How Does TikTok Read Your Mind? A Computer Scientist Explains | Cal Newport

People are always talking about the algorithm. Like it's some sort of sinister entity with which we're doing battle. It's the algorithm on TikTok. It's the algorithm on Instagram. It's the algorithm on Netflix. But how do these things actually work? And understanding how they work, how does it help us think about how to reclaim our brains? Well, back in September, I did a podcast episode in which I did a deep dive on TikTok in particular. what is happening under the covers from a computer science perspective to help TikTok decide which videos specifically to show you. I think a lot of people found that useful to get this technical explanation for this thing that otherwise plays a really big role in their everyday life. So this is the clip I'm going to show you now is me explaining exactly how the TikTok algorithm works. I think you're going to find this interesting. All right, let me tell you this. As a computer science professor who... specializes in algorithm theory. I mean, this is what I studied in grad school is distributed algorithm theory. It's what most of my academic CS papers are on. I teach algorithms at both the undergraduate and graduate level. So it's been sort of amusing and pleasing for me to see how often I'm hearing these days, non-computer science people use the word algorithm. It's sort of like my little secret world is one that everyone has been exposed to. But when it comes to discussions of social media platforms like TikTok. This term algorithm has seemed to taken on some sort of almost mystical power and capabilities. I want to play a clip here of a sort of recent discussions from last week of, in the news, people talking about the TikTok algorithm. Jessie, let's hear this first clip. And they will basically be helping TikTok to retrain its algorithm. And also to make sure that U.S. data of people, all the Americans, you know, the roughly half the country using TikTok, that all of it is secure. All right. And then I think we have another clip, right? Mm hmm. All right. From surveillance or interference by foreign adversaries and the algorithm. I know this is a question many of you have had will be secured, retrained, retrained and operated in the United States outside of ByteDance's control. All right, so there we see the way the algorithm, putting square quotes around that, is being discussed in the media right now. It's this thing, and we want Americans to be in control of it, that we're going to retrain it once it's over here. And with that sort of supervision, we can have some assurance that whatever it is that we are uncomfortable about happening with a service like TikTok, we can have some assurance that at least our interests are being preserved. But what is this algorithm that is at the core of this new deal? Well, I want to start with the mental model that I think most people have when we talk about a social media content curation algorithm. I think most people imagine it's basically like a digital version of a newspaper editor. So we have a newspaper editor who makes decisions. What's going to go on the front page today? What's worth focusing on? No, put that aside. This is going to have a banner headline. We imagine an algorithm like we're building a computer version of that, a digital version of that that can work at really high capacity and make like building a sort of like a custom newspaper for each person. So a computer program that makes decisions about what we see in the same way that like a computer editor might do. Okay, so in that model, if that's our mental model for algorithms, transferring control of the TikTok recommender algorithm to US control makes sense, right? We wouldn't want a foreign country playing the role of the editor for the social media newspapers that half of the US population is receiving through TikTok. We want an editor who has our values, who has American values, who's not going to be doing, not only promoting values of not American, but maybe conspiratorially trying to mature in candidate style, influence Americans to think one way or the other, right? We would not be happy if the New York Times was edited in a dark room in Beijing. And so we sort of feel with this mental model, we want our algorithm for this popular service to be here. But is that mental model correct? Well, I'm going to take out my computer scientist hat here, which as Jesse and I have discussed is an awesome hat. It has Spock ears, Jesse, let's be honest. And we're going to take a closer look at what really is going on, and we're going to correct our mental model for thinking about these social media algorithms. All right, so what is Oracle going to discover when they finally get their hands on TikTok's algorithm as part of this deal? You know, like most private companies in the social media space, ByteDance has not published detailed looks at exactly how their systems work. I mean, why would they? Most companies don't if they don't have to. But it's not like they've been completely secretive either. We have a pretty good sense about more or less how TikTok works. There's two sources we can draw from. One, ByteDance researchers in the last five years, they published two different major papers in academic venues. that looked at major components of their system architecture they use for multiple products, including TikTok. So that gives us a really good insight. And two, we know how recommendation, the evolution of recommendation system algorithms, that this is something we just know from a data science, computer science perspective. There's only so many ideas you can be pulling from. No one suspects there's an entirely new idea that they're using. we put these two things together, I think we can create a pretty reasonable understanding of what's actually going on here. All right, so let's get technical. First of all, when we say algorithm, that's not really the right word. We should be talking about recommender architecture or recommender system architecture. So what drives TikTok is a massive system that stores and updates information about... hundreds of billions of videos and whatever it is over one to two billion users and all this stuff is soared and accessible and this in the end what this system has to do is for each individual user as they do each swipe is have a recommendation of what next video to show them so it's actually a very large global distributed system not a single algorithm that we need to look on now we know a lot about building these type of systems This idea of using computer-driven recommendation systems really began to get a lot of attention in the late 90s and into the 2000s with Amazon and then Netflix really being pioneers and figuring out how do we use data about stuff we want to show or sell to people, data about the people themselves, put these two together, and make good recommendations. So by the 2010s, we were getting pretty good at these. What type of recommender architecture does something like TikTok probably use? Almost certainly, they're using a general approach that's known as a two-tower system. Let's get into this a little bit. I want to set the stage here before we get to the two towers that they probably utilize. Imagine I am building a system to recommend short videos to you. I have a bunch of short videos. You're a user. I want to recommend videos to you. One way I might want to do this would be to come up with a master list of properties. And I want this to be on the order of a thousand, a few hundred, maybe a couple thousand, but kind of in that order, right? Of a few hundred, maybe a thousand properties, right? That I can go to any particular video and I can assess it on each of these properties, right? So maybe you have a property that's like, it's funny or not. Or there's a property that is, you know, the content is conservative. Or there's a property that's like, this content involves like South Korean K-pop. Now, it could be more complicated than just singular properties you have or don't. It could be complicated combinations. Like maybe one of my properties I assess videos on is, okay, it's like conservative K-pop content that's not funny. Because maybe it turns out like that's a pretty important property. But this is my challenge. They have like a pretty limited list of properties that captures enough stuff about videos, like the stuff that people really tend to care about that I can start to make. good recommendations. And now I'm not going to necessarily just have like a bit for each. Is it funny or not? But maybe I have like a one to 10 or one to a hundred scale where, you know, how close, how much of this property does it have? All right. So imagine I could do that. And I go through each of the videos in my system and I have my big list of properties and I give it a score on each of those. Now assume when you join this service, I say, okay, you're a new user. I'm going to sit down and have like a detailed user interview with you. I'm going to talk to you about all sorts of stuff. What do you like? What do you don't like? Let me show you some videos. Is this the type of thing you like or not? And what I'm going to do is take that same list of properties. And now I'm going to go through and for each of those, try to measure, write down with a number, how important each of those properties is to you. In other words, when it comes to videos you like, do you want a lot of this property or not? So I find out, oh, you love funny stuff. So in that same category, I'll put down a pretty high number for funny stuff. But maybe you're really left-leaning. So you're like, I do not want to see conservative content. So I'm going to put a big zero there. That's not something you care about. So we have this master list of properties. Every video is ranked on them. There's a little tag we put on every video. And every user has the same properties. We say, OK, here's how much they care about each of these properties. Now, assume I was just able to do that. Don't ask me how, but just say I was able to do that. This now gives me a pretty good way of recommending videos to you. What I can do is say, look, here's what I want to go find. I have a hundred billion videos, fine. But what I want to look for is videos that strongly overlap properties you care about. So where you have like a high score in a property, if the video also has a high score in that property, I'm like, that looks good. But I also want to make sure that it doesn't have high scores in properties you don't care about, right? So the more it matches your preferences. it has the things you like and it doesn't have the things you don't like the better candidate i'm going to say that video is to show to you and so that's what i'm going to do to figure out a video to show to you and then i'll show it to you if i could do that the recommendations i make are going to seem like really good like yeah man how do you know like i i like baseball and you know um like like left-wing politics and i'm really into lord of the rings and my god there was a video somewhere in these like hundred billion you know videos where you have frodo making a sort of like left-wing argument using a baseball analogy right you're like wow you know me really well but it's what we really know is this video matched a bunch of things you cared about and didn't have a lot of the things that you don't A two-tower recommendation system is one way of building an automated system to do something like this. Now, the way this works, I'm going to, God help us, draw a picture, Jesse. This is like an impossible thing to diagram, and I'm a bad... Actually, I take that back. I'm using the wrong pencil. You can all tell I'm a fantastic technologist. I couldn't even figure out how to use an Apple pencil. I was actually using the wrong pencil. OK, so what I'm going to do here is I'm going to draw a picture here, make the two-tower system make this make more sense. All right, so on the screen here, I'm going to draw two on the right over here, a tower. All right, perfect diagram of a tower. So this is our first tower. This one is going to be, inside of it is going to be a collection of machine learning type of tools. There are going to be things in here. They often think about layers, neural networks, and transformers, and embedding matrices. And it's just mathematical manipulation type stuff that's complicated, but you don't need to worry about it. All right, so what we have as input to this tower is we have all of these videos. So people have just uploaded all of these videos. So we just have billions of videos. And then we're going to take these videos, and they're one by one. We input to this tower. And what's going to come out on the other side of the tower is going to be a property list for that video. And I'm going to indicate that with, I don't know, I'll just put like a big purple box. All right. So that purple box is a, it's a vector of numbers, but just think of it as if we have our thousand categories and it has a number for each of it. So it's just describing the video using whatever sort of the sort of master list of things we care about. So we can do this for each of the videos. And in the end, we'll have this, think of it like a big database of billions of videos. And we can do this when they're uploaded, like the videos don't change. And we have a way of describing this big, long list of numbers that describes the properties of each of those videos. That's tower number one. We want videos through it, and it gives us this description of those videos with the properties we care about. All right. So what's tower number two? There's a two-tower situation here. And there's no way, by the way, Jesse, that this terminology did not come out of Lord of the Rings fandom within computer science. Let's be honest. Do you even know that reference? Two Towers is the name of a Lord of the Rings book? No. You spent too much time doing sports. I might have, but I would have lost it. All right. So here's what we have over here. We have a second tower. And the way this tower works is now our input is going to be... And by people, I mean their user profiles. So it's like a description of everything we know about them, including primarily, and this is the key thing, their behavior on the platform. So things they've watched before, things they haven't watched before, how long they watch various things. So it's all this information about these people. And we run each of the people through, they have their own tower, which again, inside of it's mathematical stuff. There's neural networks, there's transformers, there's embedding matrices. Don't worry about the math. And what we get on the other side is it will also explain, describe each person with their list of when it comes to these same properties, hey, how much do they care about each of these things? So we have a common way, one tower that does nothing but describe videos. That's called the item tower. And one tower that does nothing but describe the interest of users. We call that the user tower. And the key thing is, is we describe both these things with the same way, the same list of categories. All right. So that's the idea. Now, how do we teach these towers to do it? Here's the important thing. This is done in a largely like semi-supervised manner using machine learning techniques. So it's humans don't sit down and decide what goes in each of these. What should these categories be? What are these things we are rating? Humans don't, we don't decide. Instead, what we do is we train both of these towers at the same time. In the same way as we train other sort of neural network-based systems like language models or visual vision models or other stuff that you're used to. And here the data we have to train it is we have a lot of examples of, okay, we know this user likes this video. We know this user didn't like this video. You can use that data to train these things together. And what is the goal that you're training them for? You say, come up with, I don't know how you're creating these categories and these numbers and how you're describing things. I don't know. It's not for me as a human to know. But what I want to see is the vectors you use, the descriptions of things that people care about should be pretty close to the vectors you use to describe the stuff we know they like and not too close to the things they don't like. So I don't know what's in them, but I have a bunch of examples of stuff that real people did and what videos they really like or don't like. And I want you to keep nudging and changing your internal descriptions until you get pretty good. at describing people and describing things in such a way that the vector describing the people is close to the things they like and not close to the things they don't. So we don't know what's in these vectors. We don't know how, in these two-tower recommendation systems, what they're looking at, what these neural networks and transformers, et cetera, what they're noticing in these videos, what they're noticing in our human behavior. We have no idea. It's just a list of numbers. But we know it does a good job, that when we test it and say, okay, we know this user likes this video, we say, yeah, you did these things overlap pretty well. And we know this user doesn't like this video. Yeah, they don't overlap very well. So you train these towers together. They learn, in this diagram here, the pink stuff. They learn some useful way of describing users and describing videos so that it does a pretty good job of matching them. Okay, so then if we step back, how does the final recommendation work? All right, so again, now we can draw from some of these architecture papers that ByteDance themselves had published. But what typically happens in these systems is you have so many of these items, so many videos, that what you do is you say, okay, we're going to do like a really rough first path to get some candidates of what to show the user. And we'll use this entirely... something like a distance metric, like a way of just, here's a list of numbers, here's a list of numbers, how close are these list of numbers? There's different ways of measuring this that you can do pretty quickly mathematically. And we're just going to go through and grab a bunch of videos that have a pretty close, they're close by this sort of mathematical notion of close, to the user's description, to their vector. And then there's a little bit of proprietary stuff at the end is how do we then rank these candidates and describe what's the actual one to return. And that's actually a place where recommendation systems can have a little bit of human, human oriented heuristics and rules of thumb in this final step where you're like, okay, here's a hundred videos that are like a good match. Now, which one do we want to show them? That's where you can throw in some actually like hand coded, like final little rules or tweaks or rules of thumb that would kind of happen at the end. So that's basically how these systems work. So why is TikTok so uniquely successful if that's an architecture that like other systems use? Spotify probably uses something like that. Some other social platforms are using something like this. So why is TikTok work so well? Well, we kind of have an answer to that as well. There seems to be a few things going on here, right? One is just what that service does. So short form video is a best case scenario for building one of these recommendation systems. All it does is deliver you stuff. TikTok doesn't have to deal with other complicating factors that other social services have, like your friend graph and who you follow and trying to mix in the stuff you said you're interested in with the stuff the algorithm thinks you're interested in. TikTok basically ignores that because everyone uses the For You tab. That's what we're talking about here. And so it's just pure. We're just showing you things that we think you'll like. Nothing else matters. Second, because they're short, you get a lot of feedback. An average TikTok user might go through 30-plus videos in a typical session, each one generating feedback about what they like and don't like. So you have a huge amount of data with which to get better and better at making these recommendations. Now, compare that to Netflix, where I might watch one series and one movie in a given week. It's a much, much slower data cycle. And I'm not going to try, by the way, if I'm on Netflix, I'm not going to try most of the stuff you recommend. I'm probably going to end up watching something someone told me about anyway. So I get very slow stream of data if I'm a service like Netflix. But TikTok is optimal because you only can look at what they show you. So you're giving them feedback on every single recommendation they get. All right. So that's part of it. The format is super well suited for these type of systems to work really well. The other part of the advantage is actually architectural. It's a really smart and powerful distributed system. that ByteDance actually built for their products like TikTok. So one of the things that they do that's really impressive is they can update. They update the training of the user tower almost in real time. They don't just train these two towers once, then go deploy it, now let's go use it. As you're using the app, you're getting more data, you're generating more data about yourself. Which videos did you watch and how long did you watch them? Well, they built a system, this is really pretty amazing from a distributed systems point of view, that can essentially be constantly trying to retrain your part, the user tower piece, using this new data. so that it can sort of respond in real time. This is really hard to do, but the way they do it is with this massive distributed system where it's fragmented among all these different systems. There's probably a system near you that's working on it. In the US, most of this is an Oracle Cloud Infrastructure report, so there's probably some local machine doing it. And then they transfer over the new train parameters to the production model that's actually making recommendations. really frequently and there's this whole fault tolerance system they have built up so that if this gets partitioned the system can still run it's really hard computer engineering but it allows them to continually update how it labels your infer how it labels you and what you care about like almost immediately in reaction to the stuff you're doing this is what gives tick tock for example it's amazing cold start capability where if you're a new tick tock user you just start watching things and swiping and within 10 minutes like how is this already showing me stuff that i really care about it's because they built this architecture that can retrain parts of the towers in real time alongside of what you're doing the other thing we know they've done in their system is that it's not a pure user-based history-based recommendation system they have a parallel system that's doing nothing but studying what's popular hey what's doing well on our network maybe worldwide or what's doing well in our network in a particular region or among a general group of users, and they mix this in. They call this the short-term profile versus the long-term profile, which is the user description. They join these two things together. So when they're trying to figure out what to show you, yeah, there's a bunch of candidates that really match your expressed interest, but there's also candidates that maybe are a looser fit to your expressed interest, like a reasonable fit, but a much looser fit, but are trending and are really popular right now. And so those get mixed in. And then so you can get shown, not everything you're being shown is just, here's the best match to what you've shown interest in before. It's also like, hey, this thing is really popular right now. People like it. It roughly overlaps stuff you care about. Let's throw that into the mix. And then this becomes a feedback mechanism that allows you to see things that aren't like an exact fit for things you've seen before. And maybe you like it. And it begins, you watch it for a while, and it allows the model when it's describing you to sort of learn about other interests you may or may not have. So it's why you get this mix. I mean, you know this when you use TikTok. You get this mix of, oh, this is straight shot matching to one of my clear interests. But also like, oh, this is weird and kind of compelling, but kind of off the wall. And maybe half of those things you see, you end up watching them. It's it putting in this sort of real-time popular stuff as well. So they have sort of a secret sauce for mixing those two things together. So basically, it's just no big new ideas. It's just a very well-executed system. It's a system that was built at the highest level in a format, short video, one-by-one algorithm recommendation that is perfect for this type of recommendation system. You put those two things together, and the whole thing seems pretty eerie. Like, my God, it learns me so fast. It knows more about me than I thought I knew about myself. And it works really well. All right, so that's what Oracle is going to find. There's no magic in there. These are a well-implemented distributed system. There is no magic description of a newspaper editor in somewhere that you can tweak. Just a really well-built distributed system that runs these machine learning-based categorization algorithms. So what does this teach us? Well, modern recommendation architectures like the one run by TikTok are not digital newspaper editors. They're not things that we can easily configure to reflect particular values or interests or philosophies. The machine learning techniques used in these two tower architectures are completely agnostic to what they're trying to describe. It could be videos. It could be data from a science experiment. It could be descriptions of shopping behavior, moving watching behavior. These algorithms do not care. They've just been optimized in a relentless training model for I am assigning list of numbers to things. And if these numbers are close to the numbers for this thing over here, the user, and the system tells me that's good, then I think my numbers are good. And if the system tells me they're bad, then I adjust how I do it until the system tells me it's good. There is no intention in here. There is no visibility into how things are being described or what matters or what the values are. It's just trying to win this training game of, I don't know in advance as the two towers. who this user is or what they like. So I better have described them in a way that ends up matching the things that they liked. The way these systems actually work in terms of if we want to think about what are they actually doing, if you talk to a machine learning or data scientist, they'll say, yeah, what these techniques do, you give them enough data. What they're trying to do mathematically is build approximations, mathematical approximations of whatever underlying process or systems best describe the patterns it's fed in its training data. This is why if you feed a bunch of information about traffic times or something into one of these models, and it gets good at predicting what's going to happen at given times, it has approximated maybe some sort of reality about the underlying traffic system. Like, you know what? There's a lot more cars between 4.30 and 6.30 because that's when traffic lets out from buildings. It sort of approximates these underlying systems and processes so it can do better at predicting what's going to happen. So when it comes to serving content, like videos that feature other people to other people. This method of curation, I believe, is something that should give us some pause. And the reason is, if we think about this historically, humans have always been a little uneasy and a little wary about the production of mass content. We worry about it because we know content has a real impact, people talking to other people. And mass content, since the beginning of the printing press, has had both good and bad impacts. Just look at like, the witch trials that happened all throughout Europe and eventually making its way to colonial America. A lot of this came out of some printing that got people thinking about this or that. We worry about the power of content. And the reason is, is the human psyche has dark elements. We have a hardwired affinity for hatred or violence or dehumanization. an attraction to the grizzly, an attraction to the purient. We have a lot of dark parts in our brain and we try to appeal to our better angels, especially when it comes to mass content. So we have all sorts of guardrails we put up. If I'm editing a newspaper, if I'm a producer for a television program, if I'm producing a podcast, what I'm going to say or not going to say to accomplish my goals, we're careful about it. We're careful about it because we know there's a lot of stuff in the human brain, a little more primitive that we don't want to appeal to. We integrate human values into how we curate content. And when we don't do that, we get really upset or worried about it. I mean, this was like World War II propaganda. What was that? If not, basically a group saying, throw those guardrails aside. What matters is our goal is more important. And then we look back at like World War II era propaganda or like, ah, that's not great. We don't be like, hey, what great communication? Like, no, no, no, we don't. Go there, even if it could help our cause. We're careful about it. We have human values. These type of recommendation architecture don't share our values because they don't know what values are. They don't know what they're doing. They're just producing numbers to win a game of getting positive or negative zaps from a machine learning algorithm. So when we ask a system like this, hey, do a good job of recommending stuff, it's like, great, I will do what I'm mathematically supposed to do, which is build these mathematical approximations of the underlying systems and processes that help explain the patterns I've seen. That means it's going to be building models of the dark impulses. It's inevitable. It's going to build models about the affinities for hatred or dehumanization or violence or purience or whatever it is that we're embarrassed to admit that we're wired for. These algorithms have no embarrassment. They just say, there's a thing here. We show stuff that appeals to this. It does well. You get basically not a digital newspaper editor, but a digital propagandist of the worst kind. That's what happens when you allow blind mathematical models to start doing content curation. And I think that territory, this is what we should worry about. There's not something here we easily tweak. We're not replacing one country's values with another. There's no knobs to turn about different properties we want or don't want. Machine learning-based recommendation algorithms, architectures, again, just to summarize, they are just going to model the systems at hand blindly. to win the game of finding matching interest. And when it comes to content, that really pushes back against the last 500 years of human experience with how we should deal with content production. All right, so let's move on here to my takeaways. Basically, excuse me, here's some good music. All right, let's go back to the original question. Am I glad that TikTok in America is coming under American control? Well, sure. This can't hurt. And there's places where a foreign government maybe could mess with these architectures to screw with us, especially at that last phase where you do some heuristic tweaking and the final ranking of candidates. But will this somehow allow us to fix TikTok in a fundamental way? If we have the right people controlling the algorithm, can we make these platforms behave in the right ways? No, the answer is no. Machine learning algorithms deployed in this context will relentlessly learn how best to summarize the human condition and exploit us to get it to do whatever it is, whatever goal we have given them. This technique is going to exploit our dark sides just as much as our bright because it doesn't know the difference between them. When it comes to technology in recent decades, I think we have underestimated the degree to which we... We just sort of implicitly integrate our human values and how we operate in many areas of both our personal and civic lives. And as we start conceding more control of these things to technology, technology that cannot by definition share those values, we begin to learn how unsettling things get. We don't realize how much we depended on these just humanistic moral rules of thumb, these normative standards about what's good and bad. And so it's not so innocent to say, let's let an algorithm serve our news. Let's let an algorithm serve our entertainment. Let's let an algorithm be at the center of the town squares. An algorithm is very different than a person. And we don't miss what we have in human types of moralistic thinking until we take it out of our system. So that's my main takeaway about the TikTok situation. I'm sure there's lots of national security concerns and this and that. Privacy concerns, great. But we're far away from solving the problem of social media's dark impulses. It is baked into the mathematics of how these things execute. It is not an accident or a bad feature someone added late in the process that we can remove. Hey, if you like this video, I think you'll really like this one as well. Check it out.

Cal Newport explains how TikTok works in this clip pulled from a fall, 2025 episode of the Deep Question podcast.

Buy Cal Newport's latest book, “Slow Productivity” at www.calnewport.com/slow 

What is the TikTok algorithm? How does it work? To what extent can it be controlled? In this clip, Cal looks deeper at these questions and arrives at a broader philosophical point about the role these services, in general, should play in our civic culture. 

Download my FREE Deep Life Guide HERE: https://bit.ly/3QBIcug

Listen to Episode Here:  https://www.thedeeplife.com/listen/

Get your questions answered by Cal! Here’s the link: https://bit.ly/3U3sTvo

0:00 The algorithm
4:40 Oracle and the algorithm
11:25 Cal illustrates how TikTok works
20:00 Distributed system
26:30 Mass content

Connect with Cal Newport:

🔴Visit Cal's BLOG and website:             https://calnewport.com/blog/
🔴Check out Cal's books:                         https://calnewport.com/writing/
🔴Check out The Deep Life:                     https://thedeeplife.com

About Cal Newport:
Cal Newport is a computer science professor at Georgetown University. In addition to his academic research, he writes about the intersection of digital technology and culture. Cal's particularly interested in our struggle to deploy these tools in ways that support instead of subvert the things we care about in both our personal and professional lives.

Cal is a New York Times bestselling author of seven books, including, most recently, A World Without Email, Digital Minimalism, and Deep Work. He's also the creator of The Time-Block Planner.

The videos are considered to be used under the "Fair Use Doctrine" of United States Copyright Law, Title 17 U.S. Code Sections 107-118. Videos are used for editorial and educational purposes only and I do not claim ownership of any original video content. I don't use said video clips in advertisements, marketing or for direct financial gain. All video content in each clip is considered owned by the individual broadcast companies.

#CalNewport #DeepWork #DeepLife #DeepQuestions #TimeblockPlanner
#WorldWithoutEmail #DeepQuestionsPodcast