherently like, in just a game theory of it, it's always going to be something that's a little bit neglected, a little bit underfunded, a little bit less exciting for the best engineers or best engineering managers to hitch their train on to because it has nothing really to do with growth or making money or the the cool stuff. But I think part of it is, the main thing people really don't sort of don't appreciate is again, that ratio that the computer that the computers can basically make infinite amounts of spam. And you really need to, like, if you check in a change, you're not really looking at the ratio of how much you're good or bad stuff, you're banning to how much good stuff you're sort of banning accidentally, you're really looking at your absolute amount of good stuff you're banning. And if you're banning real people, it's a real problem. You know,
yeah, I, I this hasn't happened lately. I don't know why. Maybe I'm not paying attention. Or maybe it's gotten better. But I've heard from people that I know who've been identified as bots, and you know, yeah, Twitter given was horrible. Yeah, that's like, I mean, but the reaction is, what am I doing that's like a bot. Not me - it’s never happen to me but,
You know, Razib - and then going back to the machine learning conversation we just had an A minute ago, like, those cases really can be bizarre, right? Because, you know, again, these machine learning systems are not necessarily very good out of sample, you know, if you do something novel, so think of something else. I'm gonna try not have too many tangents but like the way this was handled before good machine learning was something called like an anomaly system, right? You used to have, how would you like detect like an intruder in your system, you have anomalies like they're like this is normal activity. And then anything that's like super weird would just like trigger and shut off people's accounts. So that could be an intruder, or that could be spam. But it could also be like, maybe people don't log in and compile, like, maybe you're working at a biotech lab, they don't use some like, like very legitimate, like coding systems, and then those things would immediately be flagged and cut off. Right. So yeah, it's machines, machines, don't think like people at all. And then also, like, yeah, so some of the things they identify as a spam wouldn't make any sense to us at all whatsoever.
Mhm hmm. Yeah. I mean, so in terms of, you know, identifying spam, identifying trolls, if these sorts of things are a big deal, but, you know, from what I understand, Twitter has an enormous amount of human content moderation. Are they the last guardrail? Like, you know, you haven't been there in a while. But do you have any sense of what the ratio here is of? Because like, so for example, you know, Gmail, most of the Spam is filtered out a little bit gets through, you know, I do wonder sometimes, if there's stuff that doesn't come through, that's false identified, but I very rarely get messages about that. It's there are some cases, you know, some very few cases, like, actually one happened this week, someone sent me an attachment, a PDF attachment of their book. And they had never emailed me before. So I suspect, what Google thought was, this is a new email address. And there's a large PDF, there's a large file attached to it. What does that mean? Right. And there wasn't any text associated with it, it was just the title of his book. So it said it was spam. And I had to go into the spam filter and get it. But I mean, in terms of the Triage, or the, you know, the division of tasks. Why? Why are the all these human content moderators still around?
Oh, great. So actually, let me touch on the Google spam thing for a second, because it's a very interesting example. I'm glad you brought it up, like, so first of all, I'm going to disagree slightly, I think that Google… a lot of spam does get through and then much worse than that, there's a ton of real content that gets caught in the spam filter. So you brought up a great example of individual stuff. But like, if you're sending, you know, anybody who sends like, you know, email lists and things like that, they have this issue constantly, right? Like, also, if you use certain words, it just gonna get picked up as spam. I mean, imagine if someone is to send you a PDF about Viagra. I mean, there's no chance that is getting through. Like that's gonna get crushed by the spam filter.
Yeah, that's fair. That's fair. Things I've heard, I've heard of things like that, where people used a spam word is literally what people said, I think,
yeah, and you, and you cannot get that through. And it doesn't matter. And it's very frustrating, because because there's nothing you can really do about it, like Google doesn't care. They're not going to fix it, they're not going to change it. Like from their perspective. And again, this has to go to what I was saying before, it's not a very good from a high level, because enough people don't care. Like it's they're not going to spend 10x More money on spam detection. They're just not, you know, and yeah, it's just, it's always gonna be like this, essentially, in my opinion. But But going to the Twitter thing. So first of all, I mean, I'm not sure. Google really has that much content moderation. Actually don't think that team is at large, it definitely isn't Facebook. So Facebook has a huge team, doing content moderation, and clicking and checking stuff. I think, my perspective, my, my perspective, was a Twitter team was always a bit underfunded on that, actually. And I would imagine that may get even more so now. So you know, I actually had friends in engineering who worked on the spam filters. And going back to that for a second also, I mean, they're constantly like writing like new rules, right? So for example, in crypto, you will get these very specific spams over and over and over. And it like makes sense to write that rule that like, this should be a feature and throw that into the regression, you know, but then these rules get stale. People work around them, they accidentally pick up other things. Like there's not even a way to do it very cleanly, you know? So, yeah, I mean, I think, I'm not sure how Twitter does its content review. I mean, I guess it's good that they get back to people but. My, my perspective is actually more that Facebook is the one with a giant team. And Twitter is more like triaging and trying to keep the lights on, you know,
what, so can you talk really quickly, you know, you're talking about having to update things and, you know, the system's getting stale. A friend of mine, who works in machine learning said one thing that people don't understand, and I don't, I've used machine learning, obviously, I think we all have in the compute space that analyze data at this point, but you know, I don't have enough experience to judge this. He said, You know, one thing people don't understand machine learning is it requires a lot of human intelligence to use it properly. Right. So what do you think of that assertion?
Yes, I think that's true. And I think that's specifically true in the case of spam, because, again, the way you measure and train things, as you want to say, these are positive examples, these are negative examples, right? Like, can you get to a classifier, people will say ridiculous numbers like 99% accuracy, and like that, that's not even how we measure it. I mean, you know, this from the genetic space Razib, right, like, you know, if you're trying to call certain variants, you know, like, some things are easy, some things are hard. So you sort of look on the curve, doesn't really matter outside the scope here. But the point is, fundamentally, when you're training models and measuring them, you're saying, how many things do I get, right? And how many things do I get wrong, and they're just different ways of measuring, you know, that, you know, from a precision recall perspective, so, so if you, if you don't have the human knowledge to understand which cases are actually quite bad, or which are the way you should have designed the feature to be a little bit different to be accurate. Like, yeah, all machine learning can do is optimize the model presentation that you give it right? You know, like, for example, for example, you could, and also Twitter, you're dealing with multiple modalities. So like, one problem that I think is a lot is frankly, not very controversial at all, I would think that, you know, our team was involved in was, you know, you want it to So Twitter allows pornography. But not all kinds obviously. But they're pretty loose about it, I guess. And, but people still want to identified it because they don't want pornography to appear in search. And much more importantly, don't want a period with next to ADS, people don't like seeing their ad next to hardcore pornography. So it was sort of useful to write classifiers to identify this. And they're trying to identify it from what from like, the account activity, some of the text, you know, and also from the actual images and videos. So, you've got this sort of, you know, for a pretty simple problem, you have a pretty complicated system, we have to look at different kinds of data then making, you know, sort of joint predictions measuring those against each other. And, you know, I think, and I think, like, we needed deep learning, which was our team, because doing it without the images was, frankly, impossible to, like, the gap between doing it with and without the images was massive. But it still gets it wrong all the time. Like, like, I've tweeted out a couple of times. But there's particular queries that I do, or even like, I hit like a trending tabs, and I'll just see pornography right there show up in the search. And usually, it's the first result also, for whatever reason, because I think when it does it, I think my guess is that the model also has some sort of trade off where it's kind of like when it accidentally shows you pornography, it's going to be stuff that also got like a lot of engagement as well, for whatever reason, like a lot of people's favorited it. I don't know who's going out there. favoriting pornography on Twitter, but apparently a lot of people do.
Yeah, there's one woman, I still follow her, I think by muted her because she's a very mischievious type. That's what I'm gonna label her as, and she does retweet pornography into my timeline. Literally, she's the only one. I would not know that Twitter. You know, this is one of the things when they talk about censorship. Twitter does have a lot of, I mean, a lot of people actually don't know that Twitter has a lot of porn. But it does, you just don't ever see it. You just never
It doesn't show up in search, like at all except by accident. So even if you wanted to search for a pornographic account, you would actually like, it would be , you would actually have a hard time finding it, you'd have to either type it in directly, or you would have to search for it on Google, because of course, Google will gladly show you Twitter pornography and doesn't care at all.
Yeah, yeah, for now. So, you know, Twitter is, is a business. And you know, it's had some issues with growth. And one thing that you we've we have actually talked about the soft line earlier, but you know, I do want you to touch on it, I get some of the worst advertisements. So I think people who follow me on Twitter know that my personality is I'm not, you know, I would not be a really good candidate for diversity training advertisements, or, you know, diversity, HR advertisements, and I've been getting that a lot recently. And that's not the only thing I get. I understand with, you know, this celebrity, what you don't know about, okay, like that I can see because that's just a political. And everyone's kind of interested in that, even if you don't click all the time, but some of these misses. Are the misses just salient to me, or is this is this a huge problem?
Yeah. So I mean, yeah, we can we can do a deep dive on the Twitter advertising thing in a minimum because it's actually it's actually pretty interesting. But let me just sort of guess I'm speculating wildly here, but I would guess that there's two reasons you would see those ads. One is that They're just being shown. Like in general, like, like, for example, being in New York, I used to see a lot of political ads around elections, on Twitter. And I don't think that they're really like checking whether you would support this policy or this candidate, whatever. It's more like, look, you're active, you know, and you're in New York, right? So it could be something literally like that. In your case, though, I would actually push back a little bit. And I think, from machine perspective, I actually don't think that these are particularly bad ads for you to be honest, because you are adjacent to a lot of people who sort of care about diversity and inclusion. And like the bit flip, like the one, like the 180 thing is actually often like, it seems like very bad, but it's actually very close. Right? The person, you don't want to show that out to someone who has no, no interest in that space at all. And I would argue you actually have a lot of interest in that space. You just disagree with that point of view? You know,
Yeah. I mean, does Twitter use people you follow? Because I have a lot of reciprocal followers. Yeah, so I follow. It's changed over the years as I've gotten more followers and probably gotten a little bit more right leaning, but I've normally had a uniform ideological distribution. And in terms of who I follow, probably until very recently, I followed a good majority of liberals, because I follow scientists. And there's just liberal
that's basically it. So I think, yeah, think of it that way. So think about it as - okay, so I'm not going to do the whole deep dive, but I'll give a quick spiel on how on like, like, personalized advertising, right? So the strongest signals are not really that personal, they're gonna be more like, you know, demographic, like where you live. And then like, who's your circle? So absolutely. So if I was personalizing advertisements for you at all, at Twitter, I would be very focused on which city you are, if you either tag it in the photos that you give it to them, or you can guess from your connections. You know, I would also be interested in your sort of, like, where you are in the socio economic thing. So again, based on your connections, you know, I would guess that you will, you know, you will be you know, sort of, you know, a certain age, a certain like wealth level, things like that. That's the kind of stuff where you start. And then on top of that, you're right, the only thing that Twitter really has as any good at all, as a signal is really your connections, which are much more useful than your tweets, because people retweet random things. And like, like your, your Twitter is very good and very curated. But like that is the vast, vast, vast majority, sorry, minority of advertising recipients, the vast majority of people on Twitter, don't tweet at all. And they retweet kind of randomly, they retweet, like, the basketball team, you know, doesn't mean there.
I see there's nothing to train I like I train it a lot. But most people do not
No no no. So you just have to see which city they are in, if they tell it to you and who they follow. And that's it. These are the actual only signals if you really want to think about at a high level.
I see. I see. Okay, okay. That's interesting. Yeah. And I think you did tell me about that before. It's almost like if you want to see good ads, I gotta like follow a subset of people or something I don't know are not good ads, but ads geared towards me, I have to engage more homophily
Well,
the biggest problem is that there really are not good ads on Twitter, which is very sad. But Twitter has basically and Ben Thompson covers this very well. But I mean, I can confirm it's completely true, from my perspective is basically Twitter. Okay, so taking a step back, is the way Twitter sort of thought of themselves very arrogantly, but sort of you can't really blame them too much is that there were so many people would literally say this, there was like, Okay, first there was, you know, Google, right. Because people forget about Microsoft, there was Google that was like the company of the day, then it was Facebook. And like next is Twitter, they really thought that they were the next company, which is like hard to imagine. Now. But that's what people were thinking. I mean, when when Twitter IPO’d, they IPO’d at something like 20 or 30% of Facebook's valuation. I remember, some people were like, This is completely absurd, like you're talking about a great company, and a very flawed company, you know, like, the traffic numbers aren't comparable, the tech stack isn’t comparable. And clearly they were right. So the point is basically like, you know, they started with running a pretty sophisticated self serve ad business, which is how sort of Facebook dominates it, you know, direct response, all that stuff, but it kind of failed. And now you don't see it. So like, the real problem for Twitter is that they actually don't have that many advertisements. And then the advertisers, they have tend to be like very broad based brand advertisers, they actually spend a huge amount of money way more than Facebook, on like designing Twitter specific campaigns, like the NBA is going to do a hashtag, you know, sort of like some big soap company is going to do a hashtag, they're gonna do a custom emoji, and like, what, specifically those things changes every couple of years, but they work really hard to sign up big advertisers. And then everything else is sort of just like, you know, like it just not a big enough business to really invest in like, there's just, there's just the long tail of advertisers are just not very good. Not very deep. So, you know, it's a chicken and egg problem, right? Like, they could build a lot more better tools and better targeting to attract them but then you would also need to like get back them at the same time. And they used to have those advertisers and like everyone tried Twitter, it didn't work for them and they stopped.
Okay, so um, you know, I'm interested to ask you like a pop question kind of the, I don't know, if you've, if you're gonna be able to tell me immediately, basically, you know, you worked there, you're an engineer. You were inside the castle, so to speak. What about Twitter? Would you tell people that they do not understand thhat the people on the outside have a misimpression of?
I mean, a lot of it right. So I think And again I think people directionally people who are like on Twitter, a lot are very invested in the platform, I think directionally they're often right, by like, mechanically, they don't understand how things work at all. So for example, people are like, Oh, why does Twitter have all of these people? And I just touched on it a second ago, like, it actually takes a lot of work to sign up large advertisers, because large advertisers are gonna go to Facebook, because everyone's on Facebook. They may go to snap, because snap has a certain demographic that Facebook doesn't cover as well. They're certainly going to go to tik toc, right? Why would a big advertiser, why would a big brand advertiser go to Twitter, I mean, you can just bring them in with lower rates, because they do charge less per ad than Facebook, obviously, by, like, they actually go out there and say, Look, we're gonna design this great campaign for you, you know, and they have, and they really have, like, the, like, the big advertisers do advertise on Twitter. I mean, most of the stuff I see is brand advertising. And I think that's true. sort of broadly. So a lot of people do, like, need to do that, you know, you know, if you if you fired all those people? Okay, where does the ad income gonna come from, you know, like, it's, it's a, it's a common, you know, what you would think if you don't think through the problem that Twitter wouldn’t need to have as many people as let’s just say, Facebook working on ads, because they don't show as many ads, they're a smaller platform, but in a sense, they actually need more people, because they're further behind. You know, that's one misconception. And the other one is just like the amount of effort that it takes to just maintain the app, just keep it keep it sort of seeing through the latest iOS, and Android changes. I mean, you have to remember, I mean, Twitter is a relatively simple app, but there's still like a lot of stuff going on. There's a lot of languages, there's a lot of, you know, it actually works pretty smoothly. Like I think we've all seen, you know, who've tried like the Twitter, clone competitors, just what a disaster, those things are UI wise, livewise, rendering wise, you know, being able to find things like there's a lot of things that actually, Twitter works pretty smoothly. So we take it for granted. Just on very basic things like being able to serve you the content, you know,
all right. Okay, so I do have to I do have to say, I think, and a lot of the listeners are probably on Twitter, but not all of them. It seems like the fail whale, when the Twitter used to go down is a lot less frequent. So basically, over the last decade, they've really amped up the engineering.
Yeah, so I'm glad you brought it up. Exactly. So when when I mean, I don't remember, I don't know when you started on Twitter. I'm guessing it's around the same time I did, which was sort of like early ish, but not super early, when you will get the fail whale sometimes. There's even another thing that happens. So when I was at Twitter, you know, there was a couple rounds of layoffs. I mean, Twitter's had rolling layoffs for quite some time. And at one point, they did cut. So this is called an SRE. side reliability engineer. But basically, there's kind of fundamentally two kinds of, quote unquote, normal engineers and big faang companies. There's people who like, build, you know, build and improve and maintain
Nikolai, you just dropped FAANG, tell them what faang is,
oh, yeah, that's just Facebook, Amazon, Apple. Netflix, which is kind of a weird one there. And Google. So. So basically, yeah, like the big like, you know, the so called big tech names.
Nikolai - I gotta interrupt you again. If you dropped Netflix, it would not work.
That's right. And then if you Well, I guess, Google was alphabet, right. So it'd be just FAAA.
Anyway, go on
Yeah, you'd have to, you might have to rearrange the letters and the other the other combination, but yeah, exactly. You got FAANG you got BRICS and yes people like these things. So yeah, so So basically, you have two kinds of engineers, people who build out the product, maintain the product. And then there's site reliability engineers, which are actually the people who have something called pager duty, when things break, they like triage it, when things go down, they have to fix it. If a server dies, they bring up the new one generally speaking in Silicon Valley, like if you're a development engineer, if you're a normal engineer, or even or especially Research Engineer, which is kind of what I did most of the time, you would never be like supporting a live product somebody else would support the live product. So So yeah, I mean, Twitter needs those people to to make sure that small outages don't escalate right. And that is a fail whale thing you talked about. So again, even when I was there, like they did a little bit of layoffs there and like you know, it's Twitter got pretty unreliable for a while, you know, Yeah, so there's definitely a direct correlation between how many people and how well staffed that is and just Twitter like being live all the time. To thepoint…
Yeah. Yeah. I mean, you don't I mean, you don't think partly it's also like the cloud and just these distributed storage technologies and like, you know, computing have also gotten better.
Oh, yeah. So so. So that's, that's a good point. So remember that Twitter predates that. So Twitter was, is kind of old. Twitter was not actually built on AWS, Twitter was built on their own scaling thing. And to their enormous credit, I have to say, Twitter, I want to say last year, but like pretty recently switched off their main services from their own cloud, actually like to AWS, you know, which is expensive and difficult and sort of like a big, you know, sort of like, humbling moment where you're saying, Look, we're not actually good at running this, we're better off having someone else do it. But I mean, you can only imagine how much effort that takes to build a parallel system to test it and to swap it over. You know, while it's running at scale. Yeah. Yeah, I mean, Twitter. So that's the other thing is, if you think about it, like when you build a new service, you know, you're building, I'm not going to name them, but you're building some Twitter clone, you know, no one has experience with like, something that actually like runs at scale. You know, like, there's, there's a, there's a tiny number of services that actually have millions of users interacting, having their own stuff, etc. And, you know, and, and, yeah, I mean, if you build it from scratch, you could sort of do it better. But you would need to bring in people who have done it before, like, it's not something you can really clone, you know, and then Twitter also has all the legacy of it. So they have to sort of decide, are we going to rebuild it to take advantage of the new technology, as you pointed out? Are we going to keep it going? Sometimes you have to you have to maintain two systems in parallel for a smooth switchover, it's a lot of work.
So it almost sounds like so, you know, let's talk about, you know, more timely, you know, we don't know, you know, two years from the future, this is going to be or two years into the future, this part of the conversation, I think, is going to feel kind of like a historical note. Whereas I feel like the first half of the conversation that we've had so far is going to, you know, deep deep learning is going to be deep learning for decades. Right? So I mean, that'll be relevant. So we've just had an ownership transition at Twitter. And there were these rumors of, you know, 75%, layoffs, and, and all of this stuff going on. From your perspective, as a former Twitter engineer, did you always think that the 75% layoffs were just, you know, hot air? Because I know, people that work there now in product? And I'm not gonna name names, obviously. But you know, they they do believe there are some serious over staffing issues. Twitter doesn't do product product? Well, there's some sclerosis. So, you know, put it put your business hat on here for a second?
Yeah, um, I mean, I don't know whether 75% number comes from it sounds to me, like very made up, which is fine. I mean, obviously, it's like the right made up number because it gets people talking by, you know, I mean, even talking to sort of just experienced Silicon Valley business people, I mean, anybody would call Twitter a restructuring. So to be clear, it is bloated, it's inefficient. People are working on different things. Also, like, a lot of the best tech people have left most of them I would say, like Silicon Valley's full of great Twitter alumni, who still, many of them still like and care about the company, but like, they're not coming back. You know. And just like too many priorities, I can dig into it, because I've, because I can tell you some stories when I was there, not even very spicy stories, but just kind of like how, trying to do too many things. So I think, you know, a reasonable person would come in, I would not fire 75% of people, what they would do is they would try to figure out, okay, what do we need, which initiatives we're going to cut? Which ones do we need to support properly? You know, and how do we do that, like, you do need to bring in new engineers, new management, new priorities, but you also have to do it in a way where people believe that this is going to happen, right? Twitter's had so many changes, what would often happen is like, a new regime would come in, or some new managers, would be hired and the people who were there would be like, yeah, like, we'll go along with it, we'll listen to it, but like, these people are probably going to quit anyway. You know, I mean, when I was there, there were just so many senior managers and executives and mid level managers who would go in and, you know, they'd be there for 18 months. So people have to really believe that the new changes are actually going to happen, they're gonna be permanent, they’re going to be planned, you know?
Yeah, I mean, I guess you know, what you're alluding to is the fact that you know, for okay, how do I say this? You know, humans we can think of us as Homo economicus from an economists perspective, we’re utility maximizers, we're rationalist. But there's an element in the enterprise and the things we do that are based on kind of alignment to vision, hopes purposiveness, an almost irrational belief in what you're doing and so forth. example. I think Steve Jobs what he did in the late 2000s, before his death at Apple seems to be part of it. Yes, Apple pays well, Apple has good people, but he squeezed those people. And part of that was his management, his charisma, his leadership, whereas, you know, another company, let's say, Microsoft in the late 2000s, under Steve Ballmer, you know, their goal was always to see what other companies did imitate and exceed. And that's actually quite rational. And that worked for Microsoft, in many ways. If you look at their products,
it still does. So I mean, Ben Thompson talks about this all the time. Like basically like, like, even say, Microsoft Teams, which, you know, which no one loves, kind of crushed slack, it stopped slack in their tracks in a similar way that like Instagram, sort of cut off, you know, snaps, growth, you know, it didn't get anyone to quit slack. But what it did is it offered an alternative that was free, it was bundled. It's like, more secure. And it's like, and they're not even trying to be featured like shot for shot parody. They're like, it's good enough, you're going to use it. And it works, because it's integrated. It's part of your suite. You know, so there's they still do that extremely well at Microsoft.
Yeah. Well, I mean, I want to say you have mentioned Ben Thompson a couple times, he runs a newsletter, Stratechery. And it's one of the most prominent newsletters out there. Actually. It's very early, and a lot of people on substack, who have newsletters, even though he's not on substack , they tried to recruit him. And I think he turned them away. Because he can, you know, Ben, Ben Thompson is kind of a mentor, kind of a hero. So just just want to put that out there. People are curious about it. You know, in terms of tech news, you know, it's his daily newsletters, great. But yeah, so we have these different strategies. And now we have Elon Musk, you know, I think he's still the richest man in the world. Now, he owns Twitter. And he's, he's actually like, a power user of the platform. So why did he buy Twitter? I think, ultimately, you know, it's like the the old Hair Club for Men thing, where, you know, he's, you know, he's, he's gonna be using this and he's owning it, or, you know, I think there's other investors. But do you think I mean, because a lot of the people internally from what I see in the news, and also those people who hate the news is that they don't like Elon Musk. And they are worried and you know, they can barely breathe. I don't know. There are some people I know, internally who have no problem with Musk, and they think he's going to clean house, but I don't see that reported.
Sure.
So so. So one thing I think that's kind of interesting about Twitter. This is true about a lot of companies, but it's extremely true about Twitter is it's as a business, it's a lot less interesting and valuable than as like a thing. You know. So in that way, it is kind of like a media company. I'm not trying to, you know, say in terms of the SEC definitions, or whatever. But I mean, you know, look, when Jeff Bezos buys, buys the Washington Post, no one thinks for a second, that he's buying it, because it's like, technically a private business, and it has like good returns, you know, like, he doesn't want it to lose a ton of money, maybe he doesn't want to lose any money. But, you know, he's buying it for the influence for someone else not to have that weapon to use against him. I don't know. But like, no one in the right mind thinks that, like, The Washington Post is not important. Even though like as a business, it's like, whatever. And Twitter is among all tech companies, by far the most that way. As a business, it's okay. It's not as bad as people think. But, you know, if Elon isn't buying it, because of its like, revenues, you know, buying it because it's interesting, because it's important. I'm not saying he's doing it for his own personal brands or political reasons. I just think that he's like, it's undervalued as an asset. You know, on a non accounting basis, basically, it's not based on the PnLyou know.
Yeah, I mean, so one thing, one thing that we can talk about using economic concepts, you know, there are certain industries, certain sectors, where people, you know, certain businesses where people and corporations, institutions, they capture the value that they produce a great amount of it. And then there's other industries where they don't capture the value that they produce in the same way. So to give a concrete example, let's assume in a hypothetical world, so um, you know, not just this is not average, but there's a superstar teacher, and, you know, this teachers sending these kids, you know, onward and upward and there's, there's a long track record of value add from this teacher, okay. Okay. Well, I mean, the teacher is not going to capture the value that they add to society that the value that they add to society could be incredible. But they're salary is not rising commensurately. You know, on the other hand, there are other professions where, you know, the better you do, you're gonna scale up a lot, right? So,
Sure, like when I was when I was in the hedge fund industry, right, like, like, that's the classic example. Right? Yeah. I mean, you, you sort of you you live and breathe your PnL you know, you're as good as, you know, as your last trade. It's very stressful. But yeah, I mean, but you, but you do get a pretty good, pretty good chunk of the value you generate, yeah,
but but the teacher does generate value, it's just a positive extra. They're positive. There's, you know, spillover effects, positive externalities. So innovative technology is like that. Obviously, Apple created the iPhone. And it's like the profit maker, you know, since since it started, right. I mean, it's really driven the company stock up in their revenues. But one can argue Apple still not capturing a lot of the value of that iPhone, because iPhone transformed our lives for good or bad. So there's positive externalities, negative externalities, you can get into arguments about what they are, I think Twitter in particular, has a huge, huge amount of externalities. And it's not necessarily capturing the value of that, or, you know, a standard business, a standard, a business, PnL as you would say, are not like capturing the value that is producing culturally and socially. So I think that's what's going on where Twitter is much more important, because of its use in the elite media, and whatnot. And, you know, it's also as it has pretty high traction in places like academia, like politicians use it to communicate to other elite people, and so forth as a broadcast service. So I think, I don't know how Twitter is ever going to capture the value, but it is there, you could argue its value is actually it's actually negative, right? Maybe it's anti value, but either way, it has an effect, it has an outsize effect in relation to what it's valued at in the revenues of producers. Right.
That's right. And I think, to your point, just a couple things, I mean, you know, using Microsoft, your favorite example of the conversation, you know, they would specifically say that they want to build an ecosystem right around the office around whatever. And they even would say they want to capture about 7% of the value. They use that number all the time. I don't know, maybe they're bullshitting they probably are. But, you know, like, you know, and I think that's generally been the traditional sort of, you know, you know, the Silicon Valley View, let's just say 20 years ago, right? You're creating all this value, you're capturing a small but non trivial chunk, and like your partners are like thriving, right? And I think Ben Thompson, again, mentions just how useful Facebook ads were for like small business, right? Because they weren't great at direct response. They were great at finding like niche of people who are like, specifically going to be interested in, you know, in golf shoes, or something, you know, and Twitter was never good at that. And Facebook's amazing at that. So like, who's getting the biggest part of the value? That's actually potentially like the golf shoe makers, because they can find their customers cheaply. And they need it, you know? So I think that's sort of true broadly. And I think it's Twitter has absolutely like, not done well there at all, you know, but I think as long as they remain prominent in the prestige media in the world, and things like that, and people are always going to be interested in the long game. You know, I think the biggest Twitter, the biggest Twitter fans tend to tend to think of themselves as very long term thinkers. You hear that a lot. They're like, Well, geez, right now they're breaking even basically, Twitter doesn't actually like lose money. And if they want it to be cashflow positive, they can turn that spigot on tomorrow, like they're not like going out into the public markets and losing a bunch of money. They're not like when Uber was like losing billions, right? It's just not that kind of business. Like, you know, what they're what they're not doing is they're not, you know, they're not improving their business by and to your point, they're not capturing the value and the prominence they have, but somebody could just figure it out in a way we don't even anticipate five years from now. And the owners will benefit, right?
Yeah. Yeah. Yeah. I mean, you just mentioned Uber. I mean, Uber still hasn't posted a profit, right?
No, that’s
no, it's not true. So on an operating basis, we make money in North America.
Okay. Okay. Okay. Okay. Well, I mean, that's, that's good to know. I mean, definitely, their prices have gone up. So they're,
yeah,
they sure have. well, and then that's part of it, right? Is that like, look look at like Amazon, right? Like Amazon. That's why like this, just looking at very, very basic PnL doesn't really make sense for these growing companies, you really have to look at the, you know, like Amazon has made money for many, many years. But what they would do is they would just invest, they would just reinvest, they will take some mature projects that are like printing money and reinvest that into like, more recent projects and if they ever ran out of recent projects, and they would like acquire Twitch, right. So, I mean, mainly, I think they were they were trying to avoid You know, posting profits, maybe for tax reasons, I don't really know. But the point is they just had a very good system where they would trade off. I actually actually I even had a friend who was a was a, you know, was like an executive at Amazon. And she was on the third party seller thing. And they would even tell them at the beginning of the year, it's like, look, here's like, how many billions, you're allowed to lose this year. But like, here's what you need to do in terms of growth, you know, and they ran that for years. And eventually, they're like, Okay, now we're gonna, like, now we care about profitability. You know, like, the, just the Hi, little bean counters did their math and they're like, Okay, like, now you don't need to grow by 300% a year anymore. But you need to make money. And that's often a very easy switch, you know, in a mature thing that's still like a thriving business. You know, I wouldn't necessarily describe Twitter as a thriving business. It's more of a breakeven sort of like, it's a bit more of a stalemate.
It's a, you know, it's a sustain a sustainable enterprise. It's a going enterprise. Right?
It's very much a growing concern, but I think, yeah, the value the value in the upside is in the very clearly things like things we can define that are not business based just how important it is. I mean, the fact that even all the countries that banned Twitter, their leaders still post on Twitter, right, like I find it very, very amusing to see that.
Yeah, I mean, okay, I wanna, I want to ask you what people that you know, because you still know, people at Twitter, like you're part of the Twitter diaspore, right?
Yes.
Okay. You know what, let me just let me ask first, and then I'll end with a couple questions. But Elon Musk, like, what's going on here? What do you guys think? Do you think he's insane? Do you think he's cool? Are you guys just scratching your head? This whole thing from the outside? has just seemed like reality. I mean, we've had a lot of reality TV ask like social political business things in the last decade, we're going to get into everything. But this is just seemed off script. The simulation is being weird. It's just seems capricious, erratic out of the blue. As you said, there's not a big business logic. I don't know. It's just the whole thing is just inexplicable to me on some level- but hey, you know what, I'm not 100 billionaire or whatever.
Yeah, none of us are. I mean, I think I think opinion is very much mixed, I would say, the more the more sort of ambitious, forward looking excited people you know, tend to like Elon generally, partly because they all have Tesla's, they're sort of in that world, but but that's going to be over represented among like the brightest of the Twitter alumni, people who are still there are going to be generally a lot more anti. I mean, obviously, the anti people are gonna be more vocal, if you're pro, there's gonna be like, a lot of like, silent, there's gonna be like a lot of shy Tesla supporters is my guess. Like, there's no reason, getting on the rooftops and saying, I'm thrilled about the new management until it sort of like sticks. You know, I mean, I think Elon, including in the Twitter deal, kind of has a little bit of a history of rugging people. You know, to put my cards on the table, I'm I like Elon, but I'm not the biggest fan, I'm certainly not a fanboy. I think he's done a lot of stuff. That's bad. I think he's miss led a lot of people. You know, I think a lot of people, you know, should correctly be, you know, I mean, he tried to pull out of this deal very clearly. So I think, you know, coming in and being an Elon supporter is a little bit tough at the moment. I think also, because it is a restructuring, I think they are going to replace a lot, if not most of the management, they probably should. I mean, they probably should get people who are gonna be loyal to the new regime. You know, so that's going to be difficult. I mean, I think the dumb idea of many is that like, oh, like Elon ran these other companies, and they're doing great. So you can just do the same thing at Twitter. I mean, it's a totally different product, totally different culture. I mean, you, you know, you're not starting from scratch. So and I don't know that he has that much experience with restructurings, you know, like, I think, I think I'm very curious. It's going to be very interesting to see what kind of restructuring type executives they bring in. One thing, that's for sure is that when I was there, and I think it's gotten a little better is like Twitter was doing too many things. So when I was there, Jack Dorsey was a CEO. And I like Jack a lot, a little bit. But like, like Jack a ton, but you know, he always had too many ideas, right? So Jack would like literally, and there's also also into like radical transparency that like, even sounds insane these days. But, you know, he would send out these seven page memos, like every week, every other week of like, what are the top priorities on Twitter? And I'm like, How can you have seven pages of top priorities, it just doesn't work. So people had a long experience of things being announced things being publicized, and then just being cancelled. months, months later, I think any proper restructure executive would come in, not do anything at first keep the lights on figuring out okay, like what do we need to do? What do we what businesses do we want to get out of, and I do think Twitter in the past few years has done a good job of selling off assets, getting out of things they don't need to do like the move like sort of onto AWS from their own infrastructure was like a great move. But they're still invested in a lot of things they probably shouldn't be doing at the same time. Like you do need to keep up like Twitter spaces, I think has been a big success. Even though technology's like not great. possibly worse than clubhouse, but like it definitely was a need. It definitely captured a lot of attention. So like, you know, I don't know, I mean, I think I think I hope that Elon kind of like, delegates it to a good restructuring person and brings in good engineering people from his org. I would assume that's the plan. But like, I don't know, man. I mean, like all these people, when he first announced it, all the people that I know were like, Yeah, this is like the thing that would make me want to come back to Twitter. Like there's none of them are going to do it. They just talk, you know, and they're not going to come back to Twitter. All these alumni are doing great. They're in their current jobs are like running their own companies. And like they're, they're like retired investing in like venture capitals. Like he has a lot of goodwill, behind the scenes from people interested in Twitter, but they're not going to do anything to help him.
Yeah, yeah. It's all. It's all talk.
It's Twitter. Of course. It's all talk. It's Twitter is all talk right. Like, that's why we love it.
Yeah, that's funny. Yeah, yeah. One question that let me close out with this. Because it's kind of like out of left field. But, you know, we're talking about Twitter. It's kind of a public good, almost. I don't know how this would work. But I mean, here's the issue. The European Union, whatever censors are added, that's not her title. She's already tweeting at Elon that, you know, it's not going to be 100% free speech, because European Union obviously has some ideas. Obviously, places like Saudi Arabia, their view of Twitter is going to be different. What about the idea that the American government nationalizes Twitter, and it's an American utility, and we impose the first amendment on the rest of the world? I mean, it sounds crazy. But, you know, we have the Twitter's like, Twitter is like a going concern, because it has critical mass, it has network effects, and no one's going to take that.
So it's not it's not crazy at all, I'm actually, by the way, a very popular view around the world, is something that's not true. But it's like, also, not entirely not true, which is a lot of people around the world basically think that Twitter was either started by are completely co opted by the American intelligence community. And that's not true. But I think that there's like, a significant amount of truth in it. Like, they clearly recognize that it's an important asset, and the American Deep State or whatever you want to call it is very happy that they control it, you know, which is why exactly why he's banned in China. So I mean, Ben Thompson, you know, on, you know, his, like, longer form podcasts talked a lot about that, you know, living in Taiwan and understanding this stuff, like it's completely rational for, for China to ban Twitter, just, you know, the same way it's rational for the US government to think about, like, do we really want tiktoc you know, controlled by the Chinese government, like for example, like with, you know, relatively small state things, but you know, when they had the Daryl Morey controversy in the NBA last year, then like, all of a sudden, nothing about the Houston Rockets appeared on Tik Tok, right. So Twitter can absolutely do that, like you can just, you know, I guess the term is shadow ban, quote, unquote, like a certain topic, right? And just not not show anything about it. Like, there could be an event going on in the world. And they can do that. They don't, because they are a lot more into free speech, especially with Dorsey, but they could and I think definitely, I mean, my my take on it again, channeling Ben Thompson as well is, what you're saying is like a very American view, which is you have this thing and you're going to like solve it in like a binary black and white way, once and for all. Whereas I think maybe the rest of us whether you're coming from China, from parts of parts of Eastern Europe, or whatever, we're like, a lot more happy with things being kind of muddled and kind of vague, you know, which I think is more realistic in, especially with international relations. So I think like, like, I think it'd be a big mistake, I don't think they're going to do it for the US government to come in, like try to impose stuff in a hard way. Because then you're, you're missing the opportunity to like, like, just nudge it in a soft way. So it's like if you allow the Pakistani government and the German government to like email, Twitter, and to like, make a stink about stuff, they really don't like and negotiate with them, you're kind of keeping them on board, while still in practice they’re kind of dominated by USG. Whereas if you if you force a confrontation, right, then it's like, why, you know, like, I think people broadly generally are kind of okay with, you know, with America with American tech controlling a lot of the media landscape, as long as like, there's just no point overreaching. Like, it's already very much in USG favors. Right. And in that way, why would they want to run it? Like it's a great setup for them currently. And like musk musk, I mean, I mean, you couldn't imagine a person who is both still generally popular and has very close connections to the, you know, to the US government, right through all of their contracts and dealings in the space stuff. So I think it's a great outcome for them.
So, okay, That was going to be the last question. But I want to ask you, because you brought it up shadow banning and all these other things. How does that happen? Like, I have friends who claimed that their shadow banned and they obviously are, or that are, you know, they're throttled. In some ways your engagement drops a lot. I have other friends who have told me they have a theory, these are more conservative friends, they have a theory that Twitter and YouTube has basically chosen not to create any new conservative or right wing stars that, you know, start getting a lot of engagement in English, whereas, you know, in other languages, you know, it's a much more fully fleshed out ecosystem, because they don't engage in this sort of behavior. You were there? I mean, what is your sense of how often this happens? And I mean, how does this even happen? I mean, is there like, Is there is there a council of censorship at Twitter?
I mean, of course, there is, but it's not, it's not going to be as formal, you know, it's not going to be a bunch of, you know, very diverse people, you know, smoking cigars, you know, or vaping, or something, you know, it's never that formal, you know, but of course, of course, there is, of course, they they're aware of what's going on in the platform. And they want to have some control over it. I mean, but it's, but it's, it's again, it's like subtle, right? So for example, YouTube is a great example, like, the, you know, the CEO of YouTube is like, very, you know, it's like a very well known partisan, and her views are very well known. And yet she was CEO of, I guess, is of Twitter, which, as Jordan Peterson correctly pointed out is like, the vast majority male and just interest that are like, pretty different from management, you know? So, in the case of YouTube, they certainly push their own interest, I don't think I think you'd be really strange to think that that was not the case, you know, but still allowed kind of stuff that is people want on the platform, you know, I think a lot of the people who are shadowbanned, they're either like really on the fringes, or, you know, a lot of people are just kind of sour grapes. Like, I think a lot of their stuff, frankly, wasn't that engaging to begin with, to be quite honest. But I think there are things that matter, like you can, if you are in that world, where you think you're going to be shadow banned and censored. Like it's sort of on you to avoid own goals, because no one's going to do you any favors. I mean, I think a lot of other stuff is obviously like, and a lot of the the the Twitter's control is really like not even around that. It's more like, Look, we're not going to we don't want you to link to certain websites, right, or in general, they just don't like links in general. But it's not all political. Like, for example, in crypto, everyone has this idea whether it's true or not, but like it's true enough people think it that basically anything with the link to an ether scan, like just gets like no views. And ether scan just what people don't know, it's just it's just a transaction on the ether blockchain, you know, just a link to a transaction. And for whatever reason, they just don't want those links. So, you know, if you post those links, people are not gonna engage with your stuff, because it's not going to be shown, then you kind of have to know that. In general, Twitter has really, really, really suppressed links in general, people don't like Twitter, generally speaking, for, again, algorithmic reasons, and maybe like, human decisions just does not want content that is going to link out. So if you're promoting articles, that's not going to work out for you. Now, could you imagine that they have a rule in place where they sort of like, decrease their own rule for sort of like the official media, of course, I mean, it's kind of like Wikipedia, right? Everyone understands certain kinds of media is just preferred as a source. And that list isn't going to be like, what, maybe you and I would like, you know, but like, Twitter has an interest in not crushing the New York Times New York Times a good customer for them. Just like strategically if anything, but in general, they crush links. So it can be logic like that, like linking out is bad. But like, if you're in New York Times, you get an exception. Is that fair? No. But it’s the way it is
Well so here's a question. Here's a question. I've heard this linking out thing. It wasn't always true, though, right?
It used be true at all. It used to be the best way! But then Twitter didn't want it so they changed it.
So I feel like I don't know if you know the chronology of this. Because I still do probably every day, I probably link to a preprint or a science article. Like sometimes I'll do like five, sometimes I'll do one. I feel like the likes on those got, like crashed and 2017. And they've never really come back. Like I used to get like, and this is back when I had probably 1/5 The number of followers that I have now less than one fifth.
Razib, Do you look at your impressions? Because I mean, you know that right? Like, you can look at Twitter analytics. So are your impressions much lower than just like good messages? I would assume it's an impression issue. If their not being show to anyone.
Yeah, they are. You're right. Okay. Yeah.
No, no. So that is I mean, you can call it shadow banning, if you want to, like, I wish there was a more neutral term. I don't like these loaded terms. They just really like are bad for discourse, you know, but it's like dude, I mean, like, again, I’ll explain the Twitter algorithm in a very simple way. So it's like, certain kinds of content they prefer But then also, when you show it and people engage with it, then they show it to more people. Right? So like, part of it is that I think they're not entirely wrong, like people were like, Dude, I mean, think of it like, again, for archive, right? I mean, being in deep learning, I used to post a lot of archive stuff, and people don't click on it. Because, like, like, it's why should I click on it? You know, like, I'll read the article later. It's just like, how do I engage with it? You know, so it's, it's sort of, a lot of it's kind of benign, but it's still true, you know?
Yeahmy sub stack pieces, tend to, I mean, it depends on the picture I associate with them, to be honest people. People love maps. So I can get like, you know, 500 likes on a substack, which is just great. That's my goal. So yeah, there's definitely like variance. And, you know, you try, you get trained yourself to I mean, depending on if you really care about likes and retweets, when it comes to my regular stuff. I mean, just candidly, I mean, I don't think this is like going to be a big shock to anyone. I'm, at probably 50% of the time I tweet on Twitter, it's mostly because like, I want to get people driving to my substack with a pin tweet. And so that's the only thing that I really care about people clicking a lot. And all my other stuff, I just tweet what I really want to tweet, and you know, one out of five times, like, they might travel, and then I get a lot of engagement. That's great. But I'm not one of those people that tweets to get a specific engagement, which, you know, I think maybe some people actually I know, some people are, some people.
I was gonna say, engagement forming is a real problem, like, and Twitter isn't addressing it, and it's not. And I think they're just generally kind of hands off, which is generally good, but it has unfortunately become, there's just too much incentive to game it. So people post kind of like intellectual thirst traps, I mean, like, like those, like those ridiculous threads about like, you know, here's a thread about how, like, you know, the road gauge is related to like, the Roman chariot thing. Complete bullshit like that, because it does well, it does in the algorithm
The buckle up threads.
I and like, and like, like, everyone hates it, like, certainly in our circles. Everyone hates it, but like, it does do well. So we will post it, you know, it travels to your point. It gets sent to other people and, and it's like, yeah, I mean, I think that's the downside of algorithmic? Well, it's actually not okay, so this is like my little hobby, or some sort of struggle. And then it's like, it's it is something that sort of gives the default for the way these algorithms are trained. But it doesn't have to be. So one of the coolest thing that Jack talked about Jack Dorsey was the whole bring your algorithm thing. Do you remember when that was a bit of conversation? Or like I was like I can. So So basically, the idea is that, yeah,
I don't I don't remember that particular conversation if you want to refresh us.
Yeah. So what it was pretty cool. So basically, the idea is very simple is Twitter, right? Right. Now you have this kind of like, kind of like, like, sort of like Mexican standoff, right? Where it's like, Twitter, like, we are all like Twitter give us a better algorithm, you know what I mean? But like, they won't let us improve the algorithm or even see it, right. But like, they also don't like, they need to build an algorithm that works for everyone. So they bring your own algorithm ideas, pretty cool, which is that Twitter would have a default algorithm, but somebody like me, or, you know, a grad student, or whatever, could actually write their own sort of, like, timeline ranking algorithm that would use the public features, basically, not the full set of features, but like, a complete enough subset of the features, like basic things like, okay, you know, was this tweeting, you know, engaged with by like, someone from your close connections, right? You know, how many times have you engaged this person, the last month, just sort of these, like numerical features, just a few dozen of them. And they could build their own algorithm, they could test it. And then like, you could opt into their algorithms. So it'd be like an algorithm store. And you could be like, I want, you know, I want like Bob's algorithm to sort my tweets, which would actually kind of address a lot of these issues, potentially, because we are, like, there's just no real economic reason for them to build an algorithm that's good. For Nikolai. And for Razib, and for other people. It just doesn't make business sense, right. But they also don't allow us to, like rank our own timeline in any way whatsoever. Like they have to build for the common denominator. But like, for the 5% of us were like, way too online, like, it would not be hard for a grad student to build something that you would find more compelling for your own stuff, or you could recommend it to your own readers. Right. I think it's a very good idea. It would definitely be like, it's something that there's no, there's no reason to think it will hurt their business in any way. And that'd be really fun. But unfortunately, it's probably not going to happen, although it should. You know.
Yeah, yeah. One thing that you know, one thing is like, you know, Twitter obviously has massive media uptake. The media loves Twitter. They're the super users. And then there's the whole phenomenon. I don't know what you as a former Twitter person or what you were what you thought about this, but isn't it kind of funny and this isn't some conspiracy, but it you know, I have like, you know, friends who have hundreds of 1000s of Twitter followers, and they cannot get a blue check. And then like some. And then some guy that works at a small literary magazine has a blue check. I'm checking Thomas Chatterton Williams, who's writer at The Atlantic. Now. He's written for New York Magazine, all these places, he cannot get a blue check. He's got 128,000 followers, but there are people blue checks, who have like, you know, 2000 followers, because they worked at a particular, you know, there's,
there's people with, there's people with 300 followers that like work that right for like a local newspaper that have a blue check. Yeah, so So look, here's the thing. So very, very brief history on the blue check. And it's like very classic Twitter, like, like, it's kind of a mess. They didn't really design it, they didn't really define it. Like it's supposed it's very, it's called verified, you know what I mean? But it was like, then it became then they were just kind of lazy about it, then it just became this like this privilege. But they've gone through many, many waves. So for example, there was a time when you could apply and it got pretty, like liberal and pretty open. And a lot of people got it. So I know plenty of people who like by today's standards wouldn't get it. But as long as you don't do something super duper, duper, duper bad. Like it used to be impossible to use your verified, actually, I think I think Milo Yiannopoulos was like, famously the first person to like, be unverified. And then yeah, like - And then banned a month later, by the way, I was on Twitter when Milo was still tweeting live along with like, Marc, along with Martin Shkreli. Doing, you know, the periscopes, you know, before he went to jail, it was it was a really wild and interesting time. But, um, but yeah, so the Verify thing is ridiculous. Of course, everyone knows that. But it's like, you would need people to decide and what they want it to be, whether it's, they want it to be to tears. I mean, my my favorite is some people that we know are like, now they have they got the approval for Super follow, but not for verification, which is just very amusing. You know,
oh, yeah, cuz I got I, I'm super, I have super followers, I got verified. I think in 2017, I wouldn't have done it. But I saw someone who had far fewer followers than me, who was not even a journalist really got the blue check. And I was starting to get a little paranoid about Razib imitator accounts, shall we say?
That's right. Yeah, that's right. And that's the thing that's weird about it, of course, is that, you know, if you are big, you get imitators. I mean, the original, the original motivation, or whatever, for verification was just that with the, with the sort of Know which is the real account, but then it's sort of, you know, it sort of took on a character of its own right. Like, no one, no one can even tell you what verification means. It's like mix of things, which is like very organic and very Twitter. And like, kind of very cool from an evolutionary perspective. But but but is completely wrong and weird and doesn't make any sense.
All right, so we've been talking for a while, we should end this, but um, I guess, let me just ask you real quick. You're still on Twitter? I'm assuming you feel like you get value out of it.
I mean, of course, I mean, I think I think Twitter, Twitter, for me, does at least a couple things, which would be worth just in of itself. So first of all, you do connect to amazing people, like the number of actual, super interesting friends that I've made through Twitter, including us not, is not completely not trivial. And then also, having enough of a presence having a minimum presence does help you for, you know, business and recruiting and things like that, like, it is a way to get your stuff out. Like I actually feel very bad for people to trying to build a new Twitter account. Some people are, you have to be very talented these days, I feel like to have an account, even getting to like a few 1000 followers and like that are like engaged, you know, almost, you know, it's just so hard. But then the final thing is, of course, is just like it does give you a chance to practice writing. I think like, in a way Twitter is, especially since they switched to 280 from 140 140 was too short. It was it was like it is really like the right way to condense ideas to get actual feedback. I mean, the vast majority of us went to American schools where you never wrote for actual people you wrote for like a prompt or, and you see that with people like that, sadly, the vast majority of tech minded, you know, are mathematically minded, like people can write or communicate at all. And I think Twitter is probably still the best place to practice, right?
Yeah, for all its faults, I mean, you know, I think I will say, for me, it's been mostly positive, obviously, it's just that periodically, you get caught in some Maelstrom, and it's just like a pain in the butt. Or every literally every single person is dunking on the exact same thing in my timeline. And I just think, you know, what has gone? I mean, come on, like, you must have seen that, you know, because like, I turned off the algorithmic timeline, but it doesn't matter because everyone's talking about the exact same thing.
That's right. Yeah. And I think just to sort of a tangent on that. I think what's very interesting is the spam and the weird stuff and all of these things that happen um, There is a communication platform where we really don't see any of these problems. And that's called slack. Right? So I think an interesting bug of like email, and, you know, and Twitter and Instagram and all these other platforms is that they basically, like allowed people to just, you know, communicate with and reach out to people that they're not connected to. And, you know, I think it has, I mean, like, people talk too many times at Twitter, we even thought we even like sort of in the sort of brainstorming things like design, you know, design ways of Twitter, we could sort of have an alternative to that, but people really resisted it for historical reasons, you know, free speech, or it's so exciting when you have these like, spontaneous connections between people when, you know, different celebrities are fighting or whatever, you know, but I think a lot of the ills I mean, this is just my opinion, I think a lot of there's some positive externalities, but there's massive, massive, massive negative externalities, when you just allow people to just communicate to people like without establishing some connection first, because you can relax that right. Same with email spam thing, if you simply email was set up from the beginning, we you can’t just cold email people, you would have downsides, but you would have upsides, right? And then you can relax it, you can say, Oh, well, actually, this person is connected to a person I know. So I'm gonna, I'm gonna allow that. Right. You know? So I think I think the two real sort of, like so not existential issues we find the sort of permanent issues, I think, are this like spam and sort of unwanted harassment reach out thing to I think the right wing people minimize it. Sometimes. It's really, it's really a problem for people people get really upset. Like, and that's not going to change, you know. And then the second one again, is because of the algorithmic stuff, like just the rewards for engagement, farming are just way too high. And I wish that wasn't the case.
Yeah. All right. So Nikolai, you know, it was great talking to you former Twitter engineer and now the you know, you're leading deepNFTvalue.com. I'll put the link in the show notes obviously, but you know, that's that's your next thing. So it was great talking to you and I will see you around on Twitter.
Likewise, thanks Razib.
Alright Thanks.
Is this podcastfor kids? This is my favorite podcast!
